Datasets:
Add files using upload-large-folder tool
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +280 -0
- parse/dev/08Yk-n5l2Al/08Yk-n5l2Al.md +372 -0
- parse/dev/0YXmOFLb1wQ/0YXmOFLb1wQ_middle.json +0 -0
- parse/dev/1W8UwXAQubL/1W8UwXAQubL_middle.json +0 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf.md +523 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf_content_list.json +1493 -0
- parse/dev/4WgqjmYacAf/4WgqjmYacAf_model.json +0 -0
- parse/dev/6res1KC1Z3Z/6res1KC1Z3Z_middle.json +0 -0
- parse/dev/6res1KC1Z3Z/6res1KC1Z3Z_model.json +0 -0
- parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ.md +619 -0
- parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_content_list.json +0 -0
- parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_middle.json +0 -0
- parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_model.json +0 -0
- parse/dev/FQOC5u-1egI/FQOC5u-1egI.md +0 -0
- parse/dev/FQOC5u-1egI/FQOC5u-1egI_content_list.json +0 -0
- parse/dev/FQOC5u-1egI/FQOC5u-1egI_middle.json +0 -0
- parse/dev/FQOC5u-1egI/FQOC5u-1egI_model.json +0 -0
- parse/dev/GQjaI9mLet/GQjaI9mLet.md +493 -0
- parse/dev/GQjaI9mLet/GQjaI9mLet_content_list.json +0 -0
- parse/dev/GQjaI9mLet/GQjaI9mLet_middle.json +0 -0
- parse/dev/GQjaI9mLet/GQjaI9mLet_model.json +0 -0
- parse/dev/Jep2ykGUdS/Jep2ykGUdS.md +0 -0
- parse/dev/Jep2ykGUdS/Jep2ykGUdS_content_list.json +0 -0
- parse/dev/Jep2ykGUdS/Jep2ykGUdS_middle.json +0 -0
- parse/dev/Jep2ykGUdS/Jep2ykGUdS_model.json +0 -0
- parse/dev/QwejPcX96r/QwejPcX96r.md +0 -0
- parse/dev/QwejPcX96r/QwejPcX96r_middle.json +0 -0
- parse/dev/S7Evzt9uit3/S7Evzt9uit3.md +279 -0
- parse/dev/S7Evzt9uit3/S7Evzt9uit3_middle.json +0 -0
- parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu.md +361 -0
- parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_content_list.json +1940 -0
- parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_middle.json +0 -0
- parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_model.json +0 -0
- parse/dev/Uczck6TlSZ/Uczck6TlSZ.md +383 -0
- parse/dev/Uczck6TlSZ/Uczck6TlSZ_content_list.json +0 -0
- parse/dev/Uczck6TlSZ/Uczck6TlSZ_middle.json +0 -0
- parse/dev/Uczck6TlSZ/Uczck6TlSZ_model.json +0 -0
- parse/dev/VD-AYtP0dve/VD-AYtP0dve_model.json +0 -0
- parse/dev/WqiZJGNkjn/WqiZJGNkjn.md +204 -0
- parse/dev/WqiZJGNkjn/WqiZJGNkjn_content_list.json +842 -0
- parse/dev/WqiZJGNkjn/WqiZJGNkjn_middle.json +0 -0
- parse/dev/WqiZJGNkjn/WqiZJGNkjn_model.json +0 -0
- parse/dev/XA4ru9mfxTP/XA4ru9mfxTP.md +283 -0
- parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_content_list.json +1162 -0
- parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_middle.json +0 -0
- parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_model.json +0 -0
- parse/dev/XdDl3bFUNn5/XdDl3bFUNn5_middle.json +0 -0
- parse/dev/XtyeppctGgc/XtyeppctGgc.md +311 -0
- parse/dev/XtyeppctGgc/XtyeppctGgc_content_list.json +1264 -0
- parse/dev/XtyeppctGgc/XtyeppctGgc_middle.json +0 -0
.gitattributes
CHANGED
|
@@ -14424,3 +14424,283 @@ parse/train/aFvG-DNPNB9/aFvG-DNPNB9_layout.pdf filter=lfs diff=lfs merge=lfs -te
|
|
| 14424 |
parse/train/aFvG-DNPNB9/aFvG-DNPNB9_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14425 |
parse/train/aFvG-DNPNB9/aFvG-DNPNB9_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14426 |
parse/train/BJe932EYwS/BJe932EYwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14424 |
parse/train/aFvG-DNPNB9/aFvG-DNPNB9_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14425 |
parse/train/aFvG-DNPNB9/aFvG-DNPNB9_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14426 |
parse/train/BJe932EYwS/BJe932EYwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14427 |
+
parse/train/BJe932EYwS/BJe932EYwS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14428 |
+
parse/train/BJe932EYwS/BJe932EYwS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14429 |
+
parse/train/iKQAk8a2kM0/iKQAk8a2kM0_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14430 |
+
parse/train/iKQAk8a2kM0/iKQAk8a2kM0_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14431 |
+
parse/train/iKQAk8a2kM0/iKQAk8a2kM0_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14432 |
+
parse/train/OkFPq7ZtsQ/OkFPq7ZtsQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14433 |
+
parse/train/OkFPq7ZtsQ/OkFPq7ZtsQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14434 |
+
parse/train/OkFPq7ZtsQ/OkFPq7ZtsQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14435 |
+
parse/train/HkgaETNtDB/HkgaETNtDB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14436 |
+
parse/train/HkgaETNtDB/HkgaETNtDB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14437 |
+
parse/train/HkgaETNtDB/HkgaETNtDB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14438 |
+
parse/train/N6JECD-PI5w/N6JECD-PI5w_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14439 |
+
parse/train/N6JECD-PI5w/N6JECD-PI5w_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14440 |
+
parse/train/N6JECD-PI5w/N6JECD-PI5w_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14441 |
+
parse/train/B1MbDj0ctQ/B1MbDj0ctQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14442 |
+
parse/train/B1MbDj0ctQ/B1MbDj0ctQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14443 |
+
parse/train/B1MbDj0ctQ/B1MbDj0ctQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14444 |
+
parse/train/rsRq--gsiE/rsRq--gsiE_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14445 |
+
parse/train/rsRq--gsiE/rsRq--gsiE_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14446 |
+
parse/train/rsRq--gsiE/rsRq--gsiE_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14447 |
+
parse/train/SJi9WOeRb/SJi9WOeRb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14448 |
+
parse/train/SJi9WOeRb/SJi9WOeRb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14449 |
+
parse/train/SJi9WOeRb/SJi9WOeRb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14450 |
+
parse/train/UoVpP8R2Vn/UoVpP8R2Vn_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14451 |
+
parse/train/UoVpP8R2Vn/UoVpP8R2Vn_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14452 |
+
parse/train/UoVpP8R2Vn/UoVpP8R2Vn_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14453 |
+
parse/train/SJxbHkrKDH/SJxbHkrKDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14454 |
+
parse/train/SJxbHkrKDH/SJxbHkrKDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14455 |
+
parse/train/SJxbHkrKDH/SJxbHkrKDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14456 |
+
parse/train/SJzRZ-WCZ/SJzRZ-WCZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14457 |
+
parse/train/SJzRZ-WCZ/SJzRZ-WCZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14458 |
+
parse/train/SJzRZ-WCZ/SJzRZ-WCZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14459 |
+
parse/train/4qR3coiNaIv/4qR3coiNaIv_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14460 |
+
parse/train/4qR3coiNaIv/4qR3coiNaIv_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14461 |
+
parse/train/4qR3coiNaIv/4qR3coiNaIv_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14462 |
+
parse/train/ryfMLoCqtQ/ryfMLoCqtQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14463 |
+
parse/train/ryfMLoCqtQ/ryfMLoCqtQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14464 |
+
parse/train/ryfMLoCqtQ/ryfMLoCqtQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14465 |
+
parse/train/SJA7xfb0b/SJA7xfb0b_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14466 |
+
parse/train/SJA7xfb0b/SJA7xfb0b_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14467 |
+
parse/train/SJA7xfb0b/SJA7xfb0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14468 |
+
parse/train/SJl98sR5tX/SJl98sR5tX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14469 |
+
parse/train/SJl98sR5tX/SJl98sR5tX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14470 |
+
parse/train/SJl98sR5tX/SJl98sR5tX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14471 |
+
parse/train/BJIgi_eCZ/BJIgi_eCZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14472 |
+
parse/train/BJIgi_eCZ/BJIgi_eCZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14473 |
+
parse/train/BJIgi_eCZ/BJIgi_eCZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14474 |
+
parse/train/ByKWUeWA-/ByKWUeWA-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14475 |
+
parse/train/ByKWUeWA-/ByKWUeWA-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14476 |
+
parse/train/ByKWUeWA-/ByKWUeWA-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14477 |
+
parse/train/HyenWc5gx/HyenWc5gx_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14478 |
+
parse/train/HyenWc5gx/HyenWc5gx_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14479 |
+
parse/train/HyenWc5gx/HyenWc5gx_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14480 |
+
parse/train/Hk3mPK5gg/Hk3mPK5gg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14481 |
+
parse/train/Hk3mPK5gg/Hk3mPK5gg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14482 |
+
parse/train/Hk3mPK5gg/Hk3mPK5gg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14483 |
+
parse/train/B1xtd1HtPS/B1xtd1HtPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14484 |
+
parse/train/B1xtd1HtPS/B1xtd1HtPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14485 |
+
parse/train/B1xtd1HtPS/B1xtd1HtPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14486 |
+
parse/train/jnMjOctlfbZ/jnMjOctlfbZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14487 |
+
parse/train/jnMjOctlfbZ/jnMjOctlfbZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14488 |
+
parse/train/jnMjOctlfbZ/jnMjOctlfbZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14489 |
+
parse/train/6vWuYzkp8d/6vWuYzkp8d_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14490 |
+
parse/train/6vWuYzkp8d/6vWuYzkp8d_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14491 |
+
parse/train/6vWuYzkp8d/6vWuYzkp8d_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14492 |
+
parse/train/HJeOekHKwr/HJeOekHKwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14493 |
+
parse/train/HJeOekHKwr/HJeOekHKwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14494 |
+
parse/train/HJeOekHKwr/HJeOekHKwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14495 |
+
parse/train/BJe-DsC5Fm/BJe-DsC5Fm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14496 |
+
parse/train/BJe-DsC5Fm/BJe-DsC5Fm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14497 |
+
parse/train/BJe-DsC5Fm/BJe-DsC5Fm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14498 |
+
parse/train/SJg7KhVKPH/SJg7KhVKPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14499 |
+
parse/train/SJg7KhVKPH/SJg7KhVKPH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14500 |
+
parse/train/SJg7KhVKPH/SJg7KhVKPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14501 |
+
parse/train/LuyryrCs6Ez/LuyryrCs6Ez_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14502 |
+
parse/train/LuyryrCs6Ez/LuyryrCs6Ez_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14503 |
+
parse/train/LuyryrCs6Ez/LuyryrCs6Ez_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14504 |
+
parse/train/B1nLkl-0Z/B1nLkl-0Z_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14505 |
+
parse/train/B1nLkl-0Z/B1nLkl-0Z_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14506 |
+
parse/train/B1nLkl-0Z/B1nLkl-0Z_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14507 |
+
parse/train/SJx9GQb0-/SJx9GQb0-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14508 |
+
parse/train/SJx9GQb0-/SJx9GQb0-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14509 |
+
parse/train/SJx9GQb0-/SJx9GQb0-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14510 |
+
parse/train/ryiAv2xAZ/ryiAv2xAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14511 |
+
parse/train/ryiAv2xAZ/ryiAv2xAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14512 |
+
parse/train/ryiAv2xAZ/ryiAv2xAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14513 |
+
parse/train/HylzTiC5Km/HylzTiC5Km_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14514 |
+
parse/train/HylzTiC5Km/HylzTiC5Km_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14515 |
+
parse/train/HylzTiC5Km/HylzTiC5Km_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14516 |
+
parse/train/S1fduCl0b/S1fduCl0b_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14517 |
+
parse/train/S1fduCl0b/S1fduCl0b_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14518 |
+
parse/train/S1fduCl0b/S1fduCl0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14519 |
+
parse/train/r1HhRfWRZ/r1HhRfWRZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14520 |
+
parse/train/r1HhRfWRZ/r1HhRfWRZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14521 |
+
parse/train/r1HhRfWRZ/r1HhRfWRZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14522 |
+
parse/train/ByeUBANtvB/ByeUBANtvB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14523 |
+
parse/train/ByeUBANtvB/ByeUBANtvB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14524 |
+
parse/train/ByeUBANtvB/ByeUBANtvB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14525 |
+
parse/train/wpSWuz_hyqA/wpSWuz_hyqA_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14526 |
+
parse/train/wpSWuz_hyqA/wpSWuz_hyqA_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14527 |
+
parse/train/wpSWuz_hyqA/wpSWuz_hyqA_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14528 |
+
parse/train/H1xsSjC9Ym/H1xsSjC9Ym_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14529 |
+
parse/train/H1xsSjC9Ym/H1xsSjC9Ym_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14530 |
+
parse/train/H1xsSjC9Ym/H1xsSjC9Ym_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14531 |
+
parse/train/HkxnclHKDr/HkxnclHKDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14532 |
+
parse/train/HkxnclHKDr/HkxnclHKDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14533 |
+
parse/train/HkxnclHKDr/HkxnclHKDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14534 |
+
parse/train/rkeYL1SFvH/rkeYL1SFvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14535 |
+
parse/train/rkeYL1SFvH/rkeYL1SFvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14536 |
+
parse/train/rkeYL1SFvH/rkeYL1SFvH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14537 |
+
parse/train/wRXzOa2z5T/wRXzOa2z5T_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14538 |
+
parse/train/wRXzOa2z5T/wRXzOa2z5T_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14539 |
+
parse/train/wRXzOa2z5T/wRXzOa2z5T_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14540 |
+
parse/train/Bkgwp3NtDH/Bkgwp3NtDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14541 |
+
parse/train/Bkgwp3NtDH/Bkgwp3NtDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14542 |
+
parse/train/Bkgwp3NtDH/Bkgwp3NtDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14543 |
+
parse/train/Rt5mjXAqHrY/Rt5mjXAqHrY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14544 |
+
parse/train/Rt5mjXAqHrY/Rt5mjXAqHrY_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14545 |
+
parse/train/Rt5mjXAqHrY/Rt5mjXAqHrY_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14546 |
+
parse/train/onxoVA9FxMw/onxoVA9FxMw_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14547 |
+
parse/train/onxoVA9FxMw/onxoVA9FxMw_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14548 |
+
parse/train/onxoVA9FxMw/onxoVA9FxMw_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14549 |
+
parse/train/r1lnxTEYPS/r1lnxTEYPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14550 |
+
parse/train/r1lnxTEYPS/r1lnxTEYPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14551 |
+
parse/train/r1lnxTEYPS/r1lnxTEYPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14552 |
+
parse/train/rsf1z-JSj87/rsf1z-JSj87_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14553 |
+
parse/train/rsf1z-JSj87/rsf1z-JSj87_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14554 |
+
parse/train/rsf1z-JSj87/rsf1z-JSj87_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14555 |
+
parse/train/SJk01vogl/SJk01vogl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14556 |
+
parse/train/SJk01vogl/SJk01vogl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14557 |
+
parse/train/SJk01vogl/SJk01vogl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14558 |
+
parse/train/PdauS7wZBfC/PdauS7wZBfC_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14559 |
+
parse/train/PdauS7wZBfC/PdauS7wZBfC_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14560 |
+
parse/train/PdauS7wZBfC/PdauS7wZBfC_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14561 |
+
parse/train/rkevMnRqYQ/rkevMnRqYQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14562 |
+
parse/train/rkevMnRqYQ/rkevMnRqYQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14563 |
+
parse/train/rkevMnRqYQ/rkevMnRqYQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14564 |
+
parse/train/HJxiMAVtPH/HJxiMAVtPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14565 |
+
parse/train/HJxiMAVtPH/HJxiMAVtPH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14566 |
+
parse/train/HJxiMAVtPH/HJxiMAVtPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14567 |
+
parse/train/AY8zfZm0tDd/AY8zfZm0tDd_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14568 |
+
parse/train/AY8zfZm0tDd/AY8zfZm0tDd_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14569 |
+
parse/train/AY8zfZm0tDd/AY8zfZm0tDd_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14570 |
+
parse/train/Bkl086VYvH/Bkl086VYvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14571 |
+
parse/train/Bkl086VYvH/Bkl086VYvH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14572 |
+
parse/train/Bkl086VYvH/Bkl086VYvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14573 |
+
parse/train/v_1Soh8QUNc/v_1Soh8QUNc_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14574 |
+
parse/train/v_1Soh8QUNc/v_1Soh8QUNc_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14575 |
+
parse/train/v_1Soh8QUNc/v_1Soh8QUNc_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14576 |
+
parse/train/ryH20GbRW/ryH20GbRW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14577 |
+
parse/train/ryH20GbRW/ryH20GbRW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14578 |
+
parse/train/ryH20GbRW/ryH20GbRW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14579 |
+
parse/train/RovX-uQ1Hua/RovX-uQ1Hua_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14580 |
+
parse/train/RovX-uQ1Hua/RovX-uQ1Hua_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14581 |
+
parse/train/RovX-uQ1Hua/RovX-uQ1Hua_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14582 |
+
parse/train/Hk0olMdZOIU/Hk0olMdZOIU_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14583 |
+
parse/train/Hk0olMdZOIU/Hk0olMdZOIU_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14584 |
+
parse/train/Hk0olMdZOIU/Hk0olMdZOIU_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14585 |
+
parse/train/Goz-qsH1F14/Goz-qsH1F14_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14586 |
+
parse/train/Goz-qsH1F14/Goz-qsH1F14_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14587 |
+
parse/train/Goz-qsH1F14/Goz-qsH1F14_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14588 |
+
parse/train/BJlgNh0qKQ/BJlgNh0qKQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14589 |
+
parse/train/BJlgNh0qKQ/BJlgNh0qKQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14590 |
+
parse/train/BJlgNh0qKQ/BJlgNh0qKQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14591 |
+
parse/train/H1lVvgHKDr/H1lVvgHKDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14592 |
+
parse/train/H1lVvgHKDr/H1lVvgHKDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14593 |
+
parse/train/H1lVvgHKDr/H1lVvgHKDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14594 |
+
parse/train/cu7IUiOhujH/cu7IUiOhujH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14595 |
+
parse/train/cu7IUiOhujH/cu7IUiOhujH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14596 |
+
parse/train/cu7IUiOhujH/cu7IUiOhujH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14597 |
+
parse/train/QtTKTdVrFBB/QtTKTdVrFBB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14598 |
+
parse/train/QtTKTdVrFBB/QtTKTdVrFBB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14599 |
+
parse/train/QtTKTdVrFBB/QtTKTdVrFBB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14600 |
+
parse/train/0migj5lyUZl/0migj5lyUZl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14601 |
+
parse/train/0migj5lyUZl/0migj5lyUZl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14602 |
+
parse/train/0migj5lyUZl/0migj5lyUZl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14603 |
+
parse/train/SkeXehR9t7/SkeXehR9t7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14604 |
+
parse/train/SkeXehR9t7/SkeXehR9t7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14605 |
+
parse/train/SkeXehR9t7/SkeXehR9t7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14606 |
+
parse/train/s95BePNvykX/s95BePNvykX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14607 |
+
parse/train/s95BePNvykX/s95BePNvykX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14608 |
+
parse/train/s95BePNvykX/s95BePNvykX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14609 |
+
parse/train/SJlpYJBKvH/SJlpYJBKvH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14610 |
+
parse/train/SJlpYJBKvH/SJlpYJBKvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14611 |
+
parse/train/SJlpYJBKvH/SJlpYJBKvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14612 |
+
parse/train/Pz_dcqfcKW8/Pz_dcqfcKW8_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14613 |
+
parse/train/Pz_dcqfcKW8/Pz_dcqfcKW8_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14614 |
+
parse/train/Pz_dcqfcKW8/Pz_dcqfcKW8_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14615 |
+
parse/train/BJgnXpVYwS/BJgnXpVYwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14616 |
+
parse/train/BJgnXpVYwS/BJgnXpVYwS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14617 |
+
parse/train/BJgnXpVYwS/BJgnXpVYwS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14618 |
+
parse/train/HJeRkh05Km/HJeRkh05Km_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14619 |
+
parse/train/HJeRkh05Km/HJeRkh05Km_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14620 |
+
parse/train/HJeRkh05Km/HJeRkh05Km_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14621 |
+
parse/train/rkaT3zWCZ/rkaT3zWCZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14622 |
+
parse/train/rkaT3zWCZ/rkaT3zWCZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14623 |
+
parse/train/rkaT3zWCZ/rkaT3zWCZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14624 |
+
parse/train/iAX0l6Cz8ub/iAX0l6Cz8ub_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14625 |
+
parse/train/iAX0l6Cz8ub/iAX0l6Cz8ub_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14626 |
+
parse/train/iAX0l6Cz8ub/iAX0l6Cz8ub_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14627 |
+
parse/train/SkE6PjC9KX/SkE6PjC9KX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14628 |
+
parse/train/SkE6PjC9KX/SkE6PjC9KX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14629 |
+
parse/train/SkE6PjC9KX/SkE6PjC9KX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14630 |
+
parse/train/rJY3vK9eg/rJY3vK9eg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14631 |
+
parse/train/rJY3vK9eg/rJY3vK9eg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14632 |
+
parse/train/rJY3vK9eg/rJY3vK9eg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14633 |
+
parse/train/SJgwzCEKwH/SJgwzCEKwH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14634 |
+
parse/train/SJgwzCEKwH/SJgwzCEKwH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14635 |
+
parse/train/SJgwzCEKwH/SJgwzCEKwH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14636 |
+
parse/train/rkcQFMZRb/rkcQFMZRb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14637 |
+
parse/train/rkcQFMZRb/rkcQFMZRb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14638 |
+
parse/train/rkcQFMZRb/rkcQFMZRb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14639 |
+
parse/train/r1l1myStwr/r1l1myStwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14640 |
+
parse/train/r1l1myStwr/r1l1myStwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14641 |
+
parse/train/r1l1myStwr/r1l1myStwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14642 |
+
parse/train/HkXWCMbRW/HkXWCMbRW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14643 |
+
parse/train/HkXWCMbRW/HkXWCMbRW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14644 |
+
parse/train/HkXWCMbRW/HkXWCMbRW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14645 |
+
parse/train/SJ1nzBeA-/SJ1nzBeA-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14646 |
+
parse/train/SJ1nzBeA-/SJ1nzBeA-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14647 |
+
parse/train/SJ1nzBeA-/SJ1nzBeA-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14648 |
+
parse/train/rkl6As0cF7/rkl6As0cF7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14649 |
+
parse/train/rkl6As0cF7/rkl6As0cF7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14650 |
+
parse/train/rkl6As0cF7/rkl6As0cF7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14651 |
+
parse/train/HknbyQbC-/HknbyQbC-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14652 |
+
parse/train/HknbyQbC-/HknbyQbC-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14653 |
+
parse/train/HknbyQbC-/HknbyQbC-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14654 |
+
parse/train/H1ldNoC9tX/H1ldNoC9tX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14655 |
+
parse/train/H1ldNoC9tX/H1ldNoC9tX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14656 |
+
parse/train/H1ldNoC9tX/H1ldNoC9tX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14657 |
+
parse/train/1AoMhc_9jER/1AoMhc_9jER_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14658 |
+
parse/train/1AoMhc_9jER/1AoMhc_9jER_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14659 |
+
parse/train/1AoMhc_9jER/1AoMhc_9jER_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14660 |
+
parse/train/HJMN-xWC-/HJMN-xWC-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14661 |
+
parse/train/HJMN-xWC-/HJMN-xWC-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14662 |
+
parse/train/HJMN-xWC-/HJMN-xWC-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14663 |
+
parse/train/HkcdHtqlx/HkcdHtqlx_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14664 |
+
parse/train/HkcdHtqlx/HkcdHtqlx_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14665 |
+
parse/train/HkcdHtqlx/HkcdHtqlx_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14666 |
+
parse/train/r1ecqn4YwB/r1ecqn4YwB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14667 |
+
parse/train/r1ecqn4YwB/r1ecqn4YwB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14668 |
+
parse/train/r1ecqn4YwB/r1ecqn4YwB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14669 |
+
parse/train/52XXcK8jY0J/52XXcK8jY0J_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14670 |
+
parse/train/52XXcK8jY0J/52XXcK8jY0J_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14671 |
+
parse/train/52XXcK8jY0J/52XXcK8jY0J_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14672 |
+
parse/train/zfO1MwBFu-/zfO1MwBFu-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14673 |
+
parse/train/zfO1MwBFu-/zfO1MwBFu-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14674 |
+
parse/train/zfO1MwBFu-/zfO1MwBFu-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14675 |
+
parse/train/BylaUTNtPS/BylaUTNtPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14676 |
+
parse/train/BylaUTNtPS/BylaUTNtPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14677 |
+
parse/train/BylaUTNtPS/BylaUTNtPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14678 |
+
parse/train/2pJZSVcSZz/2pJZSVcSZz_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14679 |
+
parse/train/2pJZSVcSZz/2pJZSVcSZz_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14680 |
+
parse/train/2pJZSVcSZz/2pJZSVcSZz_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14681 |
+
parse/train/HktK4BeCZ/HktK4BeCZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14682 |
+
parse/train/HktK4BeCZ/HktK4BeCZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14683 |
+
parse/train/HktK4BeCZ/HktK4BeCZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14684 |
+
parse/train/H18uzzWAZ/H18uzzWAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14685 |
+
parse/train/H18uzzWAZ/H18uzzWAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14686 |
+
parse/train/H18uzzWAZ/H18uzzWAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14687 |
+
parse/train/NsMLjcFaO8O/NsMLjcFaO8O_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14688 |
+
parse/train/NsMLjcFaO8O/NsMLjcFaO8O_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14689 |
+
parse/train/NsMLjcFaO8O/NsMLjcFaO8O_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14690 |
+
parse/train/Eql5b1_hTE4/Eql5b1_hTE4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14691 |
+
parse/train/Eql5b1_hTE4/Eql5b1_hTE4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14692 |
+
parse/train/Eql5b1_hTE4/Eql5b1_hTE4_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14693 |
+
parse/train/H196sainb/H196sainb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14694 |
+
parse/train/H196sainb/H196sainb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14695 |
+
parse/train/H196sainb/H196sainb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14696 |
+
parse/train/qrwe7XHTmYb/qrwe7XHTmYb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14697 |
+
parse/train/qrwe7XHTmYb/qrwe7XHTmYb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14698 |
+
parse/train/qrwe7XHTmYb/qrwe7XHTmYb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14699 |
+
parse/train/Hy_o3x-0b/Hy_o3x-0b_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14700 |
+
parse/train/Hy_o3x-0b/Hy_o3x-0b_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14701 |
+
parse/train/Hy_o3x-0b/Hy_o3x-0b_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14702 |
+
parse/train/qWtmNpgjD5K/qWtmNpgjD5K_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14703 |
+
parse/train/qWtmNpgjD5K/qWtmNpgjD5K_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14704 |
+
parse/train/qWtmNpgjD5K/qWtmNpgjD5K_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14705 |
+
parse/train/XjIm8hOCxTm/XjIm8hOCxTm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 14706 |
+
parse/train/XjIm8hOCxTm/XjIm8hOCxTm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
parse/dev/08Yk-n5l2Al/08Yk-n5l2Al.md
ADDED
|
@@ -0,0 +1,372 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
|
| 2 |
+
|
| 3 |
+
# Chitwan Saharia∗, William Chan∗, Saurabh Saxena†, Lala Li†, Jay Whang†, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Raphael Gontijo-Lopes, Tim Salimans, Jonathan Ho†, David J Fleet†‡, Mohammad Norouzi∗
|
| 4 |
+
|
| 5 |
+
{sahariac,williamchan,mnorouzi}@google.com {srbs,lala,jwhang,jonathanho,davidfleet}@google.com
|
| 6 |
+
|
| 7 |
+
Google Research, Brain Team Toronto, Ontario, Canada
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
We present Imagen, a text-to-image diffusion model with an unprecedented degree of photorealism and a deep level of language understanding. Imagen builds on the power of large transformer language models in understanding text and hinges on the strength of diffusion models in high-fidelity image generation. Our key discovery is that generic large language models (e.g. T5), pretrained on text-only corpora, are surprisingly effective at encoding text for image synthesis: increasing the size of the language model in Imagen boosts both sample fidelity and image-text alignment much more than increasing the size of the image diffusion model. Imagen achieves a new state-of-the-art FID score of 7.27 on the COCO dataset, without ever training on COCO, and human raters find Imagen samples to be on par with the COCO data itself in image-text alignment. To assess text-to-image models in greater depth, we introduce DrawBench, a comprehensive and challenging benchmark for text-to-image models. With DrawBench, we compare Imagen with recent methods including VQ-GAN $^ +$ CLIP, Latent Diffusion Models, GLIDE and DALL-E 2, and find that human raters prefer Imagen over other models in side-by-side comparisons, both in terms of sample quality and image-text alignment.
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Multimodal learning has come into prominence recently, with text-to-image synthesis [55, 12, 59] and image-text contrastive learning [51, 32, 77] at the forefront. These models have transformed the research community and captured widespread public attention with creative image generation [22, 56] and editing applications [21, 43, 36]. To pursue this research direction further, we introduce Imagen, a text-to-image diffusion model that combines the power of transformer language models (LMs) [15, 54] with high-fidelity diffusion models [28, 29, 16, 43] to deliver an unprecedented degree of photorealism and a deep level of language understanding in text-to-image synthesis. In contrast to prior work that uses only image-text data for model training [e.g., 55, 43], the key finding behind Imagen is that text embeddings from large LMs [54, 15], pretrained on text-only corpora, are remarkably effective for text-to-image synthesis. See Fig. 1 for select samples.
|
| 16 |
+
|
| 17 |
+
Imagen comprises a frozen T5-XXL [54] encoder to map input text into a sequence of embeddings and a $6 4 { \times } 6 4$ image diffusion model, followed by two super-resolution diffusion models for generating
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
Sprouts in the shape of text ‘Imagen’ coming out of a fairytale book.
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
A photo of a Shiba Inu dog with a backpack riding a bike. It is wearing sunglasses and a beach hat.
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
A high contrast portrait of a very happy fuzzy panda dressed as a chef in a high end kitchen making dough. There is a painting of flowers on the wall behind him.
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
Teddy bears swimming at the Olympics $4 0 0 \mathrm { m }$ Butterfly event.
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
A cute corgi lives in a house made out of sushi.
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
A cute sloth holding a small treasure chest. A bright golden glow is coming from the chest.
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
A dragon fruit wearing karate belt in the snow.
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
A strawberry mug filled with white sesame seeds. The mug is floating in a dark chocolate sea.
|
| 44 |
+
|
| 45 |
+
A brain riding a rocketship heading towards the moon.
|
| 46 |
+
|
| 47 |
+
Figure 1: Select $1 0 2 4 \times 1 0 2 4$ Imagen samples for various text inputs. We only include photorealistic images in this figure and leave artistic content to the Appendix, since generating photorealistic images is more challenging from a technical point of view. Figs. A.1 to A.3 show more samples.
|
| 48 |
+
|
| 49 |
+
$2 5 6 \times 2 5 6$ and $1 0 2 4 \times 1 0 2 4$ images (see Fig. A.4). All diffusion models are conditioned on the text embedding sequence and use classifier-free guidance [27]. Imagen relies on new sampling techniques to allow usage of large guidance weights without sample quality degradation observed in prior work, resulting in images with higher fidelity and better image-text alignment than previously possible.
|
| 50 |
+
|
| 51 |
+
While conceptually simple and easy to train, Imagen yields surprisingly strong results. Imagen outperforms other methods on COCO [38] with zero-shot FID-30K of 7.27, significantly outperforming prior work such as GLIDE [43] (at 12.4) and the concurrent work of DALL-E 2 [56] (at 10.4). Our zero-shot FID score is also better than state-of-the-art models trained on COCO, e.g., Make-A-Scene [22] (at 7.6). Additionally, human raters indicate that generated samples from Imagen are on-par in image-text alignment to the reference images on COCO captions.
|
| 52 |
+
|
| 53 |
+
We introduce DrawBench, a new structured suite of text prompts for text-to-image evaluation. DrawBench enables deeper insights through a multi-dimensional evaluation of text-to-image models, with text prompts designed to probe different semantic properties of models. These include compositionality, cardinality, spatial relations, the ability to handle complex text prompts or prompts with rare words, and they include creative prompts that push the limits of models’ ability to generate highly implausible scenes well beyond the scope of the training data. With DrawBench, extensive human evaluation shows that Imagen outperforms other recent methods [59, 12, 56] by a significant margin. We further demonstrate some of the clear advantages of the use of large pre-trained language models [54] over multi-modal embeddings such as CLIP [51] as a text encoder for Imagen.
|
| 54 |
+
|
| 55 |
+
Key contributions of the paper include:
|
| 56 |
+
|
| 57 |
+
1. We discover that large frozen language models trained only on text data are surprisingly very effective text encoders for text-to-image generation, and that scaling the size of frozen text encoder improves sample quality significantly more than scaling the size of image diffusion model.
|
| 58 |
+
2. We introduce dynamic thresholding, a new diffusion sampling technique to leverage high guidance weights and generating more photorealistic and detailed images than previously possible.
|
| 59 |
+
3. We highlight several important diffusion architecture design choices and propose Efficient U-Net, a new architecture variant which is simpler, converges faster and is more memory efficient.
|
| 60 |
+
4. We achieve a new state-of-the-art COCO FID of 7.27. Human raters find Imagen to be on-par with the reference images in terms of image-text alignment.
|
| 61 |
+
5. We introduce DrawBench, a new comprehensive and challenging evaluation benchmark for the text-to-image task. On DrawBench human evaluation, we find Imagen to outperform all other work, including the concurrent work of DALL-E 2 [56].
|
| 62 |
+
|
| 63 |
+
# 2 Imagen
|
| 64 |
+
|
| 65 |
+
Imagen consists of a text encoder that maps text to a sequence of embeddings and a cascade of conditional diffusion models that map these embeddings to images of increasing resolutions (see Fig. A.4). In the following subsections, we describe each of these components in detail.
|
| 66 |
+
|
| 67 |
+
# 2.1 Pretrained text encoders
|
| 68 |
+
|
| 69 |
+
Text-to-image models need powerful semantic text encoders to capture the complexity and compositionality of arbitrary natural language text inputs. Text encoders trained on paired image-text data are standard in current text-to-image models; they can be trained from scratch [43, 55] or pretrained on image-text data [56] (e.g., CLIP [51]). The image-text training objectives suggest that these text encoders may encode visually semantic and meaningful representations especially relevant for the text-to-image generation task. Large language models can be another models of choice to encode text for text-to-image generation. Recent progress in large language models (e.g., BERT [15], GPT [49, 50, 7], T5 [54]) have led to leaps in textual understanding and generative capabilities. Language models are trained on text only corpus significantly larger than paired image-text data, thus being exposed to a very rich and wide distribution of text. These models are also generally much larger than text encoders in current image-text models [51, 32, 83] (e.g. PaLM [11] has 540B parameters, while CoCa [83] has a $\approx 1 \mathbf { B }$ parameter text encoder).
|
| 70 |
+
|
| 71 |
+
It thus becomes natural to explore both families of text encoders for the text-to-image task. Imagen explores pretrained text encoders: BERT [15], T5 [53] and CLIP [48]. For simplicity, we freeze the weights of these text encoders. Freezing has several advantages such as offline computation of embeddings, resulting in negligible computation or memory footprint during training of the textto-image model. In our work, we find that there is a clear conviction that scaling the text encoder size improves the quality of text-to-image generation. We also find that while T5-XXL and CLIP text encoders perform similarly on simple benchmarks such as MS-COCO, human evaluators prefer T5-XXL encoders over CLIP text encoders in both image-text alignment and image fidelity on DrawBench, a set of challenging and compositional prompts. We refer the reader to Section 4.4 for summary of our findings, and Appendix D.1 for detailed ablations.
|
| 72 |
+
|
| 73 |
+
# 2.2 Diffusion models and classifier-free guidance
|
| 74 |
+
|
| 75 |
+
Here we give a brief introduction to diffusion models; a precise description is in Appendix A. Diffusion models [66, 28, 68] are a class of generative models that convert Gaussian noise into samples from a learned data distribution via an iterative denoising process. These models can be conditional, for example on class labels, text, or low-resolution images [e.g. 16, 29, 62, 61, 78, 43, 56]. A diffusion model $\hat { \mathbf { x } } _ { \theta }$ is trained on a denoising objective of the form
|
| 76 |
+
|
| 77 |
+
$$
|
| 78 |
+
\mathbb { E } _ { \mathbf { x } , \mathbf { c } , \epsilon , t } \Big [ w _ { t } \big \| \hat { \mathbf { x } } _ { \boldsymbol { \theta } } \big ( \alpha _ { t } \mathbf { x } + \sigma _ { t } \mathbf { \epsilon } , \mathbf { c } \big ) - \mathbf { x } \big \| _ { 2 } ^ { 2 } \Big ]
|
| 79 |
+
$$
|
| 80 |
+
|
| 81 |
+
where $\displaystyle ( \mathbf { x } , \mathbf { c } )$ are data-conditioning pairs, $t \sim \mathcal { U } ( [ 0 , 1 ] )$ , $\mathbf { \epsilon } \epsilon \sim \mathcal { N } ( \mathbf { 0 } , \mathbf { I } )$ , and $\alpha _ { t } , \sigma _ { t } , w _ { t }$ are functions of $t$ that influence sample quality. Intuitively, $\hat { \mathbf { x } } _ { \theta }$ is trained to denoise $\mathbf { z } _ { t } : = \alpha _ { t } \mathbf { x } + \sigma _ { t } \mathbf { \epsilon } \mathbf { \epsilon }$ into $\mathbf { x }$ using a squared error loss, weighted to emphasize certain values of $t$ . Sampling such as the ancestral sampler [28] and DDIM [67] start from pure noise ${ \bf z } _ { 1 } \sim \mathcal { N } ( { \bf 0 } , { \bf I } )$ and iteratively generate points $\mathbf { z } _ { t _ { 1 } } , \ldots , \mathbf { z } _ { t _ { T } }$ , where $1 = t _ { 1 } > \cdot \cdot \cdot > t _ { T } = 0$ , that gradually decrease in noise content. These points are functions of the $\mathbf { x }$ -predictions $\hat { \mathbf { x } } _ { 0 } ^ { t } : = \hat { \mathbf { x } } _ { \theta } ( \mathbf { z } _ { t } , \mathbf { c } )$ .
|
| 82 |
+
|
| 83 |
+
Classifier guidance [16] is a technique to improve sample quality while reducing diversity in conditional diffusion models using gradients from a pretrained model $p ( \mathbf { c } | \mathbf { z } _ { t } )$ during sampling. Classifierfree guidance [27] is an alternative technique that avoids this pretrained model by instead jointly training a single diffusion model on conditional and unconditional objectives via randomly dropping c during training (e.g. with $10 \%$ probability). Sampling is performed using the adjusted $\mathbf { x }$ -prediction $( \mathbf { z } _ { t } - \sigma \tilde { \epsilon } _ { \theta } ) / \alpha _ { t }$ , where
|
| 84 |
+
|
| 85 |
+
$$
|
| 86 |
+
\widetilde \epsilon _ { \theta } ( \mathbf { z } _ { t } , \mathbf { c } ) = w \epsilon _ { \theta } ( \mathbf { z } _ { t } , \mathbf { c } ) + ( 1 - w ) \epsilon _ { \theta } ( \mathbf { z } _ { t } ) .
|
| 87 |
+
$$
|
| 88 |
+
|
| 89 |
+
Here, $\epsilon _ { \theta } ( \mathbf { z } _ { t } , \mathbf { c } )$ and $\epsilon _ { \theta } ( { \mathbf { z } } _ { t } )$ are conditional and unconditional $\epsilon$ -predictions, given by $\boldsymbol { \epsilon } _ { \theta } : = ( \mathbf { z } _ { t } - $ $\alpha _ { t } \hat { \mathbf { x } } _ { \theta } ) / \sigma _ { t }$ , and $w$ is the guidance weight. Setting $w = 1$ disables classifier-free guidance, while increasing $w > 1$ strengthens the effect of guidance. Imagen depends critically on classifier-free guidance for effective text conditioning.
|
| 90 |
+
|
| 91 |
+
# 2.3 Large guidance weight samplers
|
| 92 |
+
|
| 93 |
+
We corroborate the results of recent text-guided diffusion work [16, 43, 56] and find that increasing the classifier-free guidance weight improves image-text alignment, but damages image fidelity producing highly saturated and unnatural images [27]. We find that this is due to a train-test mismatch arising from high guidance weights. At each sampling step $t$ , the $\mathbf { x }$ -prediction $\hat { \mathbf { x } } _ { 0 } ^ { t }$ must be within the same bounds as training data $\mathbf { x }$ , i.e. within $[ - 1 , 1 ]$ , but we find empirically that high guidance weights cause $\mathbf { x }$ -predictions to exceed these bounds. This is a train-test mismatch, and since the diffusion model is iteratively applied on its own output throughout sampling, the sampling process produces unnatural images and sometimes even diverges. To counter this problem, we investigate static thresholding and dynamic thresholding. See Appendix Fig. A.31 for reference implementation of the techniques and Appendix Fig. A.9 for visualizations of their effects.
|
| 94 |
+
|
| 95 |
+
Static thresholding: We refer to elementwise clipping the $\mathbf { x }$ -prediction to $[ - 1 , 1 ]$ as static thresholding. This method was in fact used but not emphasized in previous work [28], and to our knowledge its importance has not been investigated in the context of guided sampling. We discover that static thresholding is essential to sampling with large guidance weights and prevents generation of blank images. Nonetheless, static thresholding still results in over-saturated and less detailed images as the guidance weight further increases.
|
| 96 |
+
|
| 97 |
+
Dynamic thresholding: We introduce a new dynamic thresholding method: at each sampling step we set $s$ to a certain percentile absolute pixel value in $\hat { \mathbf { x } } _ { 0 } ^ { t }$ , and if $s > 1$ , then we threshold $\hat { \mathbf { x } } _ { 0 } ^ { t }$ to the range $[ - s , s ]$ and then divide by $s$ . Dynamic thresholding pushes saturated pixels (those near $^ { - 1 }$ and 1) inwards, thereby actively preventing pixels from saturation at each step. We find that dynamic thresholding results in significantly better photorealism as well as better image-text alignment, especially when using very large guidance weights.
|
| 98 |
+
|
| 99 |
+
# 2.4 Robust cascaded diffusion models
|
| 100 |
+
|
| 101 |
+
Imagen utilizes a pipeline of a base $6 4 \times 6 4$ model, and two text-conditional super-resolution diffusion models to upsample a $6 4 \times 6 4$ generated image into a $2 5 6 \times 2 5 6$ image, and then to $1 0 2 4 \times 1 0 2 4$ image. Cascaded diffusion models with noise conditioning augmentation [29] have been extremely effective in progressively generating high-fidelity images. Furthermore, making the super-resolution models aware of the amount of noise added, via noise level conditioning, significantly improves the sample quality and helps improving the robustness of the super-resolution models to handle artifacts generated by lower resolution models [29]. Imagen uses noise conditioning augmentation for both the super-resolution models. We find this to be a critical for generating high fidelity images.
|
| 102 |
+
|
| 103 |
+
Given a conditioning low-resolution image and augmentation level (a.k.a aug_level) (e.g., strength of Gaussian noise or blur), we corrupt the low-resolution image with the augmentation (corresponding to aug_level), and condition the diffusion model on aug_level. During training, aug_level is chosen randomly, while during inference, we sweep over its different values to find the best sample quality. In our case, we use Gaussian noise as a form of augmentation, and apply variance preserving Gaussian noise augmentation resembling the forward process used in diffusion models (Appendix A). The augmentation level is specified using aug_level $\in [ 0 , 1 ]$ . See Fig. A.32 for reference pseudocode.
|
| 104 |
+
|
| 105 |
+
# 2.5 Neural network architecture
|
| 106 |
+
|
| 107 |
+
Base model: We adapt the U-Net [60] architecture from [42] for our base $6 4 \times 6 4$ text-to-image diffusion model. The network is conditioned on text embeddings via a pooled embedding vector, added to the diffusion timestep embedding similar to the class embedding conditioning method used in [16, 29]. We further condition on the entire sequence of text embeddings by adding cross attention [59] over the text embeddings at multiple resolutions. We study various methods of text conditioning in Appendix D.3.1. Furthermore, we found Layer Normalization [2] for text embeddings in the attention and pooling layers to help considerably improve performance.
|
| 108 |
+
|
| 109 |
+
Super-resolution models: For $6 4 \times 6 4 2 5 6 \times 2 5 6$ super-resolution, we use the U-Net model adapted from [42, 61]. We make several modifications to this U-Net model for improving memory efficiency, inference time and convergence speed (our variant is $2 { - } 3 \mathbf { x }$ faster in steps/second over the U-Net used in [42, 61]). We call this variant Efficient $U$ -Net (See Appendix B.1 for more details and comparisons). Our $2 5 6 \times 2 5 6 \to 1 0 2 4 \times 1 0 2 4$ super-resolution model trains on $6 4 \times 6 4 2 5 6 \times 2 5 6$ crops of the $1 0 2 4 \times 1 0 2 4$ image. To facilitate this, we remove the self-attention layers, however we keep the text cross-attention layers which we found to be critical. During inference, the model receives the full $2 5 6 \times 2 5 6$ low-resolution images as inputs, and returns upsampled $1 0 2 4 \times 1 0 2 4$ images as outputs. Note that we use text cross attention for both our super-resolution models.
|
| 110 |
+
|
| 111 |
+
# 3 Evaluating Text-to-Image Models
|
| 112 |
+
|
| 113 |
+
The COCO [38] validation set is the standard benchmark for evaluating text-to-image models for both the supervised [85, 22] and the zero-shot setting [55, 43]. The key automated performance metrics used are FID [26] to measure image fidelity, and CLIP score [25, 51] to measure image-text alignment. Consistent with previous works, we report zero-shot FID-30K, for which 30K prompts are drawn randomly from the validation set, and the model samples generated on these prompts are compared with reference images from the full validation set. Since guidance weight is an important ingredient to control image quality and text alignment, we report most of our ablation results using trade-off (or pareto) curves between CLIP and FID scores across a range of guidance weights.
|
| 114 |
+
|
| 115 |
+
Both FID and CLIP scores have limitations, for example FID is not fully aligned with perceptual quality [44], and CLIP is ineffective at counting [51]. Due to these limitations, we use human evaluation to assess image quality and caption similarity, with ground truth reference caption-image pairs as a baseline. We use two experimental paradigms:
|
| 116 |
+
|
| 117 |
+
1. To probe image quality, the rater is asked to select between the model generation and reference image using the question: “Which image is more photorealistic (looks more real)?”. We report the percentage of times raters choose model generations over reference images (the preference rate). 2. To probe alignment, human raters are shown an image and a prompt and asked “Does the caption accurately describe the above image?”. They must respond with “yes”, “somewhat”, or “no”. These responses are scored as 100, 50, and 0, respectively. These ratings are obtained independently for model samples and reference images, and both are reported.
|
| 118 |
+
|
| 119 |
+
|
| 120 |
+
Figure 2: Non-cherry picked Imagen samples for different categories of prompts from DrawBench.
|
| 121 |
+
|
| 122 |
+
For both cases we use 200 randomly chosen image-caption pairs from the COCO validation set. Subjects were shown batches of 50 images. We also used interleaved “control" trials, and only include rater data from those who correctly answered at least $80 \%$ of the control questions. This netted 73 and 51 ratings per image for image quality and image-text alignment evaluations, respectively.
|
| 123 |
+
|
| 124 |
+
DrawBench: While COCO is a valuable benchmark, it is increasingly clear that it has a limited spectrum of prompts that do not readily provide insight into differences between models (e.g., see Sec. 4.2). Recent work by [10] proposed a new evaluation set called PaintSkills to systematically evaluate visual reasoning skills and social biases beyond COCO. With similar motivation, we introduce DrawBench, a comprehensive and challenging set of prompts that support the evaluation and comparison of text-to-image models. DrawBench contains 11 categories of prompts, testing different capabilities of models such as the ability to faithfully render different colors, numbers of objects, spatial relations, text in the scene, and unusual interactions between objects. Categories also include complex prompts, including long, intricate textual descriptions, rare words, and also misspelled prompts. We also include sets of prompts collected from DALL-E [55], Gary Marcus et al. [40] and Reddit. Across these 11 categories, DrawBench comprises 200 prompts in total, striking a good balance between the desire for a large, comprehensive dataset, and small enough that human evaluation remains feasible. (Appendix C provides a more detailed description of DrawBench. Fig. 2 shows example prompts from DrawBench with Imagen samples.)
|
| 125 |
+
|
| 126 |
+
We use DrawBench to directly compare different models. To this end, human raters are presented with two sets of images, one from Model A and one from Model B, each of which has 8 samples. Human raters are asked to compare Model A and Model B on sample fidelity and image-text alignment. They respond with one of three choices: Prefer Model A; Indifferent; or Prefer Model B.
|
| 127 |
+
|
| 128 |
+
# 4 Experiments
|
| 129 |
+
|
| 130 |
+
Section 4.1 describes training details, Sections 4.2 and 4.3 analyze results on MS-COCO and DrawBench, and Section 4.4 summarizes our ablation studies and key findings. For all experiments below, the images are fair random samples from Imagen with no post-processing or re-ranking.
|
| 131 |
+
|
| 132 |
+
# 4.1 Training details
|
| 133 |
+
|
| 134 |
+
Unless specified, we train a 2B parameter model for the $6 4 \times 6 4$ text-to-image synthesis, and $6 0 0 \mathbf { M }$ and 400M parameter models for $6 4 \times 6 4 2 5 6 \times 2 5 6$ and $2 5 6 \times 2 5 6 \to 1 0 2 4 \times 1 0 2 4$ for superresolution respectively. We use a batch size of 2048 and $2 . 5 \mathbf { M }$ training steps for all models. We use 256 TPU-v4 chips for our base $6 4 \times 6 4$ model, and 128 TPU-v4 chips for both super-resolution models. We do not find over-fitting to be an issue, and we believe further training might improve overall performance. We use Adafactor for our base $6 4 \times 6 4$ model, because initial comparisons with Adam suggested similar performance with much smaller memory footprint for Adafactor. For superresolution models, we use Adam as we found Adafactor to hurt model quality in our initial ablations. For classifier-free guidance, we joint-train unconditionally via zeroing out the text embeddings with $10 \%$ probability for all three models. We train on a combination of internal datasets, with $\approx 4 6 0 \mathrm { M }$ image-text pairs, and the publicly available LAION-400M dataset [64], with $\approx 4 0 0 { \mathrm { M } }$ image-text pairs. There are limitations in our training data, and we refer the reader to Section 6 for details. See Appendix F for more implementation details.
|
| 135 |
+
|
| 136 |
+
Table 1: MS-COCO $2 5 6 \times 2 5 6$ FID-30K. We use a guidance weight of 1.35 for our $6 4 \times 6 4$ model, and a guidance weight of 8.0 for our super-resolution model.
|
| 137 |
+
|
| 138 |
+
<table><tr><td>Model</td><td>FID-30K</td><td>Zero-shot FID-30K</td></tr><tr><td>AttnGAN [79]</td><td>35.49</td><td></td></tr><tr><td>DM-GAN [86]</td><td>32.64</td><td></td></tr><tr><td>DF-GAN[72]</td><td>21.42</td><td></td></tr><tr><td>DM-GAN + CL [81]</td><td>20.79</td><td></td></tr><tr><td>XMC-GAN [84]</td><td>9.33</td><td></td></tr><tr><td>LAFITE [85]</td><td>8.12</td><td></td></tr><tr><td>Make-A-Scene [22]</td><td>7.55</td><td></td></tr><tr><td>DALL-E [55]</td><td></td><td>17.89</td></tr><tr><td>LAFITE [85]</td><td></td><td>26.94</td></tr><tr><td>GLIDE [43]</td><td></td><td>12.24</td></tr><tr><td>DALL-E 2 [56]</td><td></td><td>10.39</td></tr><tr><td>Imagen (Our Work)</td><td></td><td>7.27</td></tr></table>
|
| 139 |
+
|
| 140 |
+
Table 2: COCO $2 5 6 \times 2 5 6$ human evaluation comparing model outputs and original images. For the bottom part (no people), we filter out prompts containing one of man, men, woman, women, person, people, child, adult, adults, boy, boys, girl, girls, guy, lady, ladies, someone, toddler, (sport) player, workers, spectators.
|
| 141 |
+
|
| 142 |
+
<table><tr><td>Model</td><td>Photorealism个</td><td>Alignment 个</td></tr><tr><td>Original</td><td></td><td></td></tr><tr><td>Original</td><td>50.0%</td><td>91.9 ± 0.42</td></tr><tr><td>Imagen</td><td>39.5 ± 0.75%</td><td>91.4 ± 0.44</td></tr><tr><td>No people</td><td></td><td></td></tr><tr><td>Original</td><td>50.0%</td><td>92.2 ± 0.54</td></tr><tr><td>Imagen</td><td>43.9 ± 1.01%</td><td>92.1 ± 0.55</td></tr></table>
|
| 143 |
+
|
| 144 |
+
# 4.2 Results on COCO
|
| 145 |
+
|
| 146 |
+
We evaluate Imagen on the COCO validation set using FID score, similar to [55, 43]. Table 1 displays the results. Imagen achieves state of the art zero-shot FID on COCO at 7.27, outperforming the concurrent work of DALL-E 2 [56] and even models trained on COCO. Table 2 reports the human evaluation to test image quality and alignment on the COCO validation set. We report results on the original COCO validation set, as well as a filtered version in which all reference data with people have been removed. For photorealism, Imagen achieves $3 9 . 2 \%$ preference rate indicating high image quality generation. On the set with no people, there is a boost in preference rate of Imagen to $4 3 . 6 \%$ , indicating Imagen’s limited ability to generate photorealistic people. On caption similarity, Imagen’s score is on-par with the original reference images, suggesting Imagen’s ability to generate images that align well with COCO captions.
|
| 147 |
+
|
| 148 |
+
# 4.3 Results on DrawBench
|
| 149 |
+
|
| 150 |
+
Using DrawBench, we compare Imagen with DALL-E 2 (the public version) [56], GLIDE [43], Latent Diffusion [59], and CLIP-guided VQ-GAN [12]. Fig. 3 shows the human evaluation results for pairwise comparison of Imagen with each of the three models. We report the percentage of time raters prefer Model A, Model B, or are indifferent for both image fidelity and image-text alignment. We aggregate the scores across all the categories and raters. We find the human raters to exceedingly prefer Imagen over all others models in both image-text alignment and image fidelity. We refer the reader to Appendix E for a more detailed category wise comparison and qualitative comparison.
|
| 151 |
+
|
| 152 |
+
# 4.4 Analysis of Imagen
|
| 153 |
+
|
| 154 |
+
For a detailed analysis of Imagen see Appendix D. Key findings are discussed in Fig. 4 and below.
|
| 155 |
+
|
| 156 |
+
Scaling text encoder size is extremely effective. We observe that scaling the size of the text encoder leads to consistent improvement in both image-text alignment and image fidelity. Imagen trained with our largest text encoder, T5-XXL (4.6B parameters), yields the best results (Fig. 4a).
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
Figure 3: Comparison between Imagen and DALL-E 2 [56], GLIDE [43], VQ-GAN $^ +$ CLIP [12] and Latent Diffusion [59] on DrawBench: User preference rates (with $9 5 \%$ confidence intervals) for image-text alignment and image fidelity.
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
Figure 4: Summary of some of the critical findings of Imagen with pareto curves sweeping over different guidance values. See Appendix D for more details.
|
| 163 |
+
|
| 164 |
+
Scaling text encoder size is more important than U-Net size. While scaling the size of the diffusion model U-Net improves sample quality, we found scaling the text encoder size to be significantly more impactful than the U-Net size (Fig. 4b).
|
| 165 |
+
|
| 166 |
+
Dynamic thresholding is critical. We show that dynamic thresholding results in samples with significantly better photorealism and alignment with text, over static or no thresholding, especially under the presence of large classifier-free guidance weights (Fig. 4c).
|
| 167 |
+
|
| 168 |
+
Human raters prefer T5-XXL over CLIP on DrawBench. The models trained with T5-XXL and CLIP text encoders perform similarly on the COCO validation set in terms of CLIP and FID scores. However, we find that human raters prefer T5-XXL over CLIP on DrawBench across all 11 categories.
|
| 169 |
+
|
| 170 |
+
Noise conditioning augmentation is critical. We show that training the super-resolution models with noise conditioning augmentation leads to better CLIP and FID scores. We also show that noise conditioning augmentation enables stronger text conditioning for the super-resolution model, resulting in improved CLIP and FID scores at higher guidance weights. Adding noise to the low-res image during inference along with the use of large guidance weights allows the super-resolution models to generate diverse upsampled outputs while removing artifacts from the low-res image.
|
| 171 |
+
|
| 172 |
+
Text conditioning method is critical. We observe that conditioning over the sequence of text embeddings with cross attention significantly outperforms simple mean or attention based pooling in both sample fidelity as well as image-text alignment.
|
| 173 |
+
|
| 174 |
+
Efficient U-Net is critical. Our Efficient U-Net implementation uses less memory, converges faster, and has better sample quality with faster inference.
|
| 175 |
+
|
| 176 |
+
# 5 Related Work
|
| 177 |
+
|
| 178 |
+
Diffusion models have seen wide success in image generation [28, 42, 62, 16, 29, 61], outperforming GANs in fidelity and diversity, without training instability and mode collapse issues [6, 16, 29]. Autoregressive models [39], GANs [79, 84], VQ-VAE Transformer-based methods [55, 22], and diffusion models have seen remarkable progress in text-to-image [59, 43, 59], including the concurrent
|
| 179 |
+
|
| 180 |
+
DALL-E 2 [56], which uses a diffusion prior on CLIP text latents and cascaded diffusion models to generate high resolution $1 0 2 4 \times 1 0 2 4$ images; we believe Imagen is much simpler, as Imagen does not need to learn a latent prior, yet achieves better results in both MS-COCO FID and human evaluation on DrawBench. GLIDE [43] also uses cascaded diffusion models for text-to-image, but we use large pretrained frozen language models, which we found to be instrumental to both image fidelity and image-text alignment. XMC-GAN [84] also uses BERT as a text encoder, but we scale to much larger text encoders and demonstrate the effectiveness thereof. The use of cascaded models is also popular throughout the literature [14, 41] and has been used with success in diffusion models to generate high resolution images [16, 29].
|
| 181 |
+
|
| 182 |
+
# 6 Conclusions, Limitations and Societal Impact
|
| 183 |
+
|
| 184 |
+
Imagen showcases the effectiveness of frozen large pretrained language models as text encoders for the text-to-image generation using diffusion models. Our observation that scaling the size of these language models have significantly more impact than scaling the U-Net size on overall performance encourages future research directions on exploring even bigger language models as text encoders. Furthermore, through Imagen we re-emphasize the importance of classifier-free guidance, and we introduce dynamic thresholding, which allows usage of much higher guidance weights than seen in previous works. With these novel components, Imagen produces $1 0 2 4 \times 1 0 2 4$ samples with unprecedented photorealism and alignment with text.
|
| 185 |
+
|
| 186 |
+
Our primary aim with Imagen is to advance research on generative methods, using text-to-image synthesis as a test bed. While end-user applications of generative methods remain largely out of scope, we recognize the potential downstream applications of this research are varied and may impact society in complex ways. On the one hand, generative models have a great potential to complement, extend, and augment human creativity [30]. Text-to-image generation models, in particular, have the potential to extend image-editing capabilities and lead to the development of new tools for creative practitioners. On the other hand, generative methods can be leveraged for malicious purposes, including harassment and misinformation spread [20], and raise many concerns regarding social and cultural exclusion and bias [70, 65, 71]. These considerations inform our decision to not to release code or a public demo. In future work we will explore a framework for responsible externalization that balances the value of external auditing with the risks of unrestricted open-access.
|
| 187 |
+
|
| 188 |
+
Another ethical challenge relates to the large scale data requirements of text-to-image models, which have have led researchers to rely heavily on large, mostly uncurated, web-scraped datasets. While this approach has enabled rapid algorithmic advances in recent years, datasets of this nature have been critiqued and contested along various ethical dimensions. For example, public and academic discourse regarding appropriate use of public data has raised concerns regarding data subject awareness and consent [24, 18, 63, 45]. Dataset audits have revealed these datasets tend to reflect social stereotypes, oppressive viewpoints, and derogatory, or otherwise harmful, associations to marginalized identity groups [46, 4]. Training text-to-image models on this data risks reproducing these associations and causing significant representational harm that would disproportionately impact individuals and communities already experiencing marginalization, discrimination and exclusion within society. As such, there are a multitude of data challenges that must be addressed before text-to-image models like Imagen can be safely integrated into user-facing applications. While we do not directly address these challenges in this work, an awareness of the limitations of our training data guide our decision not to release Imagen for public use. We strongly caution against the use text-to-image generation methods for any user-facing tools without close care and attention to the contents of the training dataset.
|
| 189 |
+
|
| 190 |
+
Imagen’s training data was drawn from several pre-existing datasets of image and English alt-text pairs. 400 million examples came from FIT400M, a cleaned version of the Alt-Text dataset [33, 31]. This data was filtered to removed noise and undesirable content, such as pornographic imagery and toxic language. However, a recent audit of another one of our data sources, LAION-400M [64], uncovered a wide range of inappropriate content including pornographic imagery, racist slurs, and harmful social stereotypes [4]. This finding informs our assessment that Imagen is not suitable for public use at this time and also demonstrates the value of rigorous dataset audits and comprehensive dataset documentation (e.g. [23, 47]) in informing consequent decisions about the model’s appropriate and safe use. Imagen also relies on text encoders trained on uncurated web-scale data, and thus inherits the social biases and limitations of large language models [5, 3, 52].
|
| 191 |
+
|
| 192 |
+
While we leave an in-depth empirical analysis of social and cultural biases encoded by Imagen to future work, our small scale internal assessments reveal several limitations that guide our decision not to release Imagen at this time. First, all generative models, including Imagen, Imagen, may run into danger of dropping modes of the data distribution, which may further compound the social consequence of dataset bias. Second, Imagen exhibits serious limitations when generating images depicting people. Our human evaluations found Imagen obtains significantly higher preference rates when evaluated on images that do not portray people, indicating a degradation in image fidelity. Finally, our preliminary assessment also suggests Imagen encodes several social biases and stereotypes, including an overall bias towards generating images of people with lighter skin tones and a tendency for images portraying different professions to align with Western gender stereotypes. Even when we focus generations away from people, our preliminary analysis indicates Imagen encodes a range of social and cultural biases when generating images of activities, events, and objects.
|
| 193 |
+
|
| 194 |
+
While there has been extensive work auditing image-to-text and image labeling models for forms of social bias (e.g. [8, 9, 71]), there has been comparatively less work on social bias evaluation methods for text-to-image models, with the recent exception of [10]. We believe this is a critical avenue for future research and we intend to explore benchmark evaluations for social and cultural bias in future work—for example, exploring whether it is possible to generalize the normalized pointwise mutual information metric [1] to the measurement of biases in image generation models. There is also a great need to develop a conceptual vocabulary around potential harms of text-to-image models that could guide the development of evaluation metrics and inform responsible model release. We aim to address these challenges in future work.
|
| 195 |
+
|
| 196 |
+
# 7 Acknowledgements
|
| 197 |
+
|
| 198 |
+
We give thanks to Ben Poole for reviewing our manuscript, early discussions, and providing many helpful comments and suggestions throughout the project. Special thanks to Kathy Meier-Hellstern, Austin Tarango, and Sarah Laszlo for helping us incorporate important responsible AI practices around this project. We appreciate valuable feedback and support from Elizabeth Adkison, Zoubin Ghahramani, Jeff Dean, Yonghui Wu, and Eli Collins. We are grateful to Tom Small for designing the Imagen watermark. We thank Jason Baldridge, Han Zhang, and Kevin Murphy for initial discussions and feedback. We acknowledge hard work and support from Fred Alcober, Hibaq Ali, Marian Croak, Aaron Donsbach, Tulsee Doshi, Toju Duke, Douglas Eck, Jason Freidenfelds, Brian Gabriel, Molly FitzMorris, David Ha, Philip Parham, Laura Pearce, Evan Rapoport, Lauren Skelly, Johnny Soraker, Negar Rostamzadeh, Vijay Vasudevan, Tris Warkentin, Jeremy Weinstein, and Hugh Williams for giving us advice along the project and assisting us with the publication process. We thank Victor Gomes and Erica Moreira for their consistent and critical help with TPU resource allocation. We also give thanks to Shekoofeh Azizi, Harris Chan, Chris A. Lee, and Nick Ma for volunteering a considerable amount of their time for testing out DrawBench. We thank Aditya Ramesh, Prafulla Dhariwal, and Alex Nichol for allowing us to use DALL-E 2 samples and providing us with GLIDE samples. We are thankful to Matthew Johnson and Roy Frostig for starting the JAX project and to the whole JAX team for building such a fantastic system for high-performance machine learning research. Special thanks to Durk Kingma, Jascha Sohl-Dickstein, Lucas Theis and the Toronto Brain team for helpful discussions and spending time Imagening!
|
| 199 |
+
|
| 200 |
+
# References
|
| 201 |
+
|
| 202 |
+
[1] Osman Aka, Ken Burke, Alex Bauerle, Christina Greer, and Margaret Mitchell. Measuring Model Biases in the Absence of Ground Truth. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, 2021.
|
| 203 |
+
[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
|
| 204 |
+
[3] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? . In Proceedings of FAccT 2021, 2021.
|
| 205 |
+
[4] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. In arXiv:2110.01963, 2021.
|
| 206 |
+
|
| 207 |
+
[5] Shikha Bordia and Samuel R. Bowman. Identifying and Reducing Gender Bias in Word-Level Language Models. In NAACL, 2017.
|
| 208 |
+
|
| 209 |
+
[6] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018.
|
| 210 |
+
|
| 211 |
+
[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In NeurIPS, 2020.
|
| 212 |
+
|
| 213 |
+
[8] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on Fairness, Accountability and Transparency, FAT 2018, 23-24 February 2018, New York, NY, USA, Proceedings of Machine Learning Research. PMLR, 2018.
|
| 214 |
+
|
| 215 |
+
[9] Kaylee Burns, Lisa Hendricks, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In European Conference on Computer Vision (ECCV), 2018.
|
| 216 |
+
|
| 217 |
+
[10] Jaemin Cho, Abhay Zala, and Mohit Bansal. Dall-eval: Probing the reasoning skills and social biases of text-to-image generative transformers. arxiv:2202.04053, 2022.
|
| 218 |
+
|
| 219 |
+
[11] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. PaLM: Scaling Language Modeling with Pathways. In arXiv:2001.08361, 2022.
|
| 220 |
+
|
| 221 |
+
[12] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. arXiv preprint arXiv:2204.08583, 2022.
|
| 222 |
+
|
| 223 |
+
[13] Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion schrödinger bridge with applications to score-based generative modeling. Advances in Neural Information Processing Systems, 34, 2021.
|
| 224 |
+
|
| 225 |
+
[14] Emily Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks. In NIPS, 2015.
|
| 226 |
+
|
| 227 |
+
[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL, 2019.
|
| 228 |
+
|
| 229 |
+
[16] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2022.
|
| 230 |
+
|
| 231 |
+
[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34, 2021.
|
| 232 |
+
|
| 233 |
+
[18] Dulhanty, Chris. Issues in Computer Vision Data Collection: Bias, Consent, and Label Taxonomy. In UWSpace, 2020.
|
| 234 |
+
|
| 235 |
+
[19] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021.
|
| 236 |
+
|
| 237 |
+
[20] Mary Anne Franks and Ari Ezra Waldman. Sex, lies and videotape: deep fakes and free speech delusions. Maryland Law Review, 78(4):892–898, 2019.
|
| 238 |
+
|
| 239 |
+
[21] Tsu-Jui Fu, Xin Eric Wang, and William Yang Wang. Language-Driven Image Style Transfer. arXiv preprint arXiv:2106.00178, 2021.
|
| 240 |
+
|
| 241 |
+
[22] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. arXiv preprint arXiv:2203.13131, 2022.
|
| 242 |
+
|
| 243 |
+
[23] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for Datasets. arXiv:1803.09010 [cs], March 2020.
|
| 244 |
+
|
| 245 |
+
[24] Adam Harvey and Jules LaPlace. MegaPixels: Origins and endpoints of biometric datasets "In the Wild". https://megapixels.cc, 2019.
|
| 246 |
+
|
| 247 |
+
[25] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.
|
| 248 |
+
|
| 249 |
+
[26] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, 2017.
|
| 250 |
+
|
| 251 |
+
[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021.
|
| 252 |
+
|
| 253 |
+
[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. NeurIPS, 2020.
|
| 254 |
+
|
| 255 |
+
[29] Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. JMLR, 2022.
|
| 256 |
+
|
| 257 |
+
[30] Rowan T. Hughes, Liming Zhu, and Tomasz Bednarz. Generative adversarial networks-enabled human-artificial intelligence collaborative applications for creative and design industries: A systematic review of current approaches and trends. Frontiers in artificial intelligence, 4, 2021.
|
| 258 |
+
|
| 259 |
+
[31] Aashi Jain, Mandy Guo, Krishna Srinivasan, Ting Chen, Sneha Kudugunta, Chao Jia, Yinfei Yang, and Jason Baldridge. MURAL: Multimodal, multitask representations across languages. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 3449– 3463, Punta Cana, Dominican Republic, November 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.findings-emnlp.293. URL https://aclanthology.org/ 2021.findings-emnlp.293.
|
| 260 |
+
|
| 261 |
+
[32] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, YunHsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904–4916. PMLR, 2021.
|
| 262 |
+
|
| 263 |
+
[33] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, YunHsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.
|
| 264 |
+
|
| 265 |
+
[34] Zahra Kadkhodaie and Eero P Simoncelli. Solving linear inverse problems using the prior implicit in a denoiser. arXiv preprint arXiv:2007.13640, 2020.
|
| 266 |
+
|
| 267 |
+
[35] Zahra Kadkhodaie and Eero P Simoncelli. Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems, 34, 2021.
|
| 268 |
+
|
| 269 |
+
[36] Gwanghyun Kim and Jong Chul Ye. Diffusionclip: Text-guided image manipulation using diffusion models. arXiv preprint arXiv:2110.02711, 2021.
|
| 270 |
+
[37] Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021.
|
| 271 |
+
[38] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, Lawrence Zitnick, and Piotr Dollár. Microsoft COCO: Common Objects in Context. In ECCV, 2014.
|
| 272 |
+
[39] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating Images from Captions with Attention. In ICLR, 2016.
|
| 273 |
+
[40] Gary Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of DALL-E 2. In arXiv:2204.13807, 2022.
|
| 274 |
+
[41] Jacob Menick and Nal Kalchbrenner. Generating High Fidelity Images with Subscale Pixel Networks and Multidimensional Upscaling. In ICLR, 2019.
|
| 275 |
+
[42] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. arXiv preprint arXiv:2102.09672, 2021.
|
| 276 |
+
[43] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Bob McGrew Pamela Mishkin, Ilya Sutskever, and Mark Chen. GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models. In arXiv:2112.10741, 2021.
|
| 277 |
+
[44] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On Aliased Resizing and Surprising Subtleties in GAN Evaluation. In CVPR, 2022.
|
| 278 |
+
[45] Amandalynne Paullada, Inioluwa Deborah Raji, Emily M. Bender, Emily Denton, and Alex Hanna. Data and its (dis)contents: A survey of dataset development and use in machine learning research. Patterns, 2(11):100336, 2021.
|
| 279 |
+
[46] Vinay Uday Prabhu and Abeba Birhane. Large image datasets: A pyrrhic win for computer vision? arXiv:2006.16923, 2020.
|
| 280 |
+
[47] Mahima Pushkarna, Andrew Zaldivar, and Oddur Kjartansson. Data cards: Purposeful and transparent dataset documentation for responsible ai. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2022.
|
| 281 |
+
[48] Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. Learning to Generate Reviews and Discovering Sentiment. In arXiv:1704.01444, 2017.
|
| 282 |
+
[49] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving Language Understanding by Generative Pre-Training. In preprint, 2018.
|
| 283 |
+
[50] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language Models are Unsupervised Multitask Learners. In preprint, 2019.
|
| 284 |
+
[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021.
|
| 285 |
+
[52] Jack Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George Driessche, Lisa Hendricks, Maribeth Rauh, Po-Sen Huang, and Geoffrey Irving. Scaling language models: Methods, analysis & insights from training gopher. arXiv:2112.11446, 2021.
|
| 286 |
+
[53] Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, and Douglas Eck. Online and Linear-Time Attention by Enforcing Monotonic Alignments. In ICML, 2017.
|
| 287 |
+
|
| 288 |
+
[54] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. JMLR, 21(140), 2020.
|
| 289 |
+
|
| 290 |
+
[55] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-Shot Text-to-Image Generation. In ICML, 2021.
|
| 291 |
+
|
| 292 |
+
[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical Text-Conditional Image Generation with CLIP Latents. In arXiv, 2022.
|
| 293 |
+
|
| 294 |
+
[57] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. arXiv preprint arXiv:1906.00446, 2019.
|
| 295 |
+
|
| 296 |
+
[58] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060–1069. PMLR, 2016.
|
| 297 |
+
|
| 298 |
+
[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022.
|
| 299 |
+
|
| 300 |
+
[60] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015.
|
| 301 |
+
|
| 302 |
+
[61] Chitwan Saharia, William Chan, Huiwen Chang, Chris A. Lee, Jonathan Ho, Tim Salimans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-Image Diffusion Models. In arXiv:2111.05826, 2021.
|
| 303 |
+
|
| 304 |
+
[62] Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021.
|
| 305 |
+
|
| 306 |
+
[63] Morgan Klaus Scheuerman, Emily L. Denton, and A. Hanna. Do datasets have politics? disciplinary values in computer vision dataset development. Proceedings of the ACM on Human-Computer Interaction, 5:1 – 37, 2021.
|
| 307 |
+
|
| 308 |
+
[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.
|
| 309 |
+
|
| 310 |
+
[65] Lucas Sequeira, Bruno Moreschi, Amanda Jurno, and Vinicius Arruda dos Santos. Which faces can AI generate? Normativity, whiteness and lack of diversity in This Person Does Not Exist. In CVPR Workshop Beyond Fairness: Towards a Just, Equitable, and Accountable Computer Vision, 2021.
|
| 311 |
+
|
| 312 |
+
[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256–2265. PMLR, 2015.
|
| 313 |
+
|
| 314 |
+
[67] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.
|
| 315 |
+
|
| 316 |
+
[68] Yang Song and Stefano Ermon. Generative Modeling by Estimating Gradients of the Data Distribution. NeurIPS, 2019.
|
| 317 |
+
|
| 318 |
+
[69] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021.
|
| 319 |
+
|
| 320 |
+
[70] Ramya Srinivasan and Kanji Uchino. Biases in generative art: A causal look from the lens of art history. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, page 41–51, 2021.
|
| 321 |
+
|
| 322 |
+
[71] Ryan Steed and Aylin Caliskan. Image representations learned with unsupervised pre-training contain human-like biases. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’21, page 701–713. Association for Computing Machinery, 2021.
|
| 323 |
+
[72] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020.
|
| 324 |
+
[73] Belinda Tzen and Maxim Raginsky. Neural Stochastic Differential Equations: Deep Latent Gaussian Models in the Diffusion Limit. In arXiv:1905.09883, 2019.
|
| 325 |
+
[74] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.
|
| 326 |
+
[75] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Computation, 23(7):1661–1674, 2011.
|
| 327 |
+
[76] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798–8807, 2018.
|
| 328 |
+
[77] Jason Weston, Samy Bengio, and Nicolas Usunier. Wsabie: Scaling up to large vocabulary image annotation. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011.
|
| 329 |
+
[78] Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. arXiv preprint arXiv:2112.02475, 2021.
|
| 330 |
+
[79] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks. In CVPR, 2018.
|
| 331 |
+
[80] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316–1324, 2018.
|
| 332 |
+
[81] Hui Ye, Xiulong Yang, Martin Takac, Rajshekhar Sunderraman, and Shihao Ji. Improving text-to-image synthesis using contrastive learning. arXiv preprint arXiv:2107.02423, 2021.
|
| 333 |
+
[82] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021.
|
| 334 |
+
[83] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022.
|
| 335 |
+
[84] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-Modal Contrastive Learning for Text-to-Image Generation. In CVPR, 2021.
|
| 336 |
+
[85] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. arXiv preprint arXiv:2111.13792, 2021.
|
| 337 |
+
[86] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802–5810, 2019.
|
| 338 |
+
[87] Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In ICCV, 2015.
|
| 339 |
+
|
| 340 |
+
# Checklist
|
| 341 |
+
|
| 342 |
+
1. For all authors...
|
| 343 |
+
|
| 344 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 345 |
+
(b) Did you describe the limitations of your work? [Yes] See Section 6.
|
| 346 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] See Section 6.
|
| 347 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 348 |
+
|
| 349 |
+
2. If you are including theoretical results...
|
| 350 |
+
|
| 351 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 352 |
+
|
| 353 |
+
3. If you ran experiments...
|
| 354 |
+
|
| 355 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No]
|
| 356 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes]
|
| 357 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No]
|
| 358 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes]
|
| 359 |
+
|
| 360 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 361 |
+
|
| 362 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 363 |
+
(b) Did you mention the license of the assets? [Yes] We describe the license of the public data we use in Section 6.
|
| 364 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] We include DrawBench prompts in the supplemental material.
|
| 365 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [Yes] We discuss the lack of consent from data subjects in web-scraped data in Section 6.
|
| 366 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] See Section 6.
|
| 367 |
+
|
| 368 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 369 |
+
|
| 370 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [Yes]
|
| 371 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [No]
|
| 372 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [Yes]
|
parse/dev/0YXmOFLb1wQ/0YXmOFLb1wQ_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/1W8UwXAQubL/1W8UwXAQubL_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf.md
ADDED
|
@@ -0,0 +1,523 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Seeing Differently, Acting Similarly: Heterogeneously Observable Imitation Learning
|
| 2 |
+
|
| 3 |
+
Anonymous Author(s)
|
| 4 |
+
Affiliation
|
| 5 |
+
Address
|
| 6 |
+
email
|
| 7 |
+
|
| 8 |
+
# Abstract
|
| 9 |
+
|
| 10 |
+
1 In many real-world imitation learning tasks, the demonstrator and the learner have
|
| 11 |
+
2 to act under totally different observation spaces. This situation brings significant
|
| 12 |
+
3 obstacles to existing imitation learning approaches, since most of them learn poli
|
| 13 |
+
4 cies under homogeneous observation spaces. On the other hand, previous studies
|
| 14 |
+
5 under different observation spaces have strong assumptions that these two obser
|
| 15 |
+
6 vation spaces coexist during the entire learning process. However, in reality, the
|
| 16 |
+
7 observation coexistence will be limited due to the high cost of acquiring expert
|
| 17 |
+
8 observations. In this work, we study this challenging problem with limited observa
|
| 18 |
+
9 tion coexistence under heterogeneous observations: Heterogeneously Observable
|
| 19 |
+
10 Imitation Learning (HOIL). We identify two underlying issues in HOIL, i.e., the
|
| 20 |
+
11 dynamics mismatch and the support mismatch, and further propose the Impor
|
| 21 |
+
12 tance Weighting with REjection (IWRE) algorithm based on importance-weighting
|
| 22 |
+
13 and learning with rejection to solve HOIL problems. Experimental results show
|
| 23 |
+
14 that IWRE can successfully solve various HOIL tasks, including the challenging
|
| 24 |
+
15 tasks of transforming the vision-based demonstrations to random access memory
|
| 25 |
+
16 (RAM)-based policies in the Atari domain, even with limited visual observations.
|
| 26 |
+
|
| 27 |
+
# 17 1 Introduction
|
| 28 |
+
|
| 29 |
+
18 Imitation Learning (IL) studies how to learn a good policy by imitating the given expert demonstra
|
| 30 |
+
19 tions [16, 1], and has achieved great success in many domains such as autonomous driving [8], video
|
| 31 |
+
20 games [7], and continuous control [19]. In real-world IL applications, the expert and the learner
|
| 32 |
+
21 usually have their own observations of the same underlying states from the environment. For example,
|
| 33 |
+
22 in Figure 1, an autonomous agent is learning to drive by imitating a human expert. The expert takes
|
| 34 |
+
23 her actions mainly based on auditory and visual observations, which are familiar to human beings.
|
| 35 |
+
24 However, the learning agent does not necessarily use the same way to observe: it can utilize more
|
| 36 |
+
25 machine-capable sensors such as a LiDAR, radar, and bird-eye view (BEV) map to generate its
|
| 37 |
+
26 observations [20]. The key features behind this example are two-fold: First, both the expert and
|
| 38 |
+
27 the learner have their totally different observations of the same state of the environment. Thus they
|
| 39 |
+
28 essentially have to choose the same action if acting optimally. Second, the observation space of the
|
| 40 |
+
29 expert is often of high cost for the learner to utilize [6, 10]. We call this problem Heterogeneously
|
| 41 |
+
30 Observable Imitation Learning (HOIL).
|
| 42 |
+
31 There are two lines of research studying the related problems. The first line relates to domain
|
| 43 |
+
32 adaptation: the observation space of the expert and the learner are the homogeneous, while some
|
| 44 |
+
33 typical mismatches of distributions could exist: morphological mismatch, viewpoint mismatch, and
|
| 45 |
+
34 dynamics mismatch [30, 17, 26]. However, these approaches are invalid when the observation spaces
|
| 46 |
+
35 for experts and learners are completely different as in HOIL.
|
| 47 |
+
36 The second line studied IL under different observations similar to HOIL, and some representative
|
| 48 |
+
37 works include Partially Observable Imitation Learning (POIL) [14, 36] and Learning by Cheating
|
| 49 |
+
38 (LBC) [8], as depicted in Figure 2. Both POIL and LBC assume that the expert’s observations can
|
| 50 |
+
39 be easily accessed by the learner without any budget limit. However in practice, different from the
|
| 51 |
+
40 learner observations, the access to expert’s observations might be of high cost and invasive [6, 10],
|
| 52 |
+
41 hindering the wide application of these methods.
|
| 53 |
+
42 In this paper, we initialize the study of the HOIL problem. We propose a learning process across
|
| 54 |
+
43 observation spaces of experts and learners for solving this problem, and analyze the underlying issues
|
| 55 |
+
44 of HOIL, i.e., the dynamics mismatch and the support mismatch. To tackle both two issues, we resort
|
| 56 |
+
45 to the techniques of importance-weighting [12] and learning with rejection [9, 15] for active querying
|
| 57 |
+
46 to propose the Importance Weighting with REjection (IWRE) approach. We evaluate the effectiveness
|
| 58 |
+
47 of the IWRE algorithm in continuous control tasks of MuJoCo [33], and the challenging tasks of
|
| 59 |
+
48 learning random access memory (RAM)-based policies given vision-based expert demonstrations
|
| 60 |
+
49 in Atari [3] games. The results demonstrate that IWRE can significantly outperform existing IL
|
| 61 |
+
50 algorithms in HOIL tasks, with limited access to expert observations.
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
Figure 1: Autonomous driving: an example of the HOIL problem. Figures 1, 2 and 3 include some illustrations and pictures from the Internet (source: www.vecteezy.com).
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
Figure 2: Comparisons of different IL processes under different observation spaces. The targets are all to learn $\pi _ { 2 }$ based on the second observation space with an auxiliary policy $\pi _ { 1 }$ from corresponding roll-out data $\widetilde { \tau }$ and $\overline { { \mathcal { T } } }$ . (a) POIL mainly emphasized that the expert can view full observations, while the observations for the learner are partial. (b) LBC assumed that the expert’s observations contain more privileged information than the learner’s. Both POIL and LBC can observe expert’s observations all along. (c) HOIL limits the amount of expert’s observations.
|
| 68 |
+
|
| 69 |
+
# 51 2 Related Work
|
| 70 |
+
|
| 71 |
+
52 Domain-Shifted IL. For the standard IL process, where the learner and the expert share the same
|
| 72 |
+
53 observation space, current state-of-the-art methods tend to learn the policy in an adversarial style [7],
|
| 73 |
+
54 like GAIL [16]. When considering the domain mismatch problem, i.e., Domain-Shifted IL (DSIL),
|
| 74 |
+
55 the research aims at addressing the static distributional shift of the optimal policies resulted from
|
| 75 |
+
56 the environmental differences but still under homogeneous observation spaces. Stadie et al. [30],
|
| 76 |
+
57 Sermanet et al. [29], and Liu et al. [23] studied the situation where the demonstrations are in view
|
| 77 |
+
58 of a third person. Kim et al. [19] and Kim et al. [18] addressed the IL problem with morphological
|
| 78 |
+
59 mismatch between the expert’s and learner’s environment. Stadie et al. [30], Tirinzoni et al. [32], and
|
| 79 |
+
60 Desai et al. [11] focused on the calibration for the mismatch between simulators and the real world
|
| 80 |
+
61 through some transfer learning styles. There are two major differences between HOIL and DSIL:
|
| 81 |
+
62 One is that HOIL considers heterogeneous observation spaces instead of homogeneous ones; another
|
| 82 |
+
63 is that without observation heterogeneity, DSIL can directly align two fixed domains, which may
|
| 83 |
+
64 not be realistic for solving HOIL when two observation spaces are totally different. Thus HOIL is a
|
| 84 |
+
65 significantly more challenging problem than DSIL. Besides, Chen et al. [8] learned a vision-based
|
| 85 |
+
66 agent from a privileged expert. But it can obtain expert’s observations throughout the whole learning
|
| 86 |
+
67 process, so it cannot handle the problem of the support mismatch under HOIL.
|
| 87 |
+
68 POMDP. The problem of POMDPs, in which only partial observations are available for the agent(s),
|
| 88 |
+
69 has been studied in the context of multi-agent [25, 36] and imitation learning [14, 36] problems.
|
| 89 |
+
70 But distinct from HOIL, in a POMDP, the learner only have partial observations and share a same
|
| 90 |
+
71 underlying observation space with the expert, which would become an obstacle for them to make
|
| 91 |
+
72 decisions correctly. For example, Warrington et al. [36] assumed that the observation of the learner
|
| 92 |
+
73 is partial than that of the expert. Instead, in HOIL, expert’s and learner’s observations are totally
|
| 93 |
+
74 different from each other, while the learner’s observations are not belong to a part of the expert’s. For
|
| 94 |
+
75 HOIL, the main challenge is to deal with the mismatches between the observation spaces, especially
|
| 95 |
+
76 when the access to expert’s observations is strictly limited.
|
| 96 |
+
|
| 97 |
+
# 3 The HOIL Problem
|
| 98 |
+
|
| 99 |
+
In this section, we first give a formal definition of the HOIL setting, and then introduce the learning process for solving the HOIL problem.
|
| 100 |
+
|
| 101 |
+
# 3.1 Setting Definition
|
| 102 |
+
|
| 103 |
+
A HOIL problem is defined within a Markov decision process with mutiple observation spaces, i.e., $\langle \mathcal { S } , \{ \mathcal { O } \} , \mathcal { A } , \mathcal { P } , \gamma \rangle$ , where $s$ denotes the state space, $\{ \mathcal { O } \}$ denotes a set of observation spaces, $\mathcal { A }$ denotes the action space, $\mathcal { P } : \mathcal { S } \times \mathcal { A } \times \mathcal { S } \mathbb { R }$ denotes the transition probability distribution of the state and action, and $\gamma \in ( 0 , 1 ]$ denotes the discount factor. Furthermore, a policy $\pi$ over an observation space $\mathcal { O }$ is defined as a function mapping from $\mathcal { O }$ to $\mathcal { A }$ , and we denote by $\Pi _ { \mathcal { O } }$ the set of all policies over $\mathcal { O }$ . In HOIL, both the expert and the learner have their own observation spaces, which are denoted as $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ respectively. Both $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ are assumed to be produced by two bijective mappings $f _ { \mathrm { E } } : S \mathcal { O } _ { \mathrm { E } }$ , $f _ { \mathrm { L } } : S \mathcal { O } _ { \mathrm { L } }$ , which are unknown functions mapping the underlying true states to the observations. It is obvious to see that by this assumption, any policy over $\mathcal { O } _ { \mathrm { E } }$ has a unique correspondence over $\mathcal { O } _ { \mathrm { L } }$ . This makes HOIL possible since the target of HOIL is to find the corresponding policy of the expert policy under $\mathcal { O } _ { \mathrm { L } }$ .
|
| 104 |
+
|
| 105 |
+
92 A state-action pair $( s , a )$ , denoted by $x$ , is called an instance. Also, a trajectory $\mathcal { T } = \{ x _ { i } \} , i \in [ m ]$
|
| 106 |
+
93 is a set of $m$ instances. For each observation space, $\boldsymbol { \widetilde { x } } \in \mathcal { \widetilde { T } } \subseteq \mathcal { O } _ { \mathrm { E } } \times \mathcal { A }$ and $\overline { { x } } \in \overline { { \mathcal { T } } } \subseteq \mathcal { O } _ { \mathrm { L } } \times \mathcal { A }$ ,
|
| 107 |
+
94 where $\mathcal { O } _ { \mathrm { E } } = f _ { \mathrm { E } } ( \boldsymbol { S } )$ and $\mathcal { O } _ { \mathrm { L } } = f _ { \mathrm { L } } ( \mathcal { S } )$ e. Furthermore, we define the occupancy measure of a policy $\pi$
|
| 108 |
+
95 under the state space $s$ as $\rho _ { \pi } : \mathcal { S } \times \mathcal { A } \mathbb { R }$ such that $\begin{array} { r } { \rho _ { \pi } ( x ) = \pi ( a | o ) \mathrm { P r } ( o | s ) \sum _ { t = 0 } ^ { \infty } \gamma ^ { t } \mathrm { P r } ( s _ { t } = s | \pi ) } \end{array}$ .
|
| 109 |
+
96 Under HOIL, the learner accesses the expert demonstrations $\widetilde { \mathcal { T } } _ { \pi _ { \mathrm { E } } }$ , a set of instances sampled from $\rho _ { \pi _ { \mathrm { E } } }$ .
|
| 110 |
+
97 The goal of HOIL is to learn a policy $\hat { \pi }$ as the corresponding policy of $\pi _ { \mathrm { E } }$ over $\mathcal { O } _ { \mathrm { L } }$ . If $\mathcal { O } _ { \mathrm { E } } = \mathcal { O } _ { \mathrm { L } }$ ,
|
| 111 |
+
98 HOIL degenerates to standard $\mathrm { I L }$ . GAIL [16] is one of the state-of-the-art $\mathrm { I L }$ approaches under this
|
| 112 |
+
99 situation, which tries to minimize the divergence between the learner’s and the expert’s occupancy
|
| 113 |
+
100 measures $d ( \rho _ { \hat { \pi } } , \rho _ { \pi _ { \mathrm { E } } } )$ . The objective of GAIL is
|
| 114 |
+
|
| 115 |
+
$$
|
| 116 |
+
\operatorname* { m i n } _ { \hat { \pi } } \operatorname* { m a x } _ { w } \mathbb { E } _ { { x } \sim { \rho } _ { \pi _ { \mathrm { E } } } } [ \log D _ { w } ( \widetilde { x } ) ] + \mathbb { E } _ { { x } \sim { \rho } _ { \hat { \pi } } } [ \log ( 1 - D _ { w } ( \widetilde { x } ) ) ] - \mathbb { H } ( \hat { \pi } ) ,
|
| 117 |
+
$$
|
| 118 |
+
|
| 119 |
+
101 where $\mathbb { H } ( \hat { \pi } )$ is the causal entropy performed as a regularization term, and $D _ { w } : { \mathcal { O } } _ { \mathrm { E } } \times A \to [ 0 , 1 ]$ is
|
| 120 |
+
102 the discriminator of $\pi _ { \mathrm { E } }$ and $\hat { \pi }$ . GAIL solved Equation (1) by alternatively taking a gradient ascent
|
| 121 |
+
103 step to train the discriminator $D _ { w }$ , and a minimization step to learn policy $\hat { \pi }$ based on an off-the-shelf
|
| 122 |
+
104 RL algorithm with the pseudo reward $- \log D _ { w } ( \widetilde { x } )$ .
|
| 123 |
+
|
| 124 |
+
# 3.2 The Learning Process for Solving HOIL
|
| 125 |
+
|
| 126 |
+
106 In HOIL, we need to cope with the absence of the learner’s observations in demonstrations and the
|
| 127 |
+
107 high cost of collecting the expert’s observations while learning. So we introduce a learning process
|
| 128 |
+
108 with pretraining across two different observation spaces for solving HOIL, as abstracted in Figure 3.
|
| 129 |
+
109 Pretraining. Same to LBC [8], we assume that we can obtain an auxiliary policy $\pi _ { 1 }$ based on $\mathcal { O } _ { \mathrm { E } }$ at
|
| 130 |
+
110 the beginning. $\pi _ { 1 }$ can be directly provided by any sources, or trained by GAIL or behavior cloning
|
| 131 |
+
111 as did in LBC. Besides, we use this $\pi _ { 1 }$ to sample some data $\mathcal { T } _ { \pi _ { 1 } }$ , which contain both observation
|
| 132 |
+
112 under $\mathcal { O } _ { \mathrm { E } }$ (i.e., $\widetilde { \mathcal { T } } _ { \pi _ { 1 } } .$ ) and $\mathcal { O } _ { \mathrm { L } }$ (i.e., $\overline { { \mathcal { T } } } _ { \pi _ { 1 } }$ ), in order to connect these two different observation spaces.
|
| 133 |
+
113 We name ${ \mathcal { T } } _ { \pi _ { 1 } } = \{ { \widetilde { \mathcal { T } } _ { \pi _ { 1 } } , \overline { { { \mathcal { T } } } } _ { \pi _ { 1 } } } \}$ the initial data.
|
| 134 |
+
114 Training. Here we learn a policy $\pi _ { 2 }$ from the initial data $\overline { { \mathcal { T } } } _ { \pi _ { 1 } }$ and the collected data $\overline { { \mathcal { T } } } _ { \pi _ { 2 } }$ , under
|
| 135 |
+
115 $\mathcal { O } _ { \mathrm { L } }$ only. Besides, the learner is allowed for some operation of observation coexistence (OC): At
|
| 136 |
+
116 some steps of learning, besides the observations $\mathcal { O } _ { \mathrm { L } }$ , the learner could also request $\widetilde { \tau } _ { \pi _ { 2 } }$ from the
|
| 137 |
+
117 corresponding observations $\mathcal { O } _ { \mathrm { E } }$ (e.g., from the human-understandable sensors). The final objective of
|
| 138 |
+
118 HOIL is to learn a good policy $\pi _ { 2 }$ under $\mathcal { O } _ { \mathrm { L } }$ .
|
| 139 |
+
119 In practical applications, the auxiliary policy $\pi _ { 1 }$ can also come from simulation training or direct
|
| 140 |
+
120 imitation. But since $\pi _ { 1 }$ is additionally provided, it is more practical to consider $\pi _ { 1 }$ as a non-optimal
|
| 141 |
+
121 policy. During training, OC is an essential operation for solving HOIL, which helps the learner
|
| 142 |
+
122 address the issues of the dynamics mismatch and the support mismatch (especially the latter one).
|
| 143 |
+
123 Also, in reality, we do not need an oracle for actions, which still needs OC for obtaining expert
|
| 144 |
+
124 observations first, as in many active querying research [4, 8], so its cost will be relatively lower.
|
| 145 |
+
125 Besides, the related work [8] also required an initialized policy $\pi _ { 1 }$ to solve their problem, which act
|
| 146 |
+
126 as a teacher under privileged $\mathcal { O } _ { \mathrm { E } }$ in the pretraining and then learned a vision-based student from the
|
| 147 |
+
127 guidance of the teacher under both $\mathcal { O } _ { \mathrm { L } }$ and $\mathcal { O } _ { \mathrm { E } }$ . Their setting can be viewed as a variety of HOIL
|
| 148 |
+
128 with optimal $\pi _ { 1 }$ , unlimited $\mathcal { O } _ { \mathrm { E } }$ , and unlimited OC operations, so HOIL is actually a more practical
|
| 149 |
+
129 learning framework.
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
Figure 3: Illustration of a learning process across two different observation spaces for solving HOIL. $\pi _ { 1 }$ is an auxiliary policy that additionally provided.
|
| 153 |
+
|
| 154 |
+
# 4 Imitation Learning with Importance-Weighting and Rejection
|
| 155 |
+
|
| 156 |
+
In HOIL, the access frequency to $\mathcal { O } _ { \mathrm { E } }$ is strictly limited, so it is unrealistic to learn $\pi _ { 2 }$ in a Dataset Aggregation (DAgger) style [27] as in LBC. Therefore, we resort to learning $\pi _ { 2 }$ with a learned reward function by inverse reinforcement learning [1] in an adversarial learning style [16, 13].
|
| 157 |
+
|
| 158 |
+
134 In addition, both $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ are assumed to share the same latent state space $s$ as introduced in
|
| 159 |
+
135 Section 3.1, so the following analysis will be based on $s$ , while the algorithm will handle the problem
|
| 160 |
+
136 based on $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ specifically.
|
| 161 |
+
|
| 162 |
+
# 4.1 Dynamics Mismatch and Importance-Weighting
|
| 163 |
+
|
| 164 |
+
138 To analyze the learning process, we let $\rho _ { \pi _ { \mathrm { E } } } , \rho _ { \pi _ { 1 } }$ , and $\rho _ { \pi _ { 2 } }$ be the occupancy measure distributions
|
| 165 |
+
139 of the expert demonstrations, the initial data, and the data during training respectively. Since we
|
| 166 |
+
140 need to consider the sub-optimality of $\pi _ { 1 } , \rho _ { \pi _ { 1 } }$ should be a mixture distribution of the expert $\rho _ { \pi _ { \mathrm { E } } }$ and
|
| 167 |
+
141 non-expert $\rho _ { \pi _ { \mathrm { N E } } }$ , i.e., there exists some $\delta \in ( 0 , 1 )$ such that
|
| 168 |
+
|
| 169 |
+
$$
|
| 170 |
+
\rho _ { \pi _ { 1 } } = \delta \rho _ { \pi _ { \mathrm { E } } } + ( 1 - \delta ) \rho _ { \pi _ { \mathrm { N E } } } ,
|
| 171 |
+
$$
|
| 172 |
+
|
| 173 |
+
142 as depicted in Figure 4a. During training, the original objective of $\pi _ { 2 }$ is to imitate $\pi _ { \mathrm { E } }$ through demonstrations. To this end, the original objective of reward function 143 $D _ { w _ { 2 } }$ for $\pi _ { 2 }$ is to optimize
|
| 174 |
+
|
| 175 |
+
$$
|
| 176 |
+
\operatorname* { m a x } _ { w _ { 2 } } \mathbb { E } _ { { x } \sim \rho _ { \pi _ { 2 } } } [ \log D _ { w _ { 2 } } ( \overline { { x } } ) ] + \mathbb { E } _ { { x } \sim \rho _ { \pi _ { \mathrm { E } } } } [ \log ( 1 - D _ { w _ { 2 } } ( \overline { { x } } ) ) ] .
|
| 177 |
+
$$
|
| 178 |
+
|
| 179 |
+
144 But the expert demonstrations are only available under $\mathcal { O } _ { \mathrm { E } }$ . While during training, we can only utilize
|
| 180 |
+
145 the initial data $\overline { { \mathcal { T } } } _ { \pi _ { 1 } } \sim \rho _ { \pi _ { 1 } }$ to learn $\pi _ { 2 }$ and $D _ { w _ { 2 } }$ . Besides, as $\pi _ { 1 }$ is sub-optimal, directly imitating ${ \overline { { \mathcal T } } } _ { \pi _ { 1 } }$
|
| 181 |
+
146 could reduce the performance of the optimal $\pi _ { 2 }$ to that of $\pi _ { 1 }$ . So we use the importance-weighting to
|
| 182 |
+
147 calibrate this dynamics mismatch, i.e.,
|
| 183 |
+
|
| 184 |
+
$$
|
| 185 |
+
\operatorname* { m a x } _ { w _ { 2 } } \mathcal { L } ( D _ { w _ { 2 } } ) = \mathbb { E } _ { x \sim \rho _ { \pi _ { 2 } } } [ \log D _ { w _ { 2 } } ( \overline { { x } } ) ] + \mathbb { E } _ { x \sim \rho _ { \pi _ { 1 } } } [ \alpha ( x ) \log ( 1 - D _ { w _ { 2 } } ( \overline { { x } } ) ) ] ,
|
| 186 |
+
$$
|
| 187 |
+
|
| 188 |
+
|
| 189 |
+
Figure 4: The comparisons among the distributions of expert demonstrations $\rho _ { \pi _ { \mathrm { E } } }$ , initial data $\rho _ { \pi _ { 1 } }$ , and non-expert data $\rho _ { \pi _ { \mathrm { N E } } }$ . The red and blue regions denote the expert and non-expert parts of $\rho _ { \pi _ { 1 } }$ respectively. $H , O$ , and $N$ denote the latent demonstration, the observed demonstration, and the non-expert data respectively. (a) The ideal situation, where $\operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } ) \backslash \operatorname { s u p p } ( \rho _ { \pi _ { 1 } } ) = \emptyset$ ; (b) The real situation, where $\bar { H } : = \mathrm { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } ) \setminus \mathrm { s u p p } ( \rho _ { \pi _ { 1 } } ) \ne \emptyset$ in $\rho _ { \pi _ { \mathrm { E } } }$ . (c) The target output of the combined model $\mathbb { I } [ D _ { w } ^ { * } ] g ^ { * }$ . The output $+ 1$ , 0, and $- 1$ regions correspond to $H ,$ , and $N$ respectively.
|
| 190 |
+
|
| 191 |
+
148 where α(x) ≜ ρπE (x) is an importance-weighting factor [12]. So the current issue lies in how to
|
| 192 |
+
149 estimate $\frac { \rho _ { \pi _ { \mathrm { E } } } } { \rho _ { \pi _ { 1 } } }$ under $\mathcal { O } _ { \mathrm { E } }$ . To achieve this purpose, we need to bridge the expert demonstrations and
|
| 193 |
+
150 the initial data. Therefore, here we use these two data sets to train an adversarial model $D _ { w _ { 1 } }$ in the
|
| 194 |
+
151 same way as $D _ { w _ { 2 } }$ in the pretraining:
|
| 195 |
+
|
| 196 |
+
$$
|
| 197 |
+
\operatorname* { m a x } _ { w _ { 1 } } \mathcal { L } ( D _ { w _ { 1 } } ) \triangleq \mathbb { E } _ { x \sim \rho _ { \pi _ { 1 } } } [ \log D _ { w _ { 1 } } ( \widetilde { x } ) ] + \mathbb { E } _ { x \sim \rho _ { \pi _ { \mathrm { E } } } } [ \log ( 1 - D _ { w _ { 1 } } ( \widetilde { x } ) ) ] .
|
| 198 |
+
$$
|
| 199 |
+
|
| 200 |
+
152 If we write the training criterion (5) in the form of integral, i.e.,
|
| 201 |
+
|
| 202 |
+
$$
|
| 203 |
+
\operatorname* { m a x } _ { w _ { 1 } } \mathcal { L } ( D _ { w _ { 1 } } ) = \int _ { x } [ \rho _ { \pi _ { 1 } } \log D _ { w _ { 1 } } + \rho _ { \pi _ { \mathrm { E } } } \log ( 1 - D _ { w _ { 1 } } ) ] d x ,
|
| 204 |
+
$$
|
| 205 |
+
|
| 206 |
+
$\begin{array} { r } { ( \frac { \partial \mathcal { L } } { \partial D _ { w _ { 1 } } } = 0 ) } \end{array}$ $D _ { w _ { 1 } }$
|
| 207 |
+
|
| 208 |
+
$$
|
| 209 |
+
D _ { w _ { 1 } } ^ { * } = \frac { \rho _ { \pi _ { 1 } } } { \rho _ { \pi _ { 1 } } + \rho _ { \pi _ { \mathrm { E } } } } ,
|
| 210 |
+
$$
|
| 211 |
+
|
| 212 |
+
154 in which the order of differentiation and integration was changed by the Leibniz rule. Besides, we
|
| 213 |
+
155 can sufficiently train $D _ { w _ { 1 } }$ using the initial data $\widetilde { \mathcal { T } } _ { \pi _ { 1 } }$ and the expert demonstrations $\widetilde { \mathcal { T } } _ { \pi _ { \mathrm { E } } }$ . Then $D _ { w _ { 1 } }$
|
| 214 |
+
156 will be good enough to estimate the importance-weighting factor, i.e.,
|
| 215 |
+
|
| 216 |
+
$$
|
| 217 |
+
\alpha ( x ) \triangleq \frac { \rho _ { \pi _ { \mathtt { E } } } } { \rho _ { \pi _ { 1 } } } = \frac { 1 - D _ { w _ { 1 } } ^ { * } ( \widetilde { x } ) } { D _ { w _ { 1 } } ^ { * } ( \widetilde { x } ) } \approx \frac { 1 - D _ { w _ { 1 } } ( \widetilde { x } ) } { D _ { w _ { 1 } } ( \widetilde { x } ) } .
|
| 218 |
+
$$
|
| 219 |
+
|
| 220 |
+
In this way, we can use 157 $D _ { w 1 }$ , which can connect demonstrations and initial data, to calibrate the learning process of 158 $D _ { w _ { 2 } }$ . The final optimization objective for $D _ { w _ { 2 } }$ is
|
| 221 |
+
|
| 222 |
+
$$
|
| 223 |
+
\operatorname* { m a x } _ { w _ { 2 } } \mathcal { L } ( D _ { w _ { 2 } } ) = \mathbb { E } _ { x \sim \rho _ { \pi _ { 2 } } } \log D _ { w _ { 2 } } ( \overline { { x } } ) + \mathbb { E } _ { x \sim \rho _ { \pi _ { 1 } } } \frac { 1 - D _ { w _ { 1 } } ( \overline { { x } } ) } { D _ { w _ { 1 } } ( \widetilde { x } ) } \log [ 1 - D _ { w _ { 2 } } ( \overline { { x } } ) ] .
|
| 224 |
+
$$
|
| 225 |
+
|
| 226 |
+
In this way, 159 $D _ { w _ { 2 } }$ can effectively dig out the expert part of $\rho _ { \pi _ { 1 } }$ and produce efficient rewards for $\pi _ { 2 }$
|
| 227 |
+
|
| 228 |
+
# 4.2 Support Mismatch
|
| 229 |
+
|
| 230 |
+
161 So far the challenges have still been similar to homogeneously observable imitation learning. However,
|
| 231 |
+
162 our preliminary experiments demonstrated that merely importance-weighting is not enough to fix
|
| 232 |
+
163 the problem that occurred by the absence of interactions under $\mathcal { O } _ { \mathrm { E } }$ . So there exist some other issues
|
| 233 |
+
164 between the expert demonstrations and the initial data. To find out the underlying issues, we plot
|
| 234 |
+
165 the t-Distributed Stochastic Neighbor Embedding (t-SNE) [34] visualizations of these two empirical
|
| 235 |
+
166 distributions under $\mathcal { O } _ { \mathrm { E } }$ on Hopper and Walker2d, as shown in Figure 5. Twenty trajectories were
|
| 236 |
+
167 collected for both the expert demonstrations and the initial data. We can observe that there exist some
|
| 237 |
+
168 high-density regions of demonstrations in which the initial data do not cover; that is, there exist some
|
| 238 |
+
169 regions of the demonstrations that $\pi _ { 1 }$ did not explore. Wang et al. [35] found a similar phenomenon in
|
| 239 |
+
170 the standard $\mathrm { I L }$ setting. On the other hand, the importance-weighting $\alpha$ cannot calibrate this situation
|
| 240 |
+
171 where $\frac { \rho _ { \pi _ { \mathrm { E } } } } { \rho _ { \pi _ { 1 } } } = \infty$
|
| 241 |
+
172 To formulate this problem, here we introduce the Support
|
| 242 |
+
173 Set of the occupancy measure:
|
| 243 |
+
174 Definition 1 (Support Set). The support set of a occu
|
| 244 |
+
175 pancy measure $\rho$ is the subset of the domain containing
|
| 245 |
+
176 the elements which are not mapped to zero:
|
| 246 |
+
|
| 247 |
+
$$
|
| 248 |
+
\operatorname { s u p p } ( \rho ) : = \{ x \in { \mathcal { S } } \times { \mathcal { A } } | \rho ( x ) \neq 0 \} .
|
| 249 |
+
$$
|
| 250 |
+
|
| 251 |
+
177 Due to the sub-optimality of $\pi _ { 1 }$ , $\operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } ) \backslash \operatorname { s u p p } ( \rho _ { \pi _ { 1 } } ) \neq$
|
| 252 |
+
178 $\mathcal { D }$ (see Figure 4b). We call this part the Latent Demonstra
|
| 253 |
+
179 tion, defined as:
|
| 254 |
+
180 Definition 2 (Latent Demonstration). The latent demon
|
| 255 |
+
181 stration $H$ is the set of those $x \in { \mathcal { S } } \times { \mathcal { A } }$ that belong to the
|
| 256 |
+
182 relative complement of supp $\left( \rho _ { \pi _ { 1 } } \right)$ in $\operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } )$ :
|
| 257 |
+
|
| 258 |
+
|
| 259 |
+
Figure 5: t-SNE visualizations of expert demonstrations and collected data of $\pi _ { 1 }$ under $\mathcal { O } _ { \mathrm { E } }$ .
|
| 260 |
+
|
| 261 |
+
$$
|
| 262 |
+
H : = \{ x \in S \times A | \mathrm { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } ) \setminus \mathrm { s u p p } ( \rho _ { \pi _ { 1 } } ) \} .
|
| 263 |
+
$$
|
| 264 |
+
|
| 265 |
+
Also, another part of the demonstration is named the Observed Demonstration, defined as:
|
| 266 |
+
|
| 267 |
+
184 Definition 3 (Observed Demonstration). The observed demonstration $O$ is the set of those $x \in { \mathcal { S } } \times { \mathcal { A } }$ that belong to the complement of 185 $H$ in $\operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } )$ :
|
| 268 |
+
|
| 269 |
+
$$
|
| 270 |
+
O : = \{ x \in \mathcal { S } \times A \vert \mathrm { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } ) \cap \mathrm { s u p p } ( \rho _ { \pi _ { 1 } } ) \} .
|
| 271 |
+
$$
|
| 272 |
+
|
| 273 |
+
186 Besides, the data outside of demonstrations should be non-expert data:
|
| 274 |
+
|
| 275 |
+
187 Definition 4 (Non-Expert Data). The non-expert data $N$ is the set of those $x \in { \mathcal { S } } \times { \mathcal { A } }$ that out of
|
| 276 |
+
188 $\operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } )$ :
|
| 277 |
+
|
| 278 |
+
$$
|
| 279 |
+
N : = \{ x \in { \mathcal { S } } \times A | \rho _ { \pi _ { \mathrm { E } } } ( x ) = 0 \} .
|
| 280 |
+
$$
|
| 281 |
+
|
| 282 |
+
189 In other words, the sub-optimality of $\pi _ { 1 }$ will cause not only the dynamics mismatch, but also the
|
| 283 |
+
190 appearance of the latent demonstration $H$ . We call the latter one the problem of Support Mismatch.
|
| 284 |
+
191 Intuitively, when $\pi _ { 2 } \pi _ { \mathrm { E } }$ , we have $H \emptyset$ , monotonously. So in order to fix the support mismatch
|
| 285 |
+
192 between $\rho _ { \pi _ { \mathrm { E } } }$ and $\rho _ { \pi _ { 1 } }$ , guiding $\pi _ { 2 }$ to find out $H$ is the key.
|
| 286 |
+
|
| 287 |
+
In addition, the support mismatch problem can be viewed as an inverse problem of the Out Of Distribution (OOD) problem that frequently occurred in offline RL setting [21], in which they tried to avoid $\operatorname { s u p p } ( \rho _ { \pi _ { 1 } } ) \setminus \operatorname { s u p p } ( \rho _ { \pi _ { \mathrm { E } } } )$ instead.
|
| 288 |
+
|
| 289 |
+
# 4.3 Imitation Learning with Rejection
|
| 290 |
+
|
| 291 |
+
We can observe that $H \cup O \cup N = S \times { \mathcal { A } }$ . So it is desirable to filter out $H$ from $O$ and $N$ . Meanwhile, $D _ { w _ { 1 } }$ and $D _ { w _ { 2 } }$ can only classify $O \cup H$ and $N$ , under $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ respectively. Therefore, here we design two models $g _ { 1 } : \mathcal { O } _ { \mathrm { E } } \times \mathcal { A } \{ 0 , 1 \}$ and $g _ { 2 } : { \mathcal { O } } _ { \mathrm { L } } \times { \mathcal { A } } \{ 0 , 1 \}$ (Output 0: $x \in O$ and output 1: otherwise), so that given $x \sim \tau$ (corresponding $\widetilde { x } \sim \widetilde { \tau }$ and $\overline { { x } } \sim \overline { { \mathcal { T } } }$ ) they can satisfy:
|
| 292 |
+
|
| 293 |
+
$$
|
| 294 |
+
H = \{ x \in S \times \mathcal { A } | \mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \widetilde { x } ) ] g _ { 1 } ^ { * } ( \widetilde { x } ) = \mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \overline { { x } } ) ] g _ { 2 } ^ { * } ( \overline { { x } } ) = + 1 \} ,
|
| 295 |
+
$$
|
| 296 |
+
|
| 297 |
+
201
|
| 298 |
+
|
| 299 |
+
$$
|
| 300 |
+
O = \{ x \in \mathcal { S } \times \mathcal { A } | \mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \widetilde { x } ) ] g _ { 1 } ^ { * } ( \widetilde { x } ) = \mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \overline { { x } } ) ] g _ { 2 } ^ { * } ( \overline { { x } } ) = 0 \} ,
|
| 301 |
+
$$
|
| 302 |
+
|
| 303 |
+
$$
|
| 304 |
+
N = \{ x \in S \times \mathcal { A } | \mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \widetilde { x } ) ] g _ { 1 } ^ { * } ( \widetilde { x } ) = \mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \overline { { x } } ) ] g _ { 2 } ^ { * } ( \overline { { x } } ) = - 1 \} ,
|
| 305 |
+
$$
|
| 306 |
+
|
| 307 |
+
respectively, where $\mathbb { I } [ \cdot ]$ takes $+ 1$ if $\cdot > 0 . 5$ , and $- 1$ otherwise. The target combined model $\mathbb { I } [ D _ { w } ^ { * } ( x ) ] g ^ { * } ( x )$ is depicted in Figure4c.
|
| 308 |
+
|
| 309 |
+
205 To this end, both $g _ { 1 }$ and $g _ { 2 }$ should be able to cover $O$ , meanwhile $g _ { 2 }$ can be adaptive to continuously change of 206 $\rho _ { \pi _ { 2 } }$ due to the update of $\pi _ { 2 }$ . Here we learn $g _ { 1 }$ and $g _ { 2 }$ in a rejection form, to reject $O$ from
|
| 310 |
+
|
| 311 |
+
207 $O \cup H$ (where $\mathbb { I } ( D _ { w } ) = + 1$ ). Concretely, the rejection setting is the same as that in Cortes et al. [9].
|
| 312 |
+
208 Also inspired by Geifman et al. [15], the optimization objective of the combination of $D _ { w }$ and $g$ is
|
| 313 |
+
|
| 314 |
+
$$
|
| 315 |
+
\begin{array} { r } { \mathcal { L } ( D _ { w } , g ) \triangleq \hat { l } ( D _ { w } , g ) + \lambda \operatorname* { m a x } ( 0 , c - \hat { \phi } ( g ) ) ^ { 2 } , } \end{array}
|
| 316 |
+
$$
|
| 317 |
+
|
| 318 |
+
209 where $c > 0$ denotes the target coverage, and $\lambda$ denotes the factor for controlling the relative importance of rejection. Besides, the empirical coverage 210 $\hat { \phi } ( g )$ is defined as
|
| 319 |
+
|
| 320 |
+
$$
|
| 321 |
+
\hat { \phi } ( g | X ) \triangleq \frac { 1 } { m } \sum _ { i = 1 } ^ { m } g ( x _ { i } ) ,
|
| 322 |
+
$$
|
| 323 |
+
|
| 324 |
+
where a batch of data 211 $X = \{ x _ { i } \} , i \in [ m ]$ . The empirical rejection risk $\hat { l } ( D _ { w } , g )$ is the ratio between 212 the covered risk of the discriminator and the empirical coverage:
|
| 325 |
+
|
| 326 |
+
$$
|
| 327 |
+
\hat { l } ( D _ { w } , g ) \triangleq \frac { \frac { 1 } { m } \sum _ { i = 1 } ^ { m } \langle \mathcal { L } ( D _ { w } ( x _ { i } ) ) , g ( x _ { i } ) \rangle } { \hat { \phi } ( g ) } .
|
| 328 |
+
$$
|
| 329 |
+
|
| 330 |
+
Meanwhile, both $D _ { w _ { 1 } }$ and $g _ { 1 }$ can access $\rho _ { \pi _ { \mathrm { E } } }$ under $\mathcal { O } _ { \mathrm { E } }$ directly. So given $\overline { { x } } \sim \overline { { T } } _ { \pi _ { 2 } }$ under $\mathcal { O } _ { \mathrm { L } }$ , once $\langle \mathbb { I } ( D _ { w _ { 2 } } ( \overline { { x } } ) ) , g _ { 2 } ( \overline { { x } } ) \rangle = + 1$ , we can query the corresponding observations $\widetilde { x }$ of $\textstyle { \overline { { x } } }$ through OC 2operation and use $\langle \mathbb { I } ( D _ { w _ { 1 } } ( \widetilde { \boldsymbol { x } } ) ) , g _ { 1 } ( \widetilde { \boldsymbol { x } } ) \rangle$ to calibrate the output of $g _ { 2 }$ and $D _ { w _ { 2 } }$ e. In this way, $g _ { 2 }$ and $D _ { w _ { 2 } }$ e ecan be entangled together and adaptively guide $\pi _ { 2 }$ to find out the latent demonstrations $H$ under $\mathcal { O } _ { \mathrm { L } }$ .
|
| 331 |
+
|
| 332 |
+
# 4.4 IWRE
|
| 333 |
+
|
| 334 |
+
Here we combine the importance-weighting and rejection into a unified whole, to propose a novel algorithm named Importance Weighting with REjection (IWRE). Concretely, in a HOIL process:
|
| 335 |
+
|
| 336 |
+
Pretraining. We train a discriminator $D _ { w _ { 1 } }$ by Equation (5) and its corresponding rejection model $g _ { 1 }$ by Equation (17) using the initial data and the expert demonstrations.
|
| 337 |
+
|
| 338 |
+
Training. We train a discriminator $D _ { w _ { 2 } }$ by the combination of Equation (9) and Equation (17), as well as its corresponding rejection model $g _ { 2 }$ by Equation (17), using the initial data, the data collected by $\pi _ { 2 }$ , and the output of $D _ { w _ { 1 } }$ with $g _ { 1 }$ through OC operation. Also, $\pi _ { 2 }$ will be updated with $D _ { w _ { 2 } }$ and $g _ { 2 }$ asymmetrically as in GAIL.
|
| 339 |
+
|
| 340 |
+
The pseudo-code of our algorithm is provided in the supplementary material.
|
| 341 |
+
|
| 342 |
+
# 5 Experiment
|
| 343 |
+
|
| 344 |
+
In this section, we validate our algorithm in Atari 2600 [3] (GPL License) and MuJoCo [33] (Academic License) environments. The experiments were designed to investigate:
|
| 345 |
+
|
| 346 |
+
1) Can IWRE achieve significant performance under HOIL tasks?
|
| 347 |
+
2) Can IWRE deal with the support mismatch problem?
|
| 348 |
+
3) During training, is active querying for HOIL indeed necessary?
|
| 349 |
+
|
| 350 |
+
Below we first introduce the experimental setup and then investigate the above questions. More results and experimental details are included in the supplementary material.
|
| 351 |
+
|
| 352 |
+
# 5.1 Experimental Setup
|
| 353 |
+
|
| 354 |
+
Environments. We choose three pixel-memory based games in Atari and five continuous control objects in MuJoCo on OpenAI platform [5] (MIT License). Details as below:
|
| 355 |
+
|
| 356 |
+
1. Pixel-memory Atari games. $\mathcal { O } _ { \mathrm { E } }$ : $8 4 \times 8 4 \times 4$ raw pixels; $\mathcal { O } _ { \mathrm { L } }$ : 128-byte random access memories (RAM). Expert: converged DQN-based agents [24]. Atari games contain two totally isolated views: raw pixels and RAM, under the same state. Through these environments, we want to investigate whether the agent can learn an effective policy from demonstrations under completely different observation spaces. Moreover, IL with visual observations only is already very difficult [7], while learning a RAM-based policy can be even more challenging [3, 31], so few $\mathrm { I L }$ research reported desirable results on this task.
|
| 357 |
+
|
| 358 |
+
2. Continuous control MuJoCo objects. $\mathcal { O } _ { \mathrm { E } }$ : half of original observation features; $\mathcal { O } _ { \mathrm { L } }$ : another half of original observation features. Expert: converged DDPG-based agents [22]. The features of MuJoCo contain monotonous information like the direction, position, velocity, etc., of an object. Through these environments, we want to investigate whether the agent can learn from demonstrations with complementary signals under observations with missing information. Meanwhile, we make sure RL algorithms can obtain comparable performances under $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ . More details are reported in the supplementary material.
|
| 359 |
+
|
| 360 |
+
|
| 361 |
+
Figure 6: The learning curves of each method, where the shaded region indicates the standard deviation.
|
| 362 |
+
|
| 363 |
+
252 Besides, twenty expert trajectories were collected for each environment. Each result contains five
|
| 364 |
+
253 trials with different random seeds. All experiments were conducted on server clusters with NVIDIA
|
| 365 |
+
254 Tesla V100 GPUs. The summary of the environments is gathered in the supplementary material.
|
| 366 |
+
255 Baselines. Six basic contenders were included in the experiments: Vanilla GAIL [16], GAIL
|
| 367 |
+
256 with importance-weighting [12] (IW), third-person IL [30] (TPIL), generative adversarial MDP
|
| 368 |
+
257 alignment [19] (GAMA), behavioral cloning [2] (BC), and learning by cheating [8] (LBC). For
|
| 369 |
+
258 IW, we utilized the discriminator $D _ { w _ { 1 } }$ trained in the pretraining to calculate the importance weight;
|
| 370 |
+
259 also the optimization objective for $D _ { w _ { 2 } }$ during training is the same as Equation (9); TPIL learns the
|
| 371 |
+
260 third-person demonstrations by leading the cross-entropy loss into the update of the feature extractor;
|
| 372 |
+
261 GAMA learns a mapping function $\psi$ in view of adversarial training to align the observation of the
|
| 373 |
+
262 target domain into the source domain, and thereby can utilize the policy in the source domain for
|
| 374 |
+
263 zero-shot imitation. For fairness, we allowed the interaction between the policy and the environment
|
| 375 |
+
264 for GAMA under HOIL; LBC uses $\pi _ { 1 }$ learned from privileged states as a teacher to train $\pi _ { 2 }$ in a
|
| 376 |
+
265 DAgger [27] style, so here we allowed LBC to access $\mathcal { O } _ { \mathrm { E } }$ during the whole IL process. In Atari, to
|
| 377 |
+
266 investigate whether our method could achieve good performance for RAM-based control, we further
|
| 378 |
+
267 included a contender PPO-RAM, which uses proximal policy optimization (PPO) [28] to perform
|
| 379 |
+
268 RL directly with environmental true rewards under the RAM-based observations. More detailed
|
| 380 |
+
269 setup including query strategies for TPIL and GAMA, network architecture, and hyper-parameters
|
| 381 |
+
270 are reported in the supplementary material.
|
| 382 |
+
|
| 383 |
+
Learning process. To simulate the situation that $\mathcal { O } _ { \mathrm { E } }$ is costly, the steps for training $\pi _ { 1 }$ was set as 1/4 of that for training $\pi _ { 2 }$ , using GAIL [16]/HashReward [7] under the $\mathcal { O } _ { \mathrm { E } }$ space for MuJoCo/Atari environments. The learning steps were $1 0 ^ { 7 }$ for MuJoCo and $5 \times 1 0 ^ { 6 }$ for Atari environments. In the pretraining, we sampled 20 trajectories from $\pi _ { 1 }$ , and the data from each trajectory had both $\mathcal { O } _ { \mathrm { E } }$ and $\mathcal { O } _ { \mathrm { L } }$ observations. In the training, each method learned $4 \times 1 0 ^ { 7 }$ steps for MuJoCo and $2 \times 1 0 ^ { 7 }$ steps for Atari under the $\mathcal { O } _ { \mathrm { L } }$ space to obtain $\pi _ { 2 }$ .
|
| 384 |
+
|
| 385 |
+
# 5.2 Results
|
| 386 |
+
|
| 387 |
+
Experimental results are reported in Figure 6. Since the mapping function is hard to learn when input is RAM and output is raw images, we omit the results of GAMA in Atari. We can observe that while IW is better than GAIL in most environments, both GAIL and IW can hardly outperform $\pi _ { 1 }$
|
| 388 |
+
|
| 389 |
+
Because they just imitated the performance of $\pi _ { 1 }$ instead of $\pi _ { \mathrm { E } }$ , even with importance-weighting for calibration. For TPIL, its learning process was extremely unstable on Hopper, Swimmer, and Walker2d due to the continuous distribution shift. Furthermore, the performance of GAMA was not satisfactory in Hopper and Walker2d because its mapping function is hard to learn well when the support mismatch appears. The results of TPIL and GAMA demonstrate that DSIL methods will be invalid under heterogeneous observations as in HOIL tasks. On Atari environments, $\mathcal { O } _ { \mathrm { E } }$ contains more privileged information than $\mathcal { O } _ { \mathrm { L } }$ , so LBC can achieve good performance. But when $\mathcal { O } _ { \mathrm { E } }$ is not more privileged than $\mathcal { O } _ { \mathrm { L } }$ , like in most environments of MuJoCo, its performance will decrease due to the support mismatch, which would make it even worse than BC. Finally, IWRE obtained the best performance on 6/8 environments, and comparable performance with LBC on Reacher, which shows the effectiveness of our method even with limited access to $\mathcal { O } _ { \mathrm { E } }$ (LBC can access to $\mathcal { O } _ { \mathrm { E } }$ all the time). Besides, we can see that the performance differences between the GAIL/IW and IWRE/TPIL/GAMA/LBC are huge (especially on Reacher) because of the absence of queries, which demonstrates that the query operation is indeed necessary for HOIL problems.
|
| 390 |
+
|
| 391 |
+
Moreover, even learned with true rewards, PPO-RAM surprisingly failed to achieve comparable performance to IWRE, which shows that IWRE could possibly learn more effective rewards than true environmental rewards in RAM-input tasks. The results verify that, IWRE provides a powerful approach for tackling HOIL problems, even under the situation that the demonstrations are gathered from such a different observation space, meanwhile $\mathcal { O } _ { \mathrm { E } }$ is strictly limited during training.
|
| 392 |
+
|
| 393 |
+
t-SNE visualization of $\rho _ { \pi _ { 2 } }$ and $\rho _ { \pi _ { \mathrm { E } } }$ under $\mathcal { O } _ { \mathrm { E } }$ . In Section 4.2, we point that the sub-optimality of $\pi _ { 1 }$ will cause the problem of support mismatch, which is embodied as the appearance of the latent demonstration $H$ during training. Also the empirical results in Figure 5 on Hopper and Walker2d verify the existence of this problem. So we want to investigate whether the superiority of IWRE indeed comes from successfully tackling the support mismatch problem. To this end, we plotted the t-SNE visualization of the same expert demonstrations as in Section 4.2 and the collected data of $\pi _ { 2 }$ by IWRE under $\mathcal { O } _ { \mathrm { E } }$ ( $\mathcal { O } _ { \mathrm { E } }$ is hidden to $\pi _ { 2 }$ ). All setups are the same as in Section 4.2. From the results shown in Figure 7, we can see that even under $\mathcal { O } _ { \mathrm { E } }$ which cannot be obtained by $\pi _ { 2 }$ , almost all high-density regions of the demonstrations were covered by the collected data. Meanwhile, the latent demonstration $H$ is dug out nearly. The results demonstrate that IWRE basically solves the problem of support mismatch and thereby performs well in these environments.
|
| 394 |
+
|
| 395 |
+
|
| 396 |
+
Figure 7: t-SNE visualizations of expert demonstrations and collected data of $\pi _ { 2 }$ under $\mathcal { O } _ { \mathrm { E } }$ . The high-density regions of the expert demonstrations were covered by the collected data of $\pi _ { 2 }$ of IWRE.
|
| 397 |
+
|
| 398 |
+
Besides, some collected data of $\pi _ { 2 }$ of IWRE were out of the distribution of the demonstrations, which means $\pi _ { 2 }$ slightly overly explored the environment. Since $\mathcal { O } _ { \mathrm { E } }$ is hidden to $\pi _ { 2 }$ , the reward function will encourage $\pi _ { 2 }$ to explore more areas to fix the support mismatch problem. Meanwhile, the out-of-distribution problem in HOIL is not as severe as in the offline RL settings [21], so this over-exploration phenomenon makes sense.
|
| 399 |
+
|
| 400 |
+
# 6 Conclusion
|
| 401 |
+
|
| 402 |
+
In this paper, we proposed a new learning framework named Heterogeneously Observable Imitation Learning (HOIL), to formulate the situations where the observation space of demonstrations is different from that of the imitator while learning. We formally modeled a learning process of HOIL, in which the access to the observations of an expert is limited due to the high cost. Furthermore, we analyzed underlying challenges during training, i.e., the dynamics mismatch and the support mismatch, on the occupancy distributions between the demonstrations and the policy. To tackle these challenges, we proposed a new algorithm named Importance Weighting with REjection (IWRE), using importance-weighting and learning with rejection. Experimental results showed that the direct imitation and domain adaptive methods could not solve this problem, while our approach obtained promising results. In the future, we hope to involve the theoretical guarantee for our algorithm IWRE and investigate how many $\mathcal { O } _ { \mathrm { E } }$ do we need to query to learn a promising $\pi _ { 2 }$ . Furthermore, we hope to use the learning framework of HOIL and IWRE to tackle more learning scenarios with demonstrations in different spaces.
|
| 403 |
+
|
| 404 |
+
337 References [1] Pieter Abbeel and Andrew Y. Ng. Inverse reinforcement learning. In Encyclopedia of Machine Learning, pages 554–558. 2010. [2] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence 15, pages 103–129, 1996. [3] Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents. J. Artif. Intell. Res., 47:253–279, 2013. [4] Kianté Brantley, Hal Daumé III, and Amr Sharaf. Active imitation learning with noisy guidance. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel R. Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 2093–2105. Association for Computational Linguistics, 2020. [5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. CoRR, abs/1606.01540, 2016. [6] Alberto Broggi, Michele Buzzoni, Stefano Debattisti, Paolo Grisleri, Maria Chiara Laghi, Paolo Medici, and Pietro Versari. Extensive tests of autonomous driving technologies. IEEE Trans. Intell. Transp. Syst., 14(3):1403–1415, 2013. [7] Xin-Qiang Cai, Yao-Xiang Ding, Yuan Jiang, and Zhi-Hua Zhou. Imitation learning from pixel-level demonstrations by hashreward. In Proceedings of the 20th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), page 279–287, 2021. [8] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krähenbühl. Learning by cheating. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, 3rd Annual Conference on Robot Learning, CoRL 2019, Osaka, Japan, October 30 - November 1, 2019, Proceedings, volume 100 of Proceedings of Machine Learning Research, pages 66–75. PMLR, 2019. [9] Corinna Cortes, Giulia DeSalvo, and Mehryar Mohri. Learning with rejection. In Ronald Ortner, Hans Ulrich Simon, and Sandra Zilles, editors, Algorithmic Learning Theory - 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings, volume 9925 of Lecture Notes in Computer Science, pages 67–82, 2016. [10] Mark Cutler, Thomas J. Walsh, and Jonathan P. How. Reinforcement learning with multi-fidelity simulators. In 2014 IEEE International Conference on Robotics and Automation, ICRA 2014, Hong Kong, China, May 31 - June 7, 2014, pages 3888–3895. IEEE, 2014. [11] Siddharth Desai, Ishan Durugkar, Haresh Karnan, Garrett Warnell, Josiah Hanna, and Peter Stone. An imitation from observation approach to transfer learning with dynamics mismatch. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. [12] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under distribution shift. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. [13] Justin Fu, Katie Luo, and Sergey Levine. Learning robust rewards with adverserial inverse reinforcement learning. In International Conference on Learning Representations, 2018. [14] Tanmay Gangwani, Joel Lehman, Qiang Liu, and Jian Peng. Learning belief representations for imitation learning in pomdps. In Amir Globerson and Ricardo Silva, editors, Proceedings of the Thirty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI 2019, Tel Aviv, Israel, July 22-25, 2019, volume 115 of Proceedings of Machine Learning Research, pages 1061–1071. AUAI Press, 2019.
|
| 405 |
+
|
| 406 |
+
385 [15] Yonatan Geifman and Ran El-Yaniv. Selectivenet: A deep neural network with an integrated
|
| 407 |
+
386 reject option. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the
|
| 408 |
+
387 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach,
|
| 409 |
+
388 California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2151–2159.
|
| 410 |
+
389 PMLR, 2019.
|
| 411 |
+
390 [16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In Advances
|
| 412 |
+
391 in Neural Information Processing Systems 29: Annual Conference on Neural Information
|
| 413 |
+
392 Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4565–4573, 2016.
|
| 414 |
+
393 [17] Shengyi Jiang, Jing-Cheng Pang, and Yang Yu. Offline imitation learning with a misspecified
|
| 415 |
+
394 simulator. In Advances in Neural Information Processing Systems 33: Annual Conference on
|
| 416 |
+
395 Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual,
|
| 417 |
+
396 2020.
|
| 418 |
+
397 [18] Kun Ho Kim, Yihong Gu, Jiaming Song, Shengjia Zhao, and Stefano Ermon. Cross domain
|
| 419 |
+
398 imitation learning. CoRR, abs/1910.00105, 2019.
|
| 420 |
+
399 [19] Kuno Kim, Yihong Gu, Jiaming Song, Shengjia Zhao, and Stefano Ermon. Domain adaptive
|
| 421 |
+
400 imitation learning. In Proceedings of the 37th International Conference on Machine Learning,
|
| 422 |
+
401 ICML 2020, 13-18 July 2020, Virtual Event, pages 5286–5295, 2020.
|
| 423 |
+
402 [20] Bangalore Ravi Kiran, Ibrahim Sobh, Victor Talpaert, Patrick Mannion, Ahmad A. Al Sallab,
|
| 424 |
+
403 Senthil Kumar Yogamani, and Patrick Pérez. Deep reinforcement learning for autonomous
|
| 425 |
+
404 driving: A survey. CoRR, abs/2002.00444, 2020.
|
| 426 |
+
405 [21] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning:
|
| 427 |
+
406 Tutorial, review, and perspectives on open problems. CoRR, abs/2005.01643, 2020.
|
| 428 |
+
407 [22] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
|
| 429 |
+
408 Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning.
|
| 430 |
+
409 In Yoshua Bengio and Yann LeCun, editors, 4th International Conference on Learning Repre
|
| 431 |
+
410 sentations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings,
|
| 432 |
+
411 2016.
|
| 433 |
+
412 [23] Yuxuan Liu, Abhishek Gupta, Pieter Abbeel, and Sergey Levine. Imitation from observation:
|
| 434 |
+
13 Learning to imitate behaviors from raw video via context translation. In 2018 IEEE International
|
| 435 |
+
414 Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018,
|
| 436 |
+
415 pages 1118–1125, 2018.
|
| 437 |
+
416 [24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan
|
| 438 |
+
417 Wierstra, and Martin A. Riedmiller. Playing atari with deep reinforcement learning. CoRR,
|
| 439 |
+
18 abs/1312.5602, 2013.
|
| 440 |
+
419 [25] Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan P. How, and John Vian. Deep
|
| 441 |
+
420 decentralized multi-task multi-agent reinforcement learning under partial observability. In
|
| 442 |
+
421 Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney,
|
| 443 |
+
422 NSW, Australia, 6-11 August 2017, pages 2681–2690, 2017.
|
| 444 |
+
423 [26] Dripta S. Raychaudhuri, Sujoy Paul, Jeroen van Baar, and Amit K. Roy-Chowdhury. Cross
|
| 445 |
+
424 domain imitation from observations. In Marina Meila and Tong Zhang, editors, Proceedings of
|
| 446 |
+
425 the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual
|
| 447 |
+
426 Event, volume 139 of Proceedings of Machine Learning Research, pages 8902–8912. PMLR,
|
| 448 |
+
427 2021.
|
| 449 |
+
428 [27] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and
|
| 450 |
+
429 structured prediction to no-regret online learning. In Proceedings of the Fourteenth International
|
| 451 |
+
430 Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April
|
| 452 |
+
431 11-13, 2011, pages 627–635, 2011.
|
| 453 |
+
432 [28] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
|
| 454 |
+
433 policy optimization algorithms. CoRR, abs/1707.06347, 2017.
|
| 455 |
+
434 [29] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and
|
| 456 |
+
435 Sergey Levine. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE
|
| 457 |
+
436 International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May
|
| 458 |
+
437 21-25, 2018, pages 1134–1141, 2018.
|
| 459 |
+
438 [30] Bradly C. Stadie, Pieter Abbeel, and Ilya Sutskever. Third person imitation learning. In 5th
|
| 460 |
+
439 International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26,
|
| 461 |
+
440 2017, Conference Track Proceedings. OpenReview.net, 2017.
|
| 462 |
+
441 [31] Jakub Sygnowski and Henryk Michalewski. Learning from the memory of atari 2600. In Tristan
|
| 463 |
+
442 Cazenave, Mark H. M. Winands, Stefan Edelkamp, Stephan Schiffel, Michael Thielscher, and
|
| 464 |
+
443 Julian Togelius, editors, Computer Games - 5th Workshop on Computer Games, CGW 2016,
|
| 465 |
+
444 and 5th Workshop on General Intelligence in Game-Playing Agents, GIGA 2016, Held in
|
| 466 |
+
445 Conjunction with the 25th International Conference on Artificial Intelligence, IJCAI 2016, New
|
| 467 |
+
446 York City, NY, USA, July 9-10, 2016, Revised Selected Papers, volume 705 of Communications
|
| 468 |
+
447 in Computer and Information Science, pages 71–85, 2016.
|
| 469 |
+
48 [32] Andrea Tirinzoni, Andrea Sessa, Matteo Pirotta, and Marcello Restelli. Importance weighted
|
| 470 |
+
449 transfer of samples in reinforcement learning. In Jennifer G. Dy and Andreas Krause, editors,
|
| 471 |
+
450 Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stock
|
| 472 |
+
451 holmsmässan, Stockholm, Sweden, July 10-15, 2018, volume 80 of Proceedings of Machine
|
| 473 |
+
452 Learning Research, pages 4943–4952. PMLR, 2018.
|
| 474 |
+
53 [33] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based
|
| 475 |
+
454 control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS
|
| 476 |
+
455 2012, Vilamoura, Algarve, Portugal, October 7-12, 2012, pages 5026–5033, 2012.
|
| 477 |
+
456 [34] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine
|
| 478 |
+
457 Learning Research, 9:2579–2605, 2008.
|
| 479 |
+
58 [35] Ruohan Wang, Carlo Ciliberto, Pierluigi Vito Amadori, and Yiannis Demiris. Random expert
|
| 480 |
+
459 distillation: Imitation learning via expert policy support estimation. In Kamalika Chaudhuri and
|
| 481 |
+
460 Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine
|
| 482 |
+
461 Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings
|
| 483 |
+
462 of Machine Learning Research, pages 6536–6544. PMLR, 2019.
|
| 484 |
+
463 [36] Andrew Warrington, Jonathan Wilder Lavington, Adam Scibior, Mark Schmidt, and Frank ´
|
| 485 |
+
464 Wood. Robust asymmetric learning in pomdps. In Marina Meila and Tong Zhang, editors,
|
| 486 |
+
465 Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24
|
| 487 |
+
466 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages
|
| 488 |
+
467 11013–11023. PMLR, 2021.
|
| 489 |
+
|
| 490 |
+
# 468 Checklist
|
| 491 |
+
|
| 492 |
+
1. For all authors...
|
| 493 |
+
|
| 494 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 495 |
+
(b) Did you describe the limitations of your work? [Yes] See Section 6.
|
| 496 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] See supplementary material.
|
| 497 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 498 |
+
|
| 499 |
+
2. If you are including theoretical results...
|
| 500 |
+
|
| 501 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 502 |
+
|
| 503 |
+
3. If you ran experiments...
|
| 504 |
+
|
| 505 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See Section 5.
|
| 506 |
+
|
| 507 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See supplementary material.
|
| 508 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See Section 5.
|
| 509 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Section 5.
|
| 510 |
+
|
| 511 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 512 |
+
|
| 513 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 514 |
+
(b) Did you mention the license of the assets? [Yes]
|
| 515 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]
|
| 516 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 517 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A]
|
| 518 |
+
|
| 519 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 520 |
+
|
| 521 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 522 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 523 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf_content_list.json
ADDED
|
@@ -0,0 +1,1493 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Seeing Differently, Acting Similarly: Heterogeneously Observable Imitation Learning ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
207,
|
| 8 |
+
122,
|
| 9 |
+
794,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Anonymous Author(s) \nAffiliation \nAddress \nemail ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
423,
|
| 19 |
+
226,
|
| 20 |
+
580,
|
| 21 |
+
281
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
318,
|
| 32 |
+
535,
|
| 33 |
+
334
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "1 In many real-world imitation learning tasks, the demonstrator and the learner have \n2 to act under totally different observation spaces. This situation brings significant \n3 obstacles to existing imitation learning approaches, since most of them learn poli \n4 cies under homogeneous observation spaces. On the other hand, previous studies \n5 under different observation spaces have strong assumptions that these two obser \n6 vation spaces coexist during the entire learning process. However, in reality, the \n7 observation coexistence will be limited due to the high cost of acquiring expert \n8 observations. In this work, we study this challenging problem with limited observa \n9 tion coexistence under heterogeneous observations: Heterogeneously Observable \n10 Imitation Learning (HOIL). We identify two underlying issues in HOIL, i.e., the \n11 dynamics mismatch and the support mismatch, and further propose the Impor \n12 tance Weighting with REjection (IWRE) algorithm based on importance-weighting \n13 and learning with rejection to solve HOIL problems. Experimental results show \n14 that IWRE can successfully solve various HOIL tasks, including the challenging \n15 tasks of transforming the vision-based demonstrations to random access memory \n16 (RAM)-based policies in the Atari domain, even with limited visual observations. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
148,
|
| 42 |
+
349,
|
| 43 |
+
766,
|
| 44 |
+
569
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "17 1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
148,
|
| 54 |
+
587,
|
| 55 |
+
312,
|
| 56 |
+
604
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "18 Imitation Learning (IL) studies how to learn a good policy by imitating the given expert demonstra \n19 tions [16, 1], and has achieved great success in many domains such as autonomous driving [8], video \n20 games [7], and continuous control [19]. In real-world IL applications, the expert and the learner \n21 usually have their own observations of the same underlying states from the environment. For example, \n22 in Figure 1, an autonomous agent is learning to drive by imitating a human expert. The expert takes \n23 her actions mainly based on auditory and visual observations, which are familiar to human beings. \n24 However, the learning agent does not necessarily use the same way to observe: it can utilize more \n25 machine-capable sensors such as a LiDAR, radar, and bird-eye view (BEV) map to generate its \n26 observations [20]. The key features behind this example are two-fold: First, both the expert and \n27 the learner have their totally different observations of the same state of the environment. Thus they \n28 essentially have to choose the same action if acting optimally. Second, the observation space of the \n29 expert is often of high cost for the learner to utilize [6, 10]. We call this problem Heterogeneously \n30 Observable Imitation Learning (HOIL). \n31 There are two lines of research studying the related problems. The first line relates to domain \n32 adaptation: the observation space of the expert and the learner are the homogeneous, while some \n33 typical mismatches of distributions could exist: morphological mismatch, viewpoint mismatch, and \n34 dynamics mismatch [30, 17, 26]. However, these approaches are invalid when the observation spaces \n35 for experts and learners are completely different as in HOIL. \n36 The second line studied IL under different observations similar to HOIL, and some representative \n37 works include Partially Observable Imitation Learning (POIL) [14, 36] and Learning by Cheating \n38 (LBC) [8], as depicted in Figure 2. Both POIL and LBC assume that the expert’s observations can \n39 be easily accessed by the learner without any budget limit. However in practice, different from the \n40 learner observations, the access to expert’s observations might be of high cost and invasive [6, 10], \n41 hindering the wide application of these methods. \n42 In this paper, we initialize the study of the HOIL problem. We propose a learning process across \n43 observation spaces of experts and learners for solving this problem, and analyze the underlying issues \n44 of HOIL, i.e., the dynamics mismatch and the support mismatch. To tackle both two issues, we resort \n45 to the techniques of importance-weighting [12] and learning with rejection [9, 15] for active querying \n46 to propose the Importance Weighting with REjection (IWRE) approach. We evaluate the effectiveness \n47 of the IWRE algorithm in continuous control tasks of MuJoCo [33], and the challenging tasks of \n48 learning random access memory (RAM)-based policies given vision-based expert demonstrations \n49 in Atari [3] games. The results demonstrate that IWRE can significantly outperform existing IL \n50 algorithms in HOIL tasks, with limited access to expert observations. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
145,
|
| 65 |
+
611,
|
| 66 |
+
825,
|
| 67 |
+
791
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "",
|
| 74 |
+
"bbox": [
|
| 75 |
+
148,
|
| 76 |
+
796,
|
| 77 |
+
825,
|
| 78 |
+
867
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "",
|
| 85 |
+
"bbox": [
|
| 86 |
+
147,
|
| 87 |
+
872,
|
| 88 |
+
825,
|
| 89 |
+
901
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/1dc1f5e63f789d48c776d38e96873fc49a67b29d1d3c786ba655b77df5216fe3.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: Autonomous driving: an example of the HOIL problem. Figures 1, 2 and 3 include some illustrations and pictures from the Internet (source: www.vecteezy.com). "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
336,
|
| 102 |
+
88,
|
| 103 |
+
661,
|
| 104 |
+
265
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "image",
|
| 110 |
+
"img_path": "images/f4d372f2a7fe173e99bb6193ffb6f24f2fbf0a0b68b8a0c475b94f539b8abffb.jpg",
|
| 111 |
+
"image_caption": [
|
| 112 |
+
"Figure 2: Comparisons of different IL processes under different observation spaces. The targets are all to learn $\\pi _ { 2 }$ based on the second observation space with an auxiliary policy $\\pi _ { 1 }$ from corresponding roll-out data $\\widetilde { \\tau }$ and $\\overline { { \\mathcal { T } } }$ . (a) POIL mainly emphasized that the expert can view full observations, while the observations for the learner are partial. (b) LBC assumed that the expert’s observations contain more privileged information than the learner’s. Both POIL and LBC can observe expert’s observations all along. (c) HOIL limits the amount of expert’s observations. "
|
| 113 |
+
],
|
| 114 |
+
"image_footnote": [],
|
| 115 |
+
"bbox": [
|
| 116 |
+
178,
|
| 117 |
+
330,
|
| 118 |
+
823,
|
| 119 |
+
429
|
| 120 |
+
],
|
| 121 |
+
"page_idx": 1
|
| 122 |
+
},
|
| 123 |
+
{
|
| 124 |
+
"type": "text",
|
| 125 |
+
"text": "",
|
| 126 |
+
"bbox": [
|
| 127 |
+
148,
|
| 128 |
+
531,
|
| 129 |
+
825,
|
| 130 |
+
587
|
| 131 |
+
],
|
| 132 |
+
"page_idx": 1
|
| 133 |
+
},
|
| 134 |
+
{
|
| 135 |
+
"type": "text",
|
| 136 |
+
"text": "",
|
| 137 |
+
"bbox": [
|
| 138 |
+
147,
|
| 139 |
+
593,
|
| 140 |
+
825,
|
| 141 |
+
718
|
| 142 |
+
],
|
| 143 |
+
"page_idx": 1
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"type": "text",
|
| 147 |
+
"text": "51 2 Related Work ",
|
| 148 |
+
"text_level": 1,
|
| 149 |
+
"bbox": [
|
| 150 |
+
150,
|
| 151 |
+
733,
|
| 152 |
+
321,
|
| 153 |
+
750
|
| 154 |
+
],
|
| 155 |
+
"page_idx": 1
|
| 156 |
+
},
|
| 157 |
+
{
|
| 158 |
+
"type": "text",
|
| 159 |
+
"text": "52 Domain-Shifted IL. For the standard IL process, where the learner and the expert share the same \n53 observation space, current state-of-the-art methods tend to learn the policy in an adversarial style [7], \n54 like GAIL [16]. When considering the domain mismatch problem, i.e., Domain-Shifted IL (DSIL), \n55 the research aims at addressing the static distributional shift of the optimal policies resulted from \n56 the environmental differences but still under homogeneous observation spaces. Stadie et al. [30], \n57 Sermanet et al. [29], and Liu et al. [23] studied the situation where the demonstrations are in view \n58 of a third person. Kim et al. [19] and Kim et al. [18] addressed the IL problem with morphological \n59 mismatch between the expert’s and learner’s environment. Stadie et al. [30], Tirinzoni et al. [32], and \n60 Desai et al. [11] focused on the calibration for the mismatch between simulators and the real world \n61 through some transfer learning styles. There are two major differences between HOIL and DSIL: \n62 One is that HOIL considers heterogeneous observation spaces instead of homogeneous ones; another \n63 is that without observation heterogeneity, DSIL can directly align two fixed domains, which may \n64 not be realistic for solving HOIL when two observation spaces are totally different. Thus HOIL is a \n65 significantly more challenging problem than DSIL. Besides, Chen et al. [8] learned a vision-based \n66 agent from a privileged expert. But it can obtain expert’s observations throughout the whole learning \n67 process, so it cannot handle the problem of the support mismatch under HOIL. \n68 POMDP. The problem of POMDPs, in which only partial observations are available for the agent(s), \n69 has been studied in the context of multi-agent [25, 36] and imitation learning [14, 36] problems. \n70 But distinct from HOIL, in a POMDP, the learner only have partial observations and share a same \n71 underlying observation space with the expert, which would become an obstacle for them to make \n72 decisions correctly. For example, Warrington et al. [36] assumed that the observation of the learner \n73 is partial than that of the expert. Instead, in HOIL, expert’s and learner’s observations are totally \n74 different from each other, while the learner’s observations are not belong to a part of the expert’s. For \n75 HOIL, the main challenge is to deal with the mismatches between the observation spaces, especially \n76 when the access to expert’s observations is strictly limited. ",
|
| 160 |
+
"bbox": [
|
| 161 |
+
148,
|
| 162 |
+
758,
|
| 163 |
+
825,
|
| 164 |
+
911
|
| 165 |
+
],
|
| 166 |
+
"page_idx": 1
|
| 167 |
+
},
|
| 168 |
+
{
|
| 169 |
+
"type": "text",
|
| 170 |
+
"text": "",
|
| 171 |
+
"bbox": [
|
| 172 |
+
147,
|
| 173 |
+
92,
|
| 174 |
+
825,
|
| 175 |
+
160
|
| 176 |
+
],
|
| 177 |
+
"page_idx": 2
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"type": "text",
|
| 181 |
+
"text": "",
|
| 182 |
+
"bbox": [
|
| 183 |
+
151,
|
| 184 |
+
162,
|
| 185 |
+
825,
|
| 186 |
+
289
|
| 187 |
+
],
|
| 188 |
+
"page_idx": 2
|
| 189 |
+
},
|
| 190 |
+
{
|
| 191 |
+
"type": "text",
|
| 192 |
+
"text": "3 The HOIL Problem ",
|
| 193 |
+
"text_level": 1,
|
| 194 |
+
"bbox": [
|
| 195 |
+
176,
|
| 196 |
+
299,
|
| 197 |
+
370,
|
| 198 |
+
316
|
| 199 |
+
],
|
| 200 |
+
"page_idx": 2
|
| 201 |
+
},
|
| 202 |
+
{
|
| 203 |
+
"type": "text",
|
| 204 |
+
"text": "In this section, we first give a formal definition of the HOIL setting, and then introduce the learning process for solving the HOIL problem. ",
|
| 205 |
+
"bbox": [
|
| 206 |
+
173,
|
| 207 |
+
323,
|
| 208 |
+
823,
|
| 209 |
+
352
|
| 210 |
+
],
|
| 211 |
+
"page_idx": 2
|
| 212 |
+
},
|
| 213 |
+
{
|
| 214 |
+
"type": "text",
|
| 215 |
+
"text": "3.1 Setting Definition ",
|
| 216 |
+
"text_level": 1,
|
| 217 |
+
"bbox": [
|
| 218 |
+
173,
|
| 219 |
+
359,
|
| 220 |
+
339,
|
| 221 |
+
375
|
| 222 |
+
],
|
| 223 |
+
"page_idx": 2
|
| 224 |
+
},
|
| 225 |
+
{
|
| 226 |
+
"type": "text",
|
| 227 |
+
"text": "A HOIL problem is defined within a Markov decision process with mutiple observation spaces, i.e., $\\langle \\mathcal { S } , \\{ \\mathcal { O } \\} , \\mathcal { A } , \\mathcal { P } , \\gamma \\rangle$ , where $s$ denotes the state space, $\\{ \\mathcal { O } \\}$ denotes a set of observation spaces, $\\mathcal { A }$ denotes the action space, $\\mathcal { P } : \\mathcal { S } \\times \\mathcal { A } \\times \\mathcal { S } \\mathbb { R }$ denotes the transition probability distribution of the state and action, and $\\gamma \\in ( 0 , 1 ]$ denotes the discount factor. Furthermore, a policy $\\pi$ over an observation space $\\mathcal { O }$ is defined as a function mapping from $\\mathcal { O }$ to $\\mathcal { A }$ , and we denote by $\\Pi _ { \\mathcal { O } }$ the set of all policies over $\\mathcal { O }$ . In HOIL, both the expert and the learner have their own observation spaces, which are denoted as $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ respectively. Both $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ are assumed to be produced by two bijective mappings $f _ { \\mathrm { E } } : S \\mathcal { O } _ { \\mathrm { E } }$ , $f _ { \\mathrm { L } } : S \\mathcal { O } _ { \\mathrm { L } }$ , which are unknown functions mapping the underlying true states to the observations. It is obvious to see that by this assumption, any policy over $\\mathcal { O } _ { \\mathrm { E } }$ has a unique correspondence over $\\mathcal { O } _ { \\mathrm { L } }$ . This makes HOIL possible since the target of HOIL is to find the corresponding policy of the expert policy under $\\mathcal { O } _ { \\mathrm { L } }$ . ",
|
| 228 |
+
"bbox": [
|
| 229 |
+
165,
|
| 230 |
+
378,
|
| 231 |
+
825,
|
| 232 |
+
531
|
| 233 |
+
],
|
| 234 |
+
"page_idx": 2
|
| 235 |
+
},
|
| 236 |
+
{
|
| 237 |
+
"type": "text",
|
| 238 |
+
"text": "92 A state-action pair $( s , a )$ , denoted by $x$ , is called an instance. Also, a trajectory $\\mathcal { T } = \\{ x _ { i } \\} , i \\in [ m ]$ \n93 is a set of $m$ instances. For each observation space, $\\boldsymbol { \\widetilde { x } } \\in \\mathcal { \\widetilde { T } } \\subseteq \\mathcal { O } _ { \\mathrm { E } } \\times \\mathcal { A }$ and $\\overline { { x } } \\in \\overline { { \\mathcal { T } } } \\subseteq \\mathcal { O } _ { \\mathrm { L } } \\times \\mathcal { A }$ , \n94 where $\\mathcal { O } _ { \\mathrm { E } } = f _ { \\mathrm { E } } ( \\boldsymbol { S } )$ and $\\mathcal { O } _ { \\mathrm { L } } = f _ { \\mathrm { L } } ( \\mathcal { S } )$ e. Furthermore, we define the occupancy measure of a policy $\\pi$ \n95 under the state space $s$ as $\\rho _ { \\pi } : \\mathcal { S } \\times \\mathcal { A } \\mathbb { R }$ such that $\\begin{array} { r } { \\rho _ { \\pi } ( x ) = \\pi ( a | o ) \\mathrm { P r } ( o | s ) \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } \\mathrm { P r } ( s _ { t } = s | \\pi ) } \\end{array}$ . \n96 Under HOIL, the learner accesses the expert demonstrations $\\widetilde { \\mathcal { T } } _ { \\pi _ { \\mathrm { E } } }$ , a set of instances sampled from $\\rho _ { \\pi _ { \\mathrm { E } } }$ . \n97 The goal of HOIL is to learn a policy $\\hat { \\pi }$ as the corresponding policy of $\\pi _ { \\mathrm { E } }$ over $\\mathcal { O } _ { \\mathrm { L } }$ . If $\\mathcal { O } _ { \\mathrm { E } } = \\mathcal { O } _ { \\mathrm { L } }$ , \n98 HOIL degenerates to standard $\\mathrm { I L }$ . GAIL [16] is one of the state-of-the-art $\\mathrm { I L }$ approaches under this \n99 situation, which tries to minimize the divergence between the learner’s and the expert’s occupancy \n100 measures $d ( \\rho _ { \\hat { \\pi } } , \\rho _ { \\pi _ { \\mathrm { E } } } )$ . The objective of GAIL is ",
|
| 239 |
+
"bbox": [
|
| 240 |
+
143,
|
| 241 |
+
536,
|
| 242 |
+
826,
|
| 243 |
+
667
|
| 244 |
+
],
|
| 245 |
+
"page_idx": 2
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"type": "equation",
|
| 249 |
+
"img_path": "images/555b3d9d3b633ea71021e713b3de017c423abb5af669c970a3d5f2657a14b379.jpg",
|
| 250 |
+
"text": "$$\n\\operatorname* { m i n } _ { \\hat { \\pi } } \\operatorname* { m a x } _ { w } \\mathbb { E } _ { { x } \\sim { \\rho } _ { \\pi _ { \\mathrm { E } } } } [ \\log D _ { w } ( \\widetilde { x } ) ] + \\mathbb { E } _ { { x } \\sim { \\rho } _ { \\hat { \\pi } } } [ \\log ( 1 - D _ { w } ( \\widetilde { x } ) ) ] - \\mathbb { H } ( \\hat { \\pi } ) ,\n$$",
|
| 251 |
+
"text_format": "latex",
|
| 252 |
+
"bbox": [
|
| 253 |
+
279,
|
| 254 |
+
671,
|
| 255 |
+
717,
|
| 256 |
+
694
|
| 257 |
+
],
|
| 258 |
+
"page_idx": 2
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"type": "text",
|
| 262 |
+
"text": "101 where $\\mathbb { H } ( \\hat { \\pi } )$ is the causal entropy performed as a regularization term, and $D _ { w } : { \\mathcal { O } } _ { \\mathrm { E } } \\times A \\to [ 0 , 1 ]$ is \n102 the discriminator of $\\pi _ { \\mathrm { E } }$ and $\\hat { \\pi }$ . GAIL solved Equation (1) by alternatively taking a gradient ascent \n103 step to train the discriminator $D _ { w }$ , and a minimization step to learn policy $\\hat { \\pi }$ based on an off-the-shelf \n104 RL algorithm with the pseudo reward $- \\log D _ { w } ( \\widetilde { x } )$ . ",
|
| 263 |
+
"bbox": [
|
| 264 |
+
142,
|
| 265 |
+
698,
|
| 266 |
+
825,
|
| 267 |
+
755
|
| 268 |
+
],
|
| 269 |
+
"page_idx": 2
|
| 270 |
+
},
|
| 271 |
+
{
|
| 272 |
+
"type": "text",
|
| 273 |
+
"text": "3.2 The Learning Process for Solving HOIL ",
|
| 274 |
+
"text_level": 1,
|
| 275 |
+
"bbox": [
|
| 276 |
+
174,
|
| 277 |
+
762,
|
| 278 |
+
490,
|
| 279 |
+
777
|
| 280 |
+
],
|
| 281 |
+
"page_idx": 2
|
| 282 |
+
},
|
| 283 |
+
{
|
| 284 |
+
"type": "text",
|
| 285 |
+
"text": "106 In HOIL, we need to cope with the absence of the learner’s observations in demonstrations and the \n107 high cost of collecting the expert’s observations while learning. So we introduce a learning process \n108 with pretraining across two different observation spaces for solving HOIL, as abstracted in Figure 3. \n109 Pretraining. Same to LBC [8], we assume that we can obtain an auxiliary policy $\\pi _ { 1 }$ based on $\\mathcal { O } _ { \\mathrm { E } }$ at \n110 the beginning. $\\pi _ { 1 }$ can be directly provided by any sources, or trained by GAIL or behavior cloning \n111 as did in LBC. Besides, we use this $\\pi _ { 1 }$ to sample some data $\\mathcal { T } _ { \\pi _ { 1 } }$ , which contain both observation \n112 under $\\mathcal { O } _ { \\mathrm { E } }$ (i.e., $\\widetilde { \\mathcal { T } } _ { \\pi _ { 1 } } .$ ) and $\\mathcal { O } _ { \\mathrm { L } }$ (i.e., $\\overline { { \\mathcal { T } } } _ { \\pi _ { 1 } }$ ), in order to connect these two different observation spaces. \n113 We name ${ \\mathcal { T } } _ { \\pi _ { 1 } } = \\{ { \\widetilde { \\mathcal { T } } _ { \\pi _ { 1 } } , \\overline { { { \\mathcal { T } } } } _ { \\pi _ { 1 } } } \\}$ the initial data. \n114 Training. Here we learn a policy $\\pi _ { 2 }$ from the initial data $\\overline { { \\mathcal { T } } } _ { \\pi _ { 1 } }$ and the collected data $\\overline { { \\mathcal { T } } } _ { \\pi _ { 2 } }$ , under \n115 $\\mathcal { O } _ { \\mathrm { L } }$ only. Besides, the learner is allowed for some operation of observation coexistence (OC): At \n116 some steps of learning, besides the observations $\\mathcal { O } _ { \\mathrm { L } }$ , the learner could also request $\\widetilde { \\tau } _ { \\pi _ { 2 } }$ from the \n117 corresponding observations $\\mathcal { O } _ { \\mathrm { E } }$ (e.g., from the human-understandable sensors). The final objective of \n118 HOIL is to learn a good policy $\\pi _ { 2 }$ under $\\mathcal { O } _ { \\mathrm { L } }$ . \n119 In practical applications, the auxiliary policy $\\pi _ { 1 }$ can also come from simulation training or direct \n120 imitation. But since $\\pi _ { 1 }$ is additionally provided, it is more practical to consider $\\pi _ { 1 }$ as a non-optimal \n121 policy. During training, OC is an essential operation for solving HOIL, which helps the learner \n122 address the issues of the dynamics mismatch and the support mismatch (especially the latter one). \n123 Also, in reality, we do not need an oracle for actions, which still needs OC for obtaining expert \n124 observations first, as in many active querying research [4, 8], so its cost will be relatively lower. \n125 Besides, the related work [8] also required an initialized policy $\\pi _ { 1 }$ to solve their problem, which act \n126 as a teacher under privileged $\\mathcal { O } _ { \\mathrm { E } }$ in the pretraining and then learned a vision-based student from the \n127 guidance of the teacher under both $\\mathcal { O } _ { \\mathrm { L } }$ and $\\mathcal { O } _ { \\mathrm { E } }$ . Their setting can be viewed as a variety of HOIL \n128 with optimal $\\pi _ { 1 }$ , unlimited $\\mathcal { O } _ { \\mathrm { E } }$ , and unlimited OC operations, so HOIL is actually a more practical \n129 learning framework. ",
|
| 286 |
+
"bbox": [
|
| 287 |
+
142,
|
| 288 |
+
787,
|
| 289 |
+
825,
|
| 290 |
+
830
|
| 291 |
+
],
|
| 292 |
+
"page_idx": 2
|
| 293 |
+
},
|
| 294 |
+
{
|
| 295 |
+
"type": "text",
|
| 296 |
+
"text": "",
|
| 297 |
+
"bbox": [
|
| 298 |
+
142,
|
| 299 |
+
835,
|
| 300 |
+
825,
|
| 301 |
+
912
|
| 302 |
+
],
|
| 303 |
+
"page_idx": 2
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"type": "image",
|
| 307 |
+
"img_path": "images/8824fb7c81cbd26c799c8cbd3f54c6b48e41ce69d7e67e6677bda4d3254f47c3.jpg",
|
| 308 |
+
"image_caption": [
|
| 309 |
+
"Figure 3: Illustration of a learning process across two different observation spaces for solving HOIL. $\\pi _ { 1 }$ is an auxiliary policy that additionally provided. "
|
| 310 |
+
],
|
| 311 |
+
"image_footnote": [],
|
| 312 |
+
"bbox": [
|
| 313 |
+
336,
|
| 314 |
+
88,
|
| 315 |
+
658,
|
| 316 |
+
217
|
| 317 |
+
],
|
| 318 |
+
"page_idx": 3
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"type": "text",
|
| 322 |
+
"text": "",
|
| 323 |
+
"bbox": [
|
| 324 |
+
140,
|
| 325 |
+
261,
|
| 326 |
+
825,
|
| 327 |
+
335
|
| 328 |
+
],
|
| 329 |
+
"page_idx": 3
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"type": "text",
|
| 333 |
+
"text": "",
|
| 334 |
+
"bbox": [
|
| 335 |
+
142,
|
| 336 |
+
340,
|
| 337 |
+
826,
|
| 338 |
+
424
|
| 339 |
+
],
|
| 340 |
+
"page_idx": 3
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"type": "text",
|
| 344 |
+
"text": "",
|
| 345 |
+
"bbox": [
|
| 346 |
+
142,
|
| 347 |
+
430,
|
| 348 |
+
825,
|
| 349 |
+
501
|
| 350 |
+
],
|
| 351 |
+
"page_idx": 3
|
| 352 |
+
},
|
| 353 |
+
{
|
| 354 |
+
"type": "text",
|
| 355 |
+
"text": "4 Imitation Learning with Importance-Weighting and Rejection ",
|
| 356 |
+
"text_level": 1,
|
| 357 |
+
"bbox": [
|
| 358 |
+
169,
|
| 359 |
+
512,
|
| 360 |
+
722,
|
| 361 |
+
531
|
| 362 |
+
],
|
| 363 |
+
"page_idx": 3
|
| 364 |
+
},
|
| 365 |
+
{
|
| 366 |
+
"type": "text",
|
| 367 |
+
"text": "In HOIL, the access frequency to $\\mathcal { O } _ { \\mathrm { E } }$ is strictly limited, so it is unrealistic to learn $\\pi _ { 2 }$ in a Dataset Aggregation (DAgger) style [27] as in LBC. Therefore, we resort to learning $\\pi _ { 2 }$ with a learned reward function by inverse reinforcement learning [1] in an adversarial learning style [16, 13]. ",
|
| 368 |
+
"bbox": [
|
| 369 |
+
166,
|
| 370 |
+
537,
|
| 371 |
+
825,
|
| 372 |
+
580
|
| 373 |
+
],
|
| 374 |
+
"page_idx": 3
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"type": "text",
|
| 378 |
+
"text": "134 In addition, both $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ are assumed to share the same latent state space $s$ as introduced in \n135 Section 3.1, so the following analysis will be based on $s$ , while the algorithm will handle the problem \n136 based on $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ specifically. ",
|
| 379 |
+
"bbox": [
|
| 380 |
+
145,
|
| 381 |
+
585,
|
| 382 |
+
825,
|
| 383 |
+
628
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 3
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "text",
|
| 389 |
+
"text": "4.1 Dynamics Mismatch and Importance-Weighting ",
|
| 390 |
+
"text_level": 1,
|
| 391 |
+
"bbox": [
|
| 392 |
+
160,
|
| 393 |
+
635,
|
| 394 |
+
549,
|
| 395 |
+
650
|
| 396 |
+
],
|
| 397 |
+
"page_idx": 3
|
| 398 |
+
},
|
| 399 |
+
{
|
| 400 |
+
"type": "text",
|
| 401 |
+
"text": "138 To analyze the learning process, we let $\\rho _ { \\pi _ { \\mathrm { E } } } , \\rho _ { \\pi _ { 1 } }$ , and $\\rho _ { \\pi _ { 2 } }$ be the occupancy measure distributions \n139 of the expert demonstrations, the initial data, and the data during training respectively. Since we \n140 need to consider the sub-optimality of $\\pi _ { 1 } , \\rho _ { \\pi _ { 1 } }$ should be a mixture distribution of the expert $\\rho _ { \\pi _ { \\mathrm { E } } }$ and \n141 non-expert $\\rho _ { \\pi _ { \\mathrm { N E } } }$ , i.e., there exists some $\\delta \\in ( 0 , 1 )$ such that ",
|
| 402 |
+
"bbox": [
|
| 403 |
+
142,
|
| 404 |
+
661,
|
| 405 |
+
825,
|
| 406 |
+
717
|
| 407 |
+
],
|
| 408 |
+
"page_idx": 3
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"type": "equation",
|
| 412 |
+
"img_path": "images/7f26c0ec083eb44bf44b542828d4bd1d6cf7a60f22ef80eb1e7dea9de6c93cdf.jpg",
|
| 413 |
+
"text": "$$\n\\rho _ { \\pi _ { 1 } } = \\delta \\rho _ { \\pi _ { \\mathrm { E } } } + ( 1 - \\delta ) \\rho _ { \\pi _ { \\mathrm { N E } } } ,\n$$",
|
| 414 |
+
"text_format": "latex",
|
| 415 |
+
"bbox": [
|
| 416 |
+
405,
|
| 417 |
+
722,
|
| 418 |
+
591,
|
| 419 |
+
739
|
| 420 |
+
],
|
| 421 |
+
"page_idx": 3
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"type": "text",
|
| 425 |
+
"text": "142 as depicted in Figure 4a. During training, the original objective of $\\pi _ { 2 }$ is to imitate $\\pi _ { \\mathrm { E } }$ through demonstrations. To this end, the original objective of reward function 143 $D _ { w _ { 2 } }$ for $\\pi _ { 2 }$ is to optimize ",
|
| 426 |
+
"bbox": [
|
| 427 |
+
138,
|
| 428 |
+
746,
|
| 429 |
+
831,
|
| 430 |
+
775
|
| 431 |
+
],
|
| 432 |
+
"page_idx": 3
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"type": "equation",
|
| 436 |
+
"img_path": "images/b507436c776c8372a05c36d385ff2d62fac4afbfd9499d55d630e32aa9a8289b.jpg",
|
| 437 |
+
"text": "$$\n\\operatorname* { m a x } _ { w _ { 2 } } \\mathbb { E } _ { { x } \\sim \\rho _ { \\pi _ { 2 } } } [ \\log D _ { w _ { 2 } } ( \\overline { { x } } ) ] + \\mathbb { E } _ { { x } \\sim \\rho _ { \\pi _ { \\mathrm { E } } } } [ \\log ( 1 - D _ { w _ { 2 } } ( \\overline { { x } } ) ) ] .\n$$",
|
| 438 |
+
"text_format": "latex",
|
| 439 |
+
"bbox": [
|
| 440 |
+
312,
|
| 441 |
+
780,
|
| 442 |
+
684,
|
| 443 |
+
804
|
| 444 |
+
],
|
| 445 |
+
"page_idx": 3
|
| 446 |
+
},
|
| 447 |
+
{
|
| 448 |
+
"type": "text",
|
| 449 |
+
"text": "144 But the expert demonstrations are only available under $\\mathcal { O } _ { \\mathrm { E } }$ . While during training, we can only utilize \n145 the initial data $\\overline { { \\mathcal { T } } } _ { \\pi _ { 1 } } \\sim \\rho _ { \\pi _ { 1 } }$ to learn $\\pi _ { 2 }$ and $D _ { w _ { 2 } }$ . Besides, as $\\pi _ { 1 }$ is sub-optimal, directly imitating ${ \\overline { { \\mathcal T } } } _ { \\pi _ { 1 } }$ \n146 could reduce the performance of the optimal $\\pi _ { 2 }$ to that of $\\pi _ { 1 }$ . So we use the importance-weighting to \n147 calibrate this dynamics mismatch, i.e., ",
|
| 450 |
+
"bbox": [
|
| 451 |
+
142,
|
| 452 |
+
814,
|
| 453 |
+
825,
|
| 454 |
+
873
|
| 455 |
+
],
|
| 456 |
+
"page_idx": 3
|
| 457 |
+
},
|
| 458 |
+
{
|
| 459 |
+
"type": "equation",
|
| 460 |
+
"img_path": "images/cc3edf33c26b5d4f87b27fd96ac74316a30138868a3d6651207008756532485e.jpg",
|
| 461 |
+
"text": "$$\n\\operatorname* { m a x } _ { w _ { 2 } } \\mathcal { L } ( D _ { w _ { 2 } } ) = \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { 2 } } } [ \\log D _ { w _ { 2 } } ( \\overline { { x } } ) ] + \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { 1 } } } [ \\alpha ( x ) \\log ( 1 - D _ { w _ { 2 } } ( \\overline { { x } } ) ) ] ,\n$$",
|
| 462 |
+
"text_format": "latex",
|
| 463 |
+
"bbox": [
|
| 464 |
+
258,
|
| 465 |
+
885,
|
| 466 |
+
736,
|
| 467 |
+
909
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 3
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "image",
|
| 473 |
+
"img_path": "images/b0d90c1fd358a303dee0ef989b88c3851c79c8b99248c7ed098e42e050f520df.jpg",
|
| 474 |
+
"image_caption": [
|
| 475 |
+
"Figure 4: The comparisons among the distributions of expert demonstrations $\\rho _ { \\pi _ { \\mathrm { E } } }$ , initial data $\\rho _ { \\pi _ { 1 } }$ , and non-expert data $\\rho _ { \\pi _ { \\mathrm { N E } } }$ . The red and blue regions denote the expert and non-expert parts of $\\rho _ { \\pi _ { 1 } }$ respectively. $H , O$ , and $N$ denote the latent demonstration, the observed demonstration, and the non-expert data respectively. (a) The ideal situation, where $\\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } ) \\backslash \\operatorname { s u p p } ( \\rho _ { \\pi _ { 1 } } ) = \\emptyset$ ; (b) The real situation, where $\\bar { H } : = \\mathrm { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } ) \\setminus \\mathrm { s u p p } ( \\rho _ { \\pi _ { 1 } } ) \\ne \\emptyset$ in $\\rho _ { \\pi _ { \\mathrm { E } } }$ . (c) The target output of the combined model $\\mathbb { I } [ D _ { w } ^ { * } ] g ^ { * }$ . The output $+ 1$ , 0, and $- 1$ regions correspond to $H ,$ , and $N$ respectively. "
|
| 476 |
+
],
|
| 477 |
+
"image_footnote": [],
|
| 478 |
+
"bbox": [
|
| 479 |
+
187,
|
| 480 |
+
103,
|
| 481 |
+
813,
|
| 482 |
+
239
|
| 483 |
+
],
|
| 484 |
+
"page_idx": 4
|
| 485 |
+
},
|
| 486 |
+
{
|
| 487 |
+
"type": "text",
|
| 488 |
+
"text": "148 where α(x) ≜ ρπE (x) is an importance-weighting factor [12]. So the current issue lies in how to \n149 estimate $\\frac { \\rho _ { \\pi _ { \\mathrm { E } } } } { \\rho _ { \\pi _ { 1 } } }$ under $\\mathcal { O } _ { \\mathrm { E } }$ . To achieve this purpose, we need to bridge the expert demonstrations and \n150 the initial data. Therefore, here we use these two data sets to train an adversarial model $D _ { w _ { 1 } }$ in the \n151 same way as $D _ { w _ { 2 } }$ in the pretraining: ",
|
| 489 |
+
"bbox": [
|
| 490 |
+
140,
|
| 491 |
+
347,
|
| 492 |
+
826,
|
| 493 |
+
417
|
| 494 |
+
],
|
| 495 |
+
"page_idx": 4
|
| 496 |
+
},
|
| 497 |
+
{
|
| 498 |
+
"type": "equation",
|
| 499 |
+
"img_path": "images/10dd53282b26697580f000530b891e2fb433bdb0310d8c1c81188d25471d2509.jpg",
|
| 500 |
+
"text": "$$\n\\operatorname* { m a x } _ { w _ { 1 } } \\mathcal { L } ( D _ { w _ { 1 } } ) \\triangleq \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { 1 } } } [ \\log D _ { w _ { 1 } } ( \\widetilde { x } ) ] + \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { \\mathrm { E } } } } [ \\log ( 1 - D _ { w _ { 1 } } ( \\widetilde { x } ) ) ] .\n$$",
|
| 501 |
+
"text_format": "latex",
|
| 502 |
+
"bbox": [
|
| 503 |
+
274,
|
| 504 |
+
428,
|
| 505 |
+
722,
|
| 506 |
+
454
|
| 507 |
+
],
|
| 508 |
+
"page_idx": 4
|
| 509 |
+
},
|
| 510 |
+
{
|
| 511 |
+
"type": "text",
|
| 512 |
+
"text": "152 If we write the training criterion (5) in the form of integral, i.e., ",
|
| 513 |
+
"bbox": [
|
| 514 |
+
143,
|
| 515 |
+
459,
|
| 516 |
+
588,
|
| 517 |
+
474
|
| 518 |
+
],
|
| 519 |
+
"page_idx": 4
|
| 520 |
+
},
|
| 521 |
+
{
|
| 522 |
+
"type": "equation",
|
| 523 |
+
"img_path": "images/d908b3baaa54a4b4705acb73daa3a040abbd7bcf0f431a2387d1b0125a3a1224.jpg",
|
| 524 |
+
"text": "$$\n\\operatorname* { m a x } _ { w _ { 1 } } \\mathcal { L } ( D _ { w _ { 1 } } ) = \\int _ { x } [ \\rho _ { \\pi _ { 1 } } \\log D _ { w _ { 1 } } + \\rho _ { \\pi _ { \\mathrm { E } } } \\log ( 1 - D _ { w _ { 1 } } ) ] d x ,\n$$",
|
| 525 |
+
"text_format": "latex",
|
| 526 |
+
"bbox": [
|
| 527 |
+
308,
|
| 528 |
+
479,
|
| 529 |
+
687,
|
| 530 |
+
513
|
| 531 |
+
],
|
| 532 |
+
"page_idx": 4
|
| 533 |
+
},
|
| 534 |
+
{
|
| 535 |
+
"type": "text",
|
| 536 |
+
"text": "$\\begin{array} { r } { ( \\frac { \\partial \\mathcal { L } } { \\partial D _ { w _ { 1 } } } = 0 ) } \\end{array}$ $D _ { w _ { 1 } }$ ",
|
| 537 |
+
"bbox": [
|
| 538 |
+
147,
|
| 539 |
+
520,
|
| 540 |
+
825,
|
| 541 |
+
540
|
| 542 |
+
],
|
| 543 |
+
"page_idx": 4
|
| 544 |
+
},
|
| 545 |
+
{
|
| 546 |
+
"type": "equation",
|
| 547 |
+
"img_path": "images/6438bb0d10b735490d1f19378ccd88037879ac34a7123f650c90813a051f47fa.jpg",
|
| 548 |
+
"text": "$$\nD _ { w _ { 1 } } ^ { * } = \\frac { \\rho _ { \\pi _ { 1 } } } { \\rho _ { \\pi _ { 1 } } + \\rho _ { \\pi _ { \\mathrm { E } } } } ,\n$$",
|
| 549 |
+
"text_format": "latex",
|
| 550 |
+
"bbox": [
|
| 551 |
+
433,
|
| 552 |
+
546,
|
| 553 |
+
563,
|
| 554 |
+
578
|
| 555 |
+
],
|
| 556 |
+
"page_idx": 4
|
| 557 |
+
},
|
| 558 |
+
{
|
| 559 |
+
"type": "text",
|
| 560 |
+
"text": "154 in which the order of differentiation and integration was changed by the Leibniz rule. Besides, we \n155 can sufficiently train $D _ { w _ { 1 } }$ using the initial data $\\widetilde { \\mathcal { T } } _ { \\pi _ { 1 } }$ and the expert demonstrations $\\widetilde { \\mathcal { T } } _ { \\pi _ { \\mathrm { E } } }$ . Then $D _ { w _ { 1 } }$ \n156 will be good enough to estimate the importance-weighting factor, i.e., ",
|
| 561 |
+
"bbox": [
|
| 562 |
+
140,
|
| 563 |
+
583,
|
| 564 |
+
825,
|
| 565 |
+
628
|
| 566 |
+
],
|
| 567 |
+
"page_idx": 4
|
| 568 |
+
},
|
| 569 |
+
{
|
| 570 |
+
"type": "equation",
|
| 571 |
+
"img_path": "images/fc134139b8dc6827e8dca97f8cc5d1ebf2475a1f8d286b6e51597211ac010ba3.jpg",
|
| 572 |
+
"text": "$$\n\\alpha ( x ) \\triangleq \\frac { \\rho _ { \\pi _ { \\mathtt { E } } } } { \\rho _ { \\pi _ { 1 } } } = \\frac { 1 - D _ { w _ { 1 } } ^ { * } ( \\widetilde { x } ) } { D _ { w _ { 1 } } ^ { * } ( \\widetilde { x } ) } \\approx \\frac { 1 - D _ { w _ { 1 } } ( \\widetilde { x } ) } { D _ { w _ { 1 } } ( \\widetilde { x } ) } .\n$$",
|
| 573 |
+
"text_format": "latex",
|
| 574 |
+
"bbox": [
|
| 575 |
+
348,
|
| 576 |
+
633,
|
| 577 |
+
650,
|
| 578 |
+
670
|
| 579 |
+
],
|
| 580 |
+
"page_idx": 4
|
| 581 |
+
},
|
| 582 |
+
{
|
| 583 |
+
"type": "text",
|
| 584 |
+
"text": "In this way, we can use 157 $D _ { w 1 }$ , which can connect demonstrations and initial data, to calibrate the learning process of 158 $D _ { w _ { 2 } }$ . The final optimization objective for $D _ { w _ { 2 } }$ is ",
|
| 585 |
+
"bbox": [
|
| 586 |
+
142,
|
| 587 |
+
676,
|
| 588 |
+
825,
|
| 589 |
+
705
|
| 590 |
+
],
|
| 591 |
+
"page_idx": 4
|
| 592 |
+
},
|
| 593 |
+
{
|
| 594 |
+
"type": "equation",
|
| 595 |
+
"img_path": "images/013534c1c8748676a264dd55b2242800e6925010cc1bc113c0cfe87fac1e2b69.jpg",
|
| 596 |
+
"text": "$$\n\\operatorname* { m a x } _ { w _ { 2 } } \\mathcal { L } ( D _ { w _ { 2 } } ) = \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { 2 } } } \\log D _ { w _ { 2 } } ( \\overline { { x } } ) + \\mathbb { E } _ { x \\sim \\rho _ { \\pi _ { 1 } } } \\frac { 1 - D _ { w _ { 1 } } ( \\overline { { x } } ) } { D _ { w _ { 1 } } ( \\widetilde { x } ) } \\log [ 1 - D _ { w _ { 2 } } ( \\overline { { x } } ) ] .\n$$",
|
| 597 |
+
"text_format": "latex",
|
| 598 |
+
"bbox": [
|
| 599 |
+
240,
|
| 600 |
+
710,
|
| 601 |
+
754,
|
| 602 |
+
746
|
| 603 |
+
],
|
| 604 |
+
"page_idx": 4
|
| 605 |
+
},
|
| 606 |
+
{
|
| 607 |
+
"type": "text",
|
| 608 |
+
"text": "In this way, 159 $D _ { w _ { 2 } }$ can effectively dig out the expert part of $\\rho _ { \\pi _ { 1 } }$ and produce efficient rewards for $\\pi _ { 2 }$ ",
|
| 609 |
+
"bbox": [
|
| 610 |
+
145,
|
| 611 |
+
757,
|
| 612 |
+
821,
|
| 613 |
+
773
|
| 614 |
+
],
|
| 615 |
+
"page_idx": 4
|
| 616 |
+
},
|
| 617 |
+
{
|
| 618 |
+
"type": "text",
|
| 619 |
+
"text": "4.2 Support Mismatch ",
|
| 620 |
+
"text_level": 1,
|
| 621 |
+
"bbox": [
|
| 622 |
+
166,
|
| 623 |
+
789,
|
| 624 |
+
343,
|
| 625 |
+
804
|
| 626 |
+
],
|
| 627 |
+
"page_idx": 4
|
| 628 |
+
},
|
| 629 |
+
{
|
| 630 |
+
"type": "text",
|
| 631 |
+
"text": "161 So far the challenges have still been similar to homogeneously observable imitation learning. However, \n162 our preliminary experiments demonstrated that merely importance-weighting is not enough to fix \n163 the problem that occurred by the absence of interactions under $\\mathcal { O } _ { \\mathrm { E } }$ . So there exist some other issues \n164 between the expert demonstrations and the initial data. To find out the underlying issues, we plot \n165 the t-Distributed Stochastic Neighbor Embedding (t-SNE) [34] visualizations of these two empirical \n166 distributions under $\\mathcal { O } _ { \\mathrm { E } }$ on Hopper and Walker2d, as shown in Figure 5. Twenty trajectories were \n167 collected for both the expert demonstrations and the initial data. We can observe that there exist some \n168 high-density regions of demonstrations in which the initial data do not cover; that is, there exist some \n169 regions of the demonstrations that $\\pi _ { 1 }$ did not explore. Wang et al. [35] found a similar phenomenon in \n170 the standard $\\mathrm { I L }$ setting. On the other hand, the importance-weighting $\\alpha$ cannot calibrate this situation \n171 where $\\frac { \\rho _ { \\pi _ { \\mathrm { E } } } } { \\rho _ { \\pi _ { 1 } } } = \\infty$ \n172 To formulate this problem, here we introduce the Support \n173 Set of the occupancy measure: \n174 Definition 1 (Support Set). The support set of a occu \n175 pancy measure $\\rho$ is the subset of the domain containing \n176 the elements which are not mapped to zero: ",
|
| 632 |
+
"bbox": [
|
| 633 |
+
140,
|
| 634 |
+
813,
|
| 635 |
+
825,
|
| 636 |
+
912
|
| 637 |
+
],
|
| 638 |
+
"page_idx": 4
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"type": "text",
|
| 642 |
+
"text": "",
|
| 643 |
+
"bbox": [
|
| 644 |
+
140,
|
| 645 |
+
90,
|
| 646 |
+
826,
|
| 647 |
+
152
|
| 648 |
+
],
|
| 649 |
+
"page_idx": 5
|
| 650 |
+
},
|
| 651 |
+
{
|
| 652 |
+
"type": "text",
|
| 653 |
+
"text": "",
|
| 654 |
+
"bbox": [
|
| 655 |
+
140,
|
| 656 |
+
157,
|
| 657 |
+
549,
|
| 658 |
+
186
|
| 659 |
+
],
|
| 660 |
+
"page_idx": 5
|
| 661 |
+
},
|
| 662 |
+
{
|
| 663 |
+
"type": "text",
|
| 664 |
+
"text": "",
|
| 665 |
+
"bbox": [
|
| 666 |
+
140,
|
| 667 |
+
190,
|
| 668 |
+
549,
|
| 669 |
+
233
|
| 670 |
+
],
|
| 671 |
+
"page_idx": 5
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"type": "equation",
|
| 675 |
+
"img_path": "images/a4bc31cf75bd006be405d2666269231a1c941f8c8fad370e0a5caa51192b7065.jpg",
|
| 676 |
+
"text": "$$\n\\operatorname { s u p p } ( \\rho ) : = \\{ x \\in { \\mathcal { S } } \\times { \\mathcal { A } } | \\rho ( x ) \\neq 0 \\} .\n$$",
|
| 677 |
+
"text_format": "latex",
|
| 678 |
+
"bbox": [
|
| 679 |
+
238,
|
| 680 |
+
238,
|
| 681 |
+
482,
|
| 682 |
+
256
|
| 683 |
+
],
|
| 684 |
+
"page_idx": 5
|
| 685 |
+
},
|
| 686 |
+
{
|
| 687 |
+
"type": "text",
|
| 688 |
+
"text": "177 Due to the sub-optimality of $\\pi _ { 1 }$ , $\\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } ) \\backslash \\operatorname { s u p p } ( \\rho _ { \\pi _ { 1 } } ) \\neq$ \n178 $\\mathcal { D }$ (see Figure 4b). We call this part the Latent Demonstra \n179 tion, defined as: \n180 Definition 2 (Latent Demonstration). The latent demon \n181 stration $H$ is the set of those $x \\in { \\mathcal { S } } \\times { \\mathcal { A } }$ that belong to the \n182 relative complement of supp $\\left( \\rho _ { \\pi _ { 1 } } \\right)$ in $\\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } )$ : ",
|
| 689 |
+
"bbox": [
|
| 690 |
+
140,
|
| 691 |
+
268,
|
| 692 |
+
549,
|
| 693 |
+
310
|
| 694 |
+
],
|
| 695 |
+
"page_idx": 5
|
| 696 |
+
},
|
| 697 |
+
{
|
| 698 |
+
"type": "image",
|
| 699 |
+
"img_path": "images/294bf72a2bf113fa3af8d4c88f6c5ecc19cb3705bdf2b3bc98dbe4523ba8833b.jpg",
|
| 700 |
+
"image_caption": [
|
| 701 |
+
"Figure 5: t-SNE visualizations of expert demonstrations and collected data of $\\pi _ { 1 }$ under $\\mathcal { O } _ { \\mathrm { E } }$ . "
|
| 702 |
+
],
|
| 703 |
+
"image_footnote": [],
|
| 704 |
+
"bbox": [
|
| 705 |
+
568,
|
| 706 |
+
189,
|
| 707 |
+
833,
|
| 708 |
+
282
|
| 709 |
+
],
|
| 710 |
+
"page_idx": 5
|
| 711 |
+
},
|
| 712 |
+
{
|
| 713 |
+
"type": "text",
|
| 714 |
+
"text": "",
|
| 715 |
+
"bbox": [
|
| 716 |
+
140,
|
| 717 |
+
315,
|
| 718 |
+
549,
|
| 719 |
+
358
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 5
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "equation",
|
| 725 |
+
"img_path": "images/70bd2019c4108041f615fa18e30b1771fa8592d6da6ce2b7f98a3e9e53d8cc06.jpg",
|
| 726 |
+
"text": "$$\nH : = \\{ x \\in S \\times A | \\mathrm { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } ) \\setminus \\mathrm { s u p p } ( \\rho _ { \\pi _ { 1 } } ) \\} .\n$$",
|
| 727 |
+
"text_format": "latex",
|
| 728 |
+
"bbox": [
|
| 729 |
+
349,
|
| 730 |
+
363,
|
| 731 |
+
648,
|
| 732 |
+
381
|
| 733 |
+
],
|
| 734 |
+
"page_idx": 5
|
| 735 |
+
},
|
| 736 |
+
{
|
| 737 |
+
"type": "text",
|
| 738 |
+
"text": "Also, another part of the demonstration is named the Observed Demonstration, defined as: ",
|
| 739 |
+
"bbox": [
|
| 740 |
+
158,
|
| 741 |
+
393,
|
| 742 |
+
766,
|
| 743 |
+
409
|
| 744 |
+
],
|
| 745 |
+
"page_idx": 5
|
| 746 |
+
},
|
| 747 |
+
{
|
| 748 |
+
"type": "text",
|
| 749 |
+
"text": "184 Definition 3 (Observed Demonstration). The observed demonstration $O$ is the set of those $x \\in { \\mathcal { S } } \\times { \\mathcal { A } }$ that belong to the complement of 185 $H$ in $\\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } )$ : ",
|
| 750 |
+
"bbox": [
|
| 751 |
+
151,
|
| 752 |
+
412,
|
| 753 |
+
823,
|
| 754 |
+
443
|
| 755 |
+
],
|
| 756 |
+
"page_idx": 5
|
| 757 |
+
},
|
| 758 |
+
{
|
| 759 |
+
"type": "equation",
|
| 760 |
+
"img_path": "images/6beca7532a9041497e59a5f815a813e18914650a56fff07809770657347e89b8.jpg",
|
| 761 |
+
"text": "$$\nO : = \\{ x \\in \\mathcal { S } \\times A \\vert \\mathrm { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } ) \\cap \\mathrm { s u p p } ( \\rho _ { \\pi _ { 1 } } ) \\} .\n$$",
|
| 762 |
+
"text_format": "latex",
|
| 763 |
+
"bbox": [
|
| 764 |
+
349,
|
| 765 |
+
446,
|
| 766 |
+
648,
|
| 767 |
+
464
|
| 768 |
+
],
|
| 769 |
+
"page_idx": 5
|
| 770 |
+
},
|
| 771 |
+
{
|
| 772 |
+
"type": "text",
|
| 773 |
+
"text": "186 Besides, the data outside of demonstrations should be non-expert data: ",
|
| 774 |
+
"bbox": [
|
| 775 |
+
147,
|
| 776 |
+
477,
|
| 777 |
+
637,
|
| 778 |
+
492
|
| 779 |
+
],
|
| 780 |
+
"page_idx": 5
|
| 781 |
+
},
|
| 782 |
+
{
|
| 783 |
+
"type": "text",
|
| 784 |
+
"text": "187 Definition 4 (Non-Expert Data). The non-expert data $N$ is the set of those $x \\in { \\mathcal { S } } \\times { \\mathcal { A } }$ that out of \n188 $\\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } )$ : ",
|
| 785 |
+
"bbox": [
|
| 786 |
+
151,
|
| 787 |
+
496,
|
| 788 |
+
825,
|
| 789 |
+
523
|
| 790 |
+
],
|
| 791 |
+
"page_idx": 5
|
| 792 |
+
},
|
| 793 |
+
{
|
| 794 |
+
"type": "equation",
|
| 795 |
+
"img_path": "images/93ce994eb5174d28c1170320d9abbbd7cdccf1cbb49810782ed7bc0f1fbdd9e7.jpg",
|
| 796 |
+
"text": "$$\nN : = \\{ x \\in { \\mathcal { S } } \\times A | \\rho _ { \\pi _ { \\mathrm { E } } } ( x ) = 0 \\} .\n$$",
|
| 797 |
+
"text_format": "latex",
|
| 798 |
+
"bbox": [
|
| 799 |
+
388,
|
| 800 |
+
522,
|
| 801 |
+
609,
|
| 802 |
+
539
|
| 803 |
+
],
|
| 804 |
+
"page_idx": 5
|
| 805 |
+
},
|
| 806 |
+
{
|
| 807 |
+
"type": "text",
|
| 808 |
+
"text": "189 In other words, the sub-optimality of $\\pi _ { 1 }$ will cause not only the dynamics mismatch, but also the \n190 appearance of the latent demonstration $H$ . We call the latter one the problem of Support Mismatch. \n191 Intuitively, when $\\pi _ { 2 } \\pi _ { \\mathrm { E } }$ , we have $H \\emptyset$ , monotonously. So in order to fix the support mismatch \n192 between $\\rho _ { \\pi _ { \\mathrm { E } } }$ and $\\rho _ { \\pi _ { 1 } }$ , guiding $\\pi _ { 2 }$ to find out $H$ is the key. ",
|
| 809 |
+
"bbox": [
|
| 810 |
+
142,
|
| 811 |
+
547,
|
| 812 |
+
826,
|
| 813 |
+
606
|
| 814 |
+
],
|
| 815 |
+
"page_idx": 5
|
| 816 |
+
},
|
| 817 |
+
{
|
| 818 |
+
"type": "text",
|
| 819 |
+
"text": "In addition, the support mismatch problem can be viewed as an inverse problem of the Out Of Distribution (OOD) problem that frequently occurred in offline RL setting [21], in which they tried to avoid $\\operatorname { s u p p } ( \\rho _ { \\pi _ { 1 } } ) \\setminus \\operatorname { s u p p } ( \\rho _ { \\pi _ { \\mathrm { E } } } )$ instead. ",
|
| 820 |
+
"bbox": [
|
| 821 |
+
161,
|
| 822 |
+
611,
|
| 823 |
+
825,
|
| 824 |
+
654
|
| 825 |
+
],
|
| 826 |
+
"page_idx": 5
|
| 827 |
+
},
|
| 828 |
+
{
|
| 829 |
+
"type": "text",
|
| 830 |
+
"text": "4.3 Imitation Learning with Rejection ",
|
| 831 |
+
"text_level": 1,
|
| 832 |
+
"bbox": [
|
| 833 |
+
166,
|
| 834 |
+
669,
|
| 835 |
+
450,
|
| 836 |
+
684
|
| 837 |
+
],
|
| 838 |
+
"page_idx": 5
|
| 839 |
+
},
|
| 840 |
+
{
|
| 841 |
+
"type": "text",
|
| 842 |
+
"text": "We can observe that $H \\cup O \\cup N = S \\times { \\mathcal { A } }$ . So it is desirable to filter out $H$ from $O$ and $N$ . Meanwhile, $D _ { w _ { 1 } }$ and $D _ { w _ { 2 } }$ can only classify $O \\cup H$ and $N$ , under $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ respectively. Therefore, here we design two models $g _ { 1 } : \\mathcal { O } _ { \\mathrm { E } } \\times \\mathcal { A } \\{ 0 , 1 \\}$ and $g _ { 2 } : { \\mathcal { O } } _ { \\mathrm { L } } \\times { \\mathcal { A } } \\{ 0 , 1 \\}$ (Output 0: $x \\in O$ and output 1: otherwise), so that given $x \\sim \\tau$ (corresponding $\\widetilde { x } \\sim \\widetilde { \\tau }$ and $\\overline { { x } } \\sim \\overline { { \\mathcal { T } } }$ ) they can satisfy: ",
|
| 843 |
+
"bbox": [
|
| 844 |
+
161,
|
| 845 |
+
694,
|
| 846 |
+
825,
|
| 847 |
+
753
|
| 848 |
+
],
|
| 849 |
+
"page_idx": 5
|
| 850 |
+
},
|
| 851 |
+
{
|
| 852 |
+
"type": "equation",
|
| 853 |
+
"img_path": "images/2b5231b68e3ae021871914277ee2dda09dd070bfbd62df6ef3ce2eaaeb6808f5.jpg",
|
| 854 |
+
"text": "$$\nH = \\{ x \\in S \\times \\mathcal { A } | \\mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \\widetilde { x } ) ] g _ { 1 } ^ { * } ( \\widetilde { x } ) = \\mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \\overline { { x } } ) ] g _ { 2 } ^ { * } ( \\overline { { x } } ) = + 1 \\} ,\n$$",
|
| 855 |
+
"text_format": "latex",
|
| 856 |
+
"bbox": [
|
| 857 |
+
290,
|
| 858 |
+
758,
|
| 859 |
+
704,
|
| 860 |
+
777
|
| 861 |
+
],
|
| 862 |
+
"page_idx": 5
|
| 863 |
+
},
|
| 864 |
+
{
|
| 865 |
+
"type": "text",
|
| 866 |
+
"text": "201",
|
| 867 |
+
"bbox": [
|
| 868 |
+
142,
|
| 869 |
+
786,
|
| 870 |
+
158,
|
| 871 |
+
796
|
| 872 |
+
],
|
| 873 |
+
"page_idx": 5
|
| 874 |
+
},
|
| 875 |
+
{
|
| 876 |
+
"type": "equation",
|
| 877 |
+
"img_path": "images/d0de568a94606d038a9647f5a1945f2b84e3e488c5449ca84ccd66cd0823dc37.jpg",
|
| 878 |
+
"text": "$$\nO = \\{ x \\in \\mathcal { S } \\times \\mathcal { A } | \\mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \\widetilde { x } ) ] g _ { 1 } ^ { * } ( \\widetilde { x } ) = \\mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \\overline { { x } } ) ] g _ { 2 } ^ { * } ( \\overline { { x } } ) = 0 \\} ,\n$$",
|
| 879 |
+
"text_format": "latex",
|
| 880 |
+
"bbox": [
|
| 881 |
+
299,
|
| 882 |
+
795,
|
| 883 |
+
696,
|
| 884 |
+
813
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 5
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "equation",
|
| 890 |
+
"img_path": "images/dd542e38679df4822db6be3b04a064870d25ae2573b7baa367503c7903022c66.jpg",
|
| 891 |
+
"text": "$$\nN = \\{ x \\in S \\times \\mathcal { A } | \\mathbb { I } [ D _ { w _ { 1 } } ^ { * } ( \\widetilde { x } ) ] g _ { 1 } ^ { * } ( \\widetilde { x } ) = \\mathbb { I } [ D _ { w _ { 2 } } ^ { * } ( \\overline { { x } } ) ] g _ { 2 } ^ { * } ( \\overline { { x } } ) = - 1 \\} ,\n$$",
|
| 892 |
+
"text_format": "latex",
|
| 893 |
+
"bbox": [
|
| 894 |
+
290,
|
| 895 |
+
827,
|
| 896 |
+
705,
|
| 897 |
+
845
|
| 898 |
+
],
|
| 899 |
+
"page_idx": 5
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"type": "text",
|
| 903 |
+
"text": "respectively, where $\\mathbb { I } [ \\cdot ]$ takes $+ 1$ if $\\cdot > 0 . 5$ , and $- 1$ otherwise. The target combined model $\\mathbb { I } [ D _ { w } ^ { * } ( x ) ] g ^ { * } ( x )$ is depicted in Figure4c. ",
|
| 904 |
+
"bbox": [
|
| 905 |
+
163,
|
| 906 |
+
848,
|
| 907 |
+
823,
|
| 908 |
+
877
|
| 909 |
+
],
|
| 910 |
+
"page_idx": 5
|
| 911 |
+
},
|
| 912 |
+
{
|
| 913 |
+
"type": "text",
|
| 914 |
+
"text": "205 To this end, both $g _ { 1 }$ and $g _ { 2 }$ should be able to cover $O$ , meanwhile $g _ { 2 }$ can be adaptive to continuously change of 206 $\\rho _ { \\pi _ { 2 } }$ due to the update of $\\pi _ { 2 }$ . Here we learn $g _ { 1 }$ and $g _ { 2 }$ in a rejection form, to reject $O$ from ",
|
| 915 |
+
"bbox": [
|
| 916 |
+
147,
|
| 917 |
+
882,
|
| 918 |
+
825,
|
| 919 |
+
912
|
| 920 |
+
],
|
| 921 |
+
"page_idx": 5
|
| 922 |
+
},
|
| 923 |
+
{
|
| 924 |
+
"type": "text",
|
| 925 |
+
"text": "207 $O \\cup H$ (where $\\mathbb { I } ( D _ { w } ) = + 1$ ). Concretely, the rejection setting is the same as that in Cortes et al. [9]. \n208 Also inspired by Geifman et al. [15], the optimization objective of the combination of $D _ { w }$ and $g$ is ",
|
| 926 |
+
"bbox": [
|
| 927 |
+
138,
|
| 928 |
+
90,
|
| 929 |
+
825,
|
| 930 |
+
121
|
| 931 |
+
],
|
| 932 |
+
"page_idx": 6
|
| 933 |
+
},
|
| 934 |
+
{
|
| 935 |
+
"type": "equation",
|
| 936 |
+
"img_path": "images/60da34f74758b03f1b582cd4d18fc253cbc1f46dbe9c411781517b9e7e46151f.jpg",
|
| 937 |
+
"text": "$$\n\\begin{array} { r } { \\mathcal { L } ( D _ { w } , g ) \\triangleq \\hat { l } ( D _ { w } , g ) + \\lambda \\operatorname* { m a x } ( 0 , c - \\hat { \\phi } ( g ) ) ^ { 2 } , } \\end{array}\n$$",
|
| 938 |
+
"text_format": "latex",
|
| 939 |
+
"bbox": [
|
| 940 |
+
344,
|
| 941 |
+
126,
|
| 942 |
+
650,
|
| 943 |
+
146
|
| 944 |
+
],
|
| 945 |
+
"page_idx": 6
|
| 946 |
+
},
|
| 947 |
+
{
|
| 948 |
+
"type": "text",
|
| 949 |
+
"text": "209 where $c > 0$ denotes the target coverage, and $\\lambda$ denotes the factor for controlling the relative importance of rejection. Besides, the empirical coverage 210 $\\hat { \\phi } ( g )$ is defined as ",
|
| 950 |
+
"bbox": [
|
| 951 |
+
142,
|
| 952 |
+
151,
|
| 953 |
+
825,
|
| 954 |
+
183
|
| 955 |
+
],
|
| 956 |
+
"page_idx": 6
|
| 957 |
+
},
|
| 958 |
+
{
|
| 959 |
+
"type": "equation",
|
| 960 |
+
"img_path": "images/7913d63efc6973629b1c50a23e9a3c7d50db3282b6c6dcd956c923a3aeea0cb8.jpg",
|
| 961 |
+
"text": "$$\n\\hat { \\phi } ( g | X ) \\triangleq \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } g ( x _ { i } ) ,\n$$",
|
| 962 |
+
"text_format": "latex",
|
| 963 |
+
"bbox": [
|
| 964 |
+
416,
|
| 965 |
+
189,
|
| 966 |
+
580,
|
| 967 |
+
231
|
| 968 |
+
],
|
| 969 |
+
"page_idx": 6
|
| 970 |
+
},
|
| 971 |
+
{
|
| 972 |
+
"type": "text",
|
| 973 |
+
"text": "where a batch of data 211 $X = \\{ x _ { i } \\} , i \\in [ m ]$ . The empirical rejection risk $\\hat { l } ( D _ { w } , g )$ is the ratio between 212 the covered risk of the discriminator and the empirical coverage: ",
|
| 974 |
+
"bbox": [
|
| 975 |
+
138,
|
| 976 |
+
238,
|
| 977 |
+
826,
|
| 978 |
+
268
|
| 979 |
+
],
|
| 980 |
+
"page_idx": 6
|
| 981 |
+
},
|
| 982 |
+
{
|
| 983 |
+
"type": "equation",
|
| 984 |
+
"img_path": "images/afe3724859746e013d9106fd23b6641ceedd437ce671fbc624a679ba96479f58.jpg",
|
| 985 |
+
"text": "$$\n\\hat { l } ( D _ { w } , g ) \\triangleq \\frac { \\frac { 1 } { m } \\sum _ { i = 1 } ^ { m } \\langle \\mathcal { L } ( D _ { w } ( x _ { i } ) ) , g ( x _ { i } ) \\rangle } { \\hat { \\phi } ( g ) } .\n$$",
|
| 986 |
+
"text_format": "latex",
|
| 987 |
+
"bbox": [
|
| 988 |
+
357,
|
| 989 |
+
273,
|
| 990 |
+
640,
|
| 991 |
+
313
|
| 992 |
+
],
|
| 993 |
+
"page_idx": 6
|
| 994 |
+
},
|
| 995 |
+
{
|
| 996 |
+
"type": "text",
|
| 997 |
+
"text": "Meanwhile, both $D _ { w _ { 1 } }$ and $g _ { 1 }$ can access $\\rho _ { \\pi _ { \\mathrm { E } } }$ under $\\mathcal { O } _ { \\mathrm { E } }$ directly. So given $\\overline { { x } } \\sim \\overline { { T } } _ { \\pi _ { 2 } }$ under $\\mathcal { O } _ { \\mathrm { L } }$ , once $\\langle \\mathbb { I } ( D _ { w _ { 2 } } ( \\overline { { x } } ) ) , g _ { 2 } ( \\overline { { x } } ) \\rangle = + 1$ , we can query the corresponding observations $\\widetilde { x }$ of $\\textstyle { \\overline { { x } } }$ through OC 2operation and use $\\langle \\mathbb { I } ( D _ { w _ { 1 } } ( \\widetilde { \\boldsymbol { x } } ) ) , g _ { 1 } ( \\widetilde { \\boldsymbol { x } } ) \\rangle$ to calibrate the output of $g _ { 2 }$ and $D _ { w _ { 2 } }$ e. In this way, $g _ { 2 }$ and $D _ { w _ { 2 } }$ e ecan be entangled together and adaptively guide $\\pi _ { 2 }$ to find out the latent demonstrations $H$ under $\\mathcal { O } _ { \\mathrm { L } }$ . ",
|
| 998 |
+
"bbox": [
|
| 999 |
+
140,
|
| 1000 |
+
325,
|
| 1001 |
+
826,
|
| 1002 |
+
383
|
| 1003 |
+
],
|
| 1004 |
+
"page_idx": 6
|
| 1005 |
+
},
|
| 1006 |
+
{
|
| 1007 |
+
"type": "text",
|
| 1008 |
+
"text": "4.4 IWRE ",
|
| 1009 |
+
"text_level": 1,
|
| 1010 |
+
"bbox": [
|
| 1011 |
+
174,
|
| 1012 |
+
391,
|
| 1013 |
+
258,
|
| 1014 |
+
405
|
| 1015 |
+
],
|
| 1016 |
+
"page_idx": 6
|
| 1017 |
+
},
|
| 1018 |
+
{
|
| 1019 |
+
"type": "text",
|
| 1020 |
+
"text": "Here we combine the importance-weighting and rejection into a unified whole, to propose a novel algorithm named Importance Weighting with REjection (IWRE). Concretely, in a HOIL process: ",
|
| 1021 |
+
"bbox": [
|
| 1022 |
+
173,
|
| 1023 |
+
410,
|
| 1024 |
+
821,
|
| 1025 |
+
439
|
| 1026 |
+
],
|
| 1027 |
+
"page_idx": 6
|
| 1028 |
+
},
|
| 1029 |
+
{
|
| 1030 |
+
"type": "text",
|
| 1031 |
+
"text": "Pretraining. We train a discriminator $D _ { w _ { 1 } }$ by Equation (5) and its corresponding rejection model $g _ { 1 }$ by Equation (17) using the initial data and the expert demonstrations. ",
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
169,
|
| 1034 |
+
444,
|
| 1035 |
+
820,
|
| 1036 |
+
473
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 6
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "text",
|
| 1042 |
+
"text": "Training. We train a discriminator $D _ { w _ { 2 } }$ by the combination of Equation (9) and Equation (17), as well as its corresponding rejection model $g _ { 2 }$ by Equation (17), using the initial data, the data collected by $\\pi _ { 2 }$ , and the output of $D _ { w _ { 1 } }$ with $g _ { 1 }$ through OC operation. Also, $\\pi _ { 2 }$ will be updated with $D _ { w _ { 2 } }$ and $g _ { 2 }$ asymmetrically as in GAIL. ",
|
| 1043 |
+
"bbox": [
|
| 1044 |
+
173,
|
| 1045 |
+
478,
|
| 1046 |
+
825,
|
| 1047 |
+
535
|
| 1048 |
+
],
|
| 1049 |
+
"page_idx": 6
|
| 1050 |
+
},
|
| 1051 |
+
{
|
| 1052 |
+
"type": "text",
|
| 1053 |
+
"text": "The pseudo-code of our algorithm is provided in the supplementary material. ",
|
| 1054 |
+
"bbox": [
|
| 1055 |
+
174,
|
| 1056 |
+
540,
|
| 1057 |
+
676,
|
| 1058 |
+
556
|
| 1059 |
+
],
|
| 1060 |
+
"page_idx": 6
|
| 1061 |
+
},
|
| 1062 |
+
{
|
| 1063 |
+
"type": "text",
|
| 1064 |
+
"text": "5 Experiment ",
|
| 1065 |
+
"text_level": 1,
|
| 1066 |
+
"bbox": [
|
| 1067 |
+
173,
|
| 1068 |
+
568,
|
| 1069 |
+
305,
|
| 1070 |
+
584
|
| 1071 |
+
],
|
| 1072 |
+
"page_idx": 6
|
| 1073 |
+
},
|
| 1074 |
+
{
|
| 1075 |
+
"type": "text",
|
| 1076 |
+
"text": "In this section, we validate our algorithm in Atari 2600 [3] (GPL License) and MuJoCo [33] (Academic License) environments. The experiments were designed to investigate: ",
|
| 1077 |
+
"bbox": [
|
| 1078 |
+
166,
|
| 1079 |
+
592,
|
| 1080 |
+
823,
|
| 1081 |
+
621
|
| 1082 |
+
],
|
| 1083 |
+
"page_idx": 6
|
| 1084 |
+
},
|
| 1085 |
+
{
|
| 1086 |
+
"type": "text",
|
| 1087 |
+
"text": "1) Can IWRE achieve significant performance under HOIL tasks? \n2) Can IWRE deal with the support mismatch problem? \n3) During training, is active querying for HOIL indeed necessary? ",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
210,
|
| 1090 |
+
623,
|
| 1091 |
+
648,
|
| 1092 |
+
676
|
| 1093 |
+
],
|
| 1094 |
+
"page_idx": 6
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"type": "text",
|
| 1098 |
+
"text": "Below we first introduce the experimental setup and then investigate the above questions. More results and experimental details are included in the supplementary material. ",
|
| 1099 |
+
"bbox": [
|
| 1100 |
+
174,
|
| 1101 |
+
686,
|
| 1102 |
+
823,
|
| 1103 |
+
715
|
| 1104 |
+
],
|
| 1105 |
+
"page_idx": 6
|
| 1106 |
+
},
|
| 1107 |
+
{
|
| 1108 |
+
"type": "text",
|
| 1109 |
+
"text": "5.1 Experimental Setup ",
|
| 1110 |
+
"text_level": 1,
|
| 1111 |
+
"bbox": [
|
| 1112 |
+
174,
|
| 1113 |
+
724,
|
| 1114 |
+
352,
|
| 1115 |
+
739
|
| 1116 |
+
],
|
| 1117 |
+
"page_idx": 6
|
| 1118 |
+
},
|
| 1119 |
+
{
|
| 1120 |
+
"type": "text",
|
| 1121 |
+
"text": "Environments. We choose three pixel-memory based games in Atari and five continuous control objects in MuJoCo on OpenAI platform [5] (MIT License). Details as below: ",
|
| 1122 |
+
"bbox": [
|
| 1123 |
+
163,
|
| 1124 |
+
742,
|
| 1125 |
+
823,
|
| 1126 |
+
771
|
| 1127 |
+
],
|
| 1128 |
+
"page_idx": 6
|
| 1129 |
+
},
|
| 1130 |
+
{
|
| 1131 |
+
"type": "text",
|
| 1132 |
+
"text": "1. Pixel-memory Atari games. $\\mathcal { O } _ { \\mathrm { E } }$ : $8 4 \\times 8 4 \\times 4$ raw pixels; $\\mathcal { O } _ { \\mathrm { L } }$ : 128-byte random access memories (RAM). Expert: converged DQN-based agents [24]. Atari games contain two totally isolated views: raw pixels and RAM, under the same state. Through these environments, we want to investigate whether the agent can learn an effective policy from demonstrations under completely different observation spaces. Moreover, IL with visual observations only is already very difficult [7], while learning a RAM-based policy can be even more challenging [3, 31], so few $\\mathrm { I L }$ research reported desirable results on this task. ",
|
| 1133 |
+
"bbox": [
|
| 1134 |
+
214,
|
| 1135 |
+
781,
|
| 1136 |
+
825,
|
| 1137 |
+
878
|
| 1138 |
+
],
|
| 1139 |
+
"page_idx": 6
|
| 1140 |
+
},
|
| 1141 |
+
{
|
| 1142 |
+
"type": "text",
|
| 1143 |
+
"text": "2. Continuous control MuJoCo objects. $\\mathcal { O } _ { \\mathrm { E } }$ : half of original observation features; $\\mathcal { O } _ { \\mathrm { L } }$ : another half of original observation features. Expert: converged DDPG-based agents [22]. The features of MuJoCo contain monotonous information like the direction, position, velocity, etc., of an object. Through these environments, we want to investigate whether the agent can learn from demonstrations with complementary signals under observations with missing information. Meanwhile, we make sure RL algorithms can obtain comparable performances under $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ . More details are reported in the supplementary material. ",
|
| 1144 |
+
"bbox": [
|
| 1145 |
+
212,
|
| 1146 |
+
882,
|
| 1147 |
+
825,
|
| 1148 |
+
911
|
| 1149 |
+
],
|
| 1150 |
+
"page_idx": 6
|
| 1151 |
+
},
|
| 1152 |
+
{
|
| 1153 |
+
"type": "image",
|
| 1154 |
+
"img_path": "images/f5ce93ddf7cef3cc04f4113541e8e8daeeef8bbe14b00517c1c3ac7639dfe84e.jpg",
|
| 1155 |
+
"image_caption": [
|
| 1156 |
+
"Figure 6: The learning curves of each method, where the shaded region indicates the standard deviation. "
|
| 1157 |
+
],
|
| 1158 |
+
"image_footnote": [],
|
| 1159 |
+
"bbox": [
|
| 1160 |
+
173,
|
| 1161 |
+
111,
|
| 1162 |
+
826,
|
| 1163 |
+
348
|
| 1164 |
+
],
|
| 1165 |
+
"page_idx": 7
|
| 1166 |
+
},
|
| 1167 |
+
{
|
| 1168 |
+
"type": "text",
|
| 1169 |
+
"text": "",
|
| 1170 |
+
"bbox": [
|
| 1171 |
+
230,
|
| 1172 |
+
393,
|
| 1173 |
+
823,
|
| 1174 |
+
463
|
| 1175 |
+
],
|
| 1176 |
+
"page_idx": 7
|
| 1177 |
+
},
|
| 1178 |
+
{
|
| 1179 |
+
"type": "text",
|
| 1180 |
+
"text": "252 Besides, twenty expert trajectories were collected for each environment. Each result contains five \n253 trials with different random seeds. All experiments were conducted on server clusters with NVIDIA \n254 Tesla V100 GPUs. The summary of the environments is gathered in the supplementary material. \n255 Baselines. Six basic contenders were included in the experiments: Vanilla GAIL [16], GAIL \n256 with importance-weighting [12] (IW), third-person IL [30] (TPIL), generative adversarial MDP \n257 alignment [19] (GAMA), behavioral cloning [2] (BC), and learning by cheating [8] (LBC). For \n258 IW, we utilized the discriminator $D _ { w _ { 1 } }$ trained in the pretraining to calculate the importance weight; \n259 also the optimization objective for $D _ { w _ { 2 } }$ during training is the same as Equation (9); TPIL learns the \n260 third-person demonstrations by leading the cross-entropy loss into the update of the feature extractor; \n261 GAMA learns a mapping function $\\psi$ in view of adversarial training to align the observation of the \n262 target domain into the source domain, and thereby can utilize the policy in the source domain for \n263 zero-shot imitation. For fairness, we allowed the interaction between the policy and the environment \n264 for GAMA under HOIL; LBC uses $\\pi _ { 1 }$ learned from privileged states as a teacher to train $\\pi _ { 2 }$ in a \n265 DAgger [27] style, so here we allowed LBC to access $\\mathcal { O } _ { \\mathrm { E } }$ during the whole IL process. In Atari, to \n266 investigate whether our method could achieve good performance for RAM-based control, we further \n267 included a contender PPO-RAM, which uses proximal policy optimization (PPO) [28] to perform \n268 RL directly with environmental true rewards under the RAM-based observations. More detailed \n269 setup including query strategies for TPIL and GAMA, network architecture, and hyper-parameters \n270 are reported in the supplementary material. ",
|
| 1181 |
+
"bbox": [
|
| 1182 |
+
147,
|
| 1183 |
+
474,
|
| 1184 |
+
826,
|
| 1185 |
+
517
|
| 1186 |
+
],
|
| 1187 |
+
"page_idx": 7
|
| 1188 |
+
},
|
| 1189 |
+
{
|
| 1190 |
+
"type": "text",
|
| 1191 |
+
"text": "",
|
| 1192 |
+
"bbox": [
|
| 1193 |
+
143,
|
| 1194 |
+
523,
|
| 1195 |
+
825,
|
| 1196 |
+
743
|
| 1197 |
+
],
|
| 1198 |
+
"page_idx": 7
|
| 1199 |
+
},
|
| 1200 |
+
{
|
| 1201 |
+
"type": "text",
|
| 1202 |
+
"text": "Learning process. To simulate the situation that $\\mathcal { O } _ { \\mathrm { E } }$ is costly, the steps for training $\\pi _ { 1 }$ was set as 1/4 of that for training $\\pi _ { 2 }$ , using GAIL [16]/HashReward [7] under the $\\mathcal { O } _ { \\mathrm { E } }$ space for MuJoCo/Atari environments. The learning steps were $1 0 ^ { 7 }$ for MuJoCo and $5 \\times 1 0 ^ { 6 }$ for Atari environments. In the pretraining, we sampled 20 trajectories from $\\pi _ { 1 }$ , and the data from each trajectory had both $\\mathcal { O } _ { \\mathrm { E } }$ and $\\mathcal { O } _ { \\mathrm { L } }$ observations. In the training, each method learned $4 \\times 1 0 ^ { 7 }$ steps for MuJoCo and $2 \\times 1 0 ^ { 7 }$ steps for Atari under the $\\mathcal { O } _ { \\mathrm { L } }$ space to obtain $\\pi _ { 2 }$ . ",
|
| 1203 |
+
"bbox": [
|
| 1204 |
+
171,
|
| 1205 |
+
751,
|
| 1206 |
+
823,
|
| 1207 |
+
833
|
| 1208 |
+
],
|
| 1209 |
+
"page_idx": 7
|
| 1210 |
+
},
|
| 1211 |
+
{
|
| 1212 |
+
"type": "text",
|
| 1213 |
+
"text": "5.2 Results ",
|
| 1214 |
+
"text_level": 1,
|
| 1215 |
+
"bbox": [
|
| 1216 |
+
173,
|
| 1217 |
+
852,
|
| 1218 |
+
264,
|
| 1219 |
+
864
|
| 1220 |
+
],
|
| 1221 |
+
"page_idx": 7
|
| 1222 |
+
},
|
| 1223 |
+
{
|
| 1224 |
+
"type": "text",
|
| 1225 |
+
"text": "Experimental results are reported in Figure 6. Since the mapping function is hard to learn when input is RAM and output is raw images, we omit the results of GAMA in Atari. We can observe that while IW is better than GAIL in most environments, both GAIL and IW can hardly outperform $\\pi _ { 1 }$ ",
|
| 1226 |
+
"bbox": [
|
| 1227 |
+
173,
|
| 1228 |
+
869,
|
| 1229 |
+
825,
|
| 1230 |
+
911
|
| 1231 |
+
],
|
| 1232 |
+
"page_idx": 7
|
| 1233 |
+
},
|
| 1234 |
+
{
|
| 1235 |
+
"type": "text",
|
| 1236 |
+
"text": "Because they just imitated the performance of $\\pi _ { 1 }$ instead of $\\pi _ { \\mathrm { E } }$ , even with importance-weighting for calibration. For TPIL, its learning process was extremely unstable on Hopper, Swimmer, and Walker2d due to the continuous distribution shift. Furthermore, the performance of GAMA was not satisfactory in Hopper and Walker2d because its mapping function is hard to learn well when the support mismatch appears. The results of TPIL and GAMA demonstrate that DSIL methods will be invalid under heterogeneous observations as in HOIL tasks. On Atari environments, $\\mathcal { O } _ { \\mathrm { E } }$ contains more privileged information than $\\mathcal { O } _ { \\mathrm { L } }$ , so LBC can achieve good performance. But when $\\mathcal { O } _ { \\mathrm { E } }$ is not more privileged than $\\mathcal { O } _ { \\mathrm { L } }$ , like in most environments of MuJoCo, its performance will decrease due to the support mismatch, which would make it even worse than BC. Finally, IWRE obtained the best performance on 6/8 environments, and comparable performance with LBC on Reacher, which shows the effectiveness of our method even with limited access to $\\mathcal { O } _ { \\mathrm { E } }$ (LBC can access to $\\mathcal { O } _ { \\mathrm { E } }$ all the time). Besides, we can see that the performance differences between the GAIL/IW and IWRE/TPIL/GAMA/LBC are huge (especially on Reacher) because of the absence of queries, which demonstrates that the query operation is indeed necessary for HOIL problems. ",
|
| 1237 |
+
"bbox": [
|
| 1238 |
+
138,
|
| 1239 |
+
90,
|
| 1240 |
+
825,
|
| 1241 |
+
285
|
| 1242 |
+
],
|
| 1243 |
+
"page_idx": 8
|
| 1244 |
+
},
|
| 1245 |
+
{
|
| 1246 |
+
"type": "text",
|
| 1247 |
+
"text": "Moreover, even learned with true rewards, PPO-RAM surprisingly failed to achieve comparable performance to IWRE, which shows that IWRE could possibly learn more effective rewards than true environmental rewards in RAM-input tasks. The results verify that, IWRE provides a powerful approach for tackling HOIL problems, even under the situation that the demonstrations are gathered from such a different observation space, meanwhile $\\mathcal { O } _ { \\mathrm { E } }$ is strictly limited during training. ",
|
| 1248 |
+
"bbox": [
|
| 1249 |
+
140,
|
| 1250 |
+
291,
|
| 1251 |
+
825,
|
| 1252 |
+
361
|
| 1253 |
+
],
|
| 1254 |
+
"page_idx": 8
|
| 1255 |
+
},
|
| 1256 |
+
{
|
| 1257 |
+
"type": "text",
|
| 1258 |
+
"text": "t-SNE visualization of $\\rho _ { \\pi _ { 2 } }$ and $\\rho _ { \\pi _ { \\mathrm { E } } }$ under $\\mathcal { O } _ { \\mathrm { E } }$ . In Section 4.2, we point that the sub-optimality of $\\pi _ { 1 }$ will cause the problem of support mismatch, which is embodied as the appearance of the latent demonstration $H$ during training. Also the empirical results in Figure 5 on Hopper and Walker2d verify the existence of this problem. So we want to investigate whether the superiority of IWRE indeed comes from successfully tackling the support mismatch problem. To this end, we plotted the t-SNE visualization of the same expert demonstrations as in Section 4.2 and the collected data of $\\pi _ { 2 }$ by IWRE under $\\mathcal { O } _ { \\mathrm { E } }$ ( $\\mathcal { O } _ { \\mathrm { E } }$ is hidden to $\\pi _ { 2 }$ ). All setups are the same as in Section 4.2. From the results shown in Figure 7, we can see that even under $\\mathcal { O } _ { \\mathrm { E } }$ which cannot be obtained by $\\pi _ { 2 }$ , almost all high-density regions of the demonstrations were covered by the collected data. Meanwhile, the latent demonstration $H$ is dug out nearly. The results demonstrate that IWRE basically solves the problem of support mismatch and thereby performs well in these environments. ",
|
| 1259 |
+
"bbox": [
|
| 1260 |
+
174,
|
| 1261 |
+
367,
|
| 1262 |
+
549,
|
| 1263 |
+
587
|
| 1264 |
+
],
|
| 1265 |
+
"page_idx": 8
|
| 1266 |
+
},
|
| 1267 |
+
{
|
| 1268 |
+
"type": "image",
|
| 1269 |
+
"img_path": "images/39284efc6bf18c77ace181d7b3be7faa19ca73e478ef3101cd9225f50036f763.jpg",
|
| 1270 |
+
"image_caption": [
|
| 1271 |
+
"Figure 7: t-SNE visualizations of expert demonstrations and collected data of $\\pi _ { 2 }$ under $\\mathcal { O } _ { \\mathrm { E } }$ . The high-density regions of the expert demonstrations were covered by the collected data of $\\pi _ { 2 }$ of IWRE. "
|
| 1272 |
+
],
|
| 1273 |
+
"image_footnote": [],
|
| 1274 |
+
"bbox": [
|
| 1275 |
+
568,
|
| 1276 |
+
397,
|
| 1277 |
+
833,
|
| 1278 |
+
491
|
| 1279 |
+
],
|
| 1280 |
+
"page_idx": 8
|
| 1281 |
+
},
|
| 1282 |
+
{
|
| 1283 |
+
"type": "text",
|
| 1284 |
+
"text": "",
|
| 1285 |
+
"bbox": [
|
| 1286 |
+
173,
|
| 1287 |
+
588,
|
| 1288 |
+
823,
|
| 1289 |
+
614
|
| 1290 |
+
],
|
| 1291 |
+
"page_idx": 8
|
| 1292 |
+
},
|
| 1293 |
+
{
|
| 1294 |
+
"type": "text",
|
| 1295 |
+
"text": "Besides, some collected data of $\\pi _ { 2 }$ of IWRE were out of the distribution of the demonstrations, which means $\\pi _ { 2 }$ slightly overly explored the environment. Since $\\mathcal { O } _ { \\mathrm { E } }$ is hidden to $\\pi _ { 2 }$ , the reward function will encourage $\\pi _ { 2 }$ to explore more areas to fix the support mismatch problem. Meanwhile, the out-of-distribution problem in HOIL is not as severe as in the offline RL settings [21], so this over-exploration phenomenon makes sense. ",
|
| 1296 |
+
"bbox": [
|
| 1297 |
+
173,
|
| 1298 |
+
622,
|
| 1299 |
+
825,
|
| 1300 |
+
690
|
| 1301 |
+
],
|
| 1302 |
+
"page_idx": 8
|
| 1303 |
+
},
|
| 1304 |
+
{
|
| 1305 |
+
"type": "text",
|
| 1306 |
+
"text": "6 Conclusion ",
|
| 1307 |
+
"text_level": 1,
|
| 1308 |
+
"bbox": [
|
| 1309 |
+
173,
|
| 1310 |
+
707,
|
| 1311 |
+
299,
|
| 1312 |
+
723
|
| 1313 |
+
],
|
| 1314 |
+
"page_idx": 8
|
| 1315 |
+
},
|
| 1316 |
+
{
|
| 1317 |
+
"type": "text",
|
| 1318 |
+
"text": "In this paper, we proposed a new learning framework named Heterogeneously Observable Imitation Learning (HOIL), to formulate the situations where the observation space of demonstrations is different from that of the imitator while learning. We formally modeled a learning process of HOIL, in which the access to the observations of an expert is limited due to the high cost. Furthermore, we analyzed underlying challenges during training, i.e., the dynamics mismatch and the support mismatch, on the occupancy distributions between the demonstrations and the policy. To tackle these challenges, we proposed a new algorithm named Importance Weighting with REjection (IWRE), using importance-weighting and learning with rejection. Experimental results showed that the direct imitation and domain adaptive methods could not solve this problem, while our approach obtained promising results. In the future, we hope to involve the theoretical guarantee for our algorithm IWRE and investigate how many $\\mathcal { O } _ { \\mathrm { E } }$ do we need to query to learn a promising $\\pi _ { 2 }$ . Furthermore, we hope to use the learning framework of HOIL and IWRE to tackle more learning scenarios with demonstrations in different spaces. ",
|
| 1319 |
+
"bbox": [
|
| 1320 |
+
171,
|
| 1321 |
+
732,
|
| 1322 |
+
825,
|
| 1323 |
+
911
|
| 1324 |
+
],
|
| 1325 |
+
"page_idx": 8
|
| 1326 |
+
},
|
| 1327 |
+
{
|
| 1328 |
+
"type": "text",
|
| 1329 |
+
"text": "337 References [1] Pieter Abbeel and Andrew Y. Ng. Inverse reinforcement learning. In Encyclopedia of Machine Learning, pages 554–558. 2010. [2] Michael Bain and Claude Sammut. A framework for behavioural cloning. In Machine Intelligence 15, pages 103–129, 1996. [3] Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents. J. Artif. Intell. Res., 47:253–279, 2013. [4] Kianté Brantley, Hal Daumé III, and Amr Sharaf. Active imitation learning with noisy guidance. In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel R. Tetreault, editors, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 2093–2105. Association for Computational Linguistics, 2020. [5] Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. CoRR, abs/1606.01540, 2016. [6] Alberto Broggi, Michele Buzzoni, Stefano Debattisti, Paolo Grisleri, Maria Chiara Laghi, Paolo Medici, and Pietro Versari. Extensive tests of autonomous driving technologies. IEEE Trans. Intell. Transp. Syst., 14(3):1403–1415, 2013. [7] Xin-Qiang Cai, Yao-Xiang Ding, Yuan Jiang, and Zhi-Hua Zhou. Imitation learning from pixel-level demonstrations by hashreward. In Proceedings of the 20th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), page 279–287, 2021. [8] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krähenbühl. Learning by cheating. In Leslie Pack Kaelbling, Danica Kragic, and Komei Sugiura, editors, 3rd Annual Conference on Robot Learning, CoRL 2019, Osaka, Japan, October 30 - November 1, 2019, Proceedings, volume 100 of Proceedings of Machine Learning Research, pages 66–75. PMLR, 2019. [9] Corinna Cortes, Giulia DeSalvo, and Mehryar Mohri. Learning with rejection. In Ronald Ortner, Hans Ulrich Simon, and Sandra Zilles, editors, Algorithmic Learning Theory - 27th International Conference, ALT 2016, Bari, Italy, October 19-21, 2016, Proceedings, volume 9925 of Lecture Notes in Computer Science, pages 67–82, 2016. [10] Mark Cutler, Thomas J. Walsh, and Jonathan P. How. Reinforcement learning with multi-fidelity simulators. In 2014 IEEE International Conference on Robotics and Automation, ICRA 2014, Hong Kong, China, May 31 - June 7, 2014, pages 3888–3895. IEEE, 2014. [11] Siddharth Desai, Ishan Durugkar, Haresh Karnan, Garrett Warnell, Josiah Hanna, and Peter Stone. An imitation from observation approach to transfer learning with dynamics mismatch. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. [12] Tongtong Fang, Nan Lu, Gang Niu, and Masashi Sugiyama. Rethinking importance weighting for deep learning under distribution shift. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. [13] Justin Fu, Katie Luo, and Sergey Levine. Learning robust rewards with adverserial inverse reinforcement learning. In International Conference on Learning Representations, 2018. [14] Tanmay Gangwani, Joel Lehman, Qiang Liu, and Jian Peng. Learning belief representations for imitation learning in pomdps. In Amir Globerson and Ricardo Silva, editors, Proceedings of the Thirty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI 2019, Tel Aviv, Israel, July 22-25, 2019, volume 115 of Proceedings of Machine Learning Research, pages 1061–1071. AUAI Press, 2019. ",
|
| 1330 |
+
"bbox": [
|
| 1331 |
+
151,
|
| 1332 |
+
37,
|
| 1333 |
+
826,
|
| 1334 |
+
920
|
| 1335 |
+
],
|
| 1336 |
+
"page_idx": 9
|
| 1337 |
+
},
|
| 1338 |
+
{
|
| 1339 |
+
"type": "text",
|
| 1340 |
+
"text": "385 [15] Yonatan Geifman and Ran El-Yaniv. Selectivenet: A deep neural network with an integrated \n386 reject option. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the \n387 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, \n388 California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2151–2159. \n389 PMLR, 2019. \n390 [16] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. In Advances \n391 in Neural Information Processing Systems 29: Annual Conference on Neural Information \n392 Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 4565–4573, 2016. \n393 [17] Shengyi Jiang, Jing-Cheng Pang, and Yang Yu. Offline imitation learning with a misspecified \n394 simulator. In Advances in Neural Information Processing Systems 33: Annual Conference on \n395 Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, \n396 2020. \n397 [18] Kun Ho Kim, Yihong Gu, Jiaming Song, Shengjia Zhao, and Stefano Ermon. Cross domain \n398 imitation learning. CoRR, abs/1910.00105, 2019. \n399 [19] Kuno Kim, Yihong Gu, Jiaming Song, Shengjia Zhao, and Stefano Ermon. Domain adaptive \n400 imitation learning. In Proceedings of the 37th International Conference on Machine Learning, \n401 ICML 2020, 13-18 July 2020, Virtual Event, pages 5286–5295, 2020. \n402 [20] Bangalore Ravi Kiran, Ibrahim Sobh, Victor Talpaert, Patrick Mannion, Ahmad A. Al Sallab, \n403 Senthil Kumar Yogamani, and Patrick Pérez. Deep reinforcement learning for autonomous \n404 driving: A survey. CoRR, abs/2002.00444, 2020. \n405 [21] Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: \n406 Tutorial, review, and perspectives on open problems. CoRR, abs/2005.01643, 2020. \n407 [22] Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval \n408 Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. \n409 In Yoshua Bengio and Yann LeCun, editors, 4th International Conference on Learning Repre \n410 sentations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings, \n411 2016. \n412 [23] Yuxuan Liu, Abhishek Gupta, Pieter Abbeel, and Sergey Levine. Imitation from observation: \n13 Learning to imitate behaviors from raw video via context translation. In 2018 IEEE International \n414 Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May 21-25, 2018, \n415 pages 1118–1125, 2018. \n416 [24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan \n417 Wierstra, and Martin A. Riedmiller. Playing atari with deep reinforcement learning. CoRR, \n18 abs/1312.5602, 2013. \n419 [25] Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan P. How, and John Vian. Deep \n420 decentralized multi-task multi-agent reinforcement learning under partial observability. In \n421 Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, \n422 NSW, Australia, 6-11 August 2017, pages 2681–2690, 2017. \n423 [26] Dripta S. Raychaudhuri, Sujoy Paul, Jeroen van Baar, and Amit K. Roy-Chowdhury. Cross \n424 domain imitation from observations. In Marina Meila and Tong Zhang, editors, Proceedings of \n425 the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual \n426 Event, volume 139 of Proceedings of Machine Learning Research, pages 8902–8912. PMLR, \n427 2021. \n428 [27] Stéphane Ross, Geoffrey J. Gordon, and Drew Bagnell. A reduction of imitation learning and \n429 structured prediction to no-regret online learning. In Proceedings of the Fourteenth International \n430 Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April \n431 11-13, 2011, pages 627–635, 2011. \n432 [28] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal \n433 policy optimization algorithms. CoRR, abs/1707.06347, 2017. \n434 [29] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, and \n435 Sergey Levine. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE \n436 International Conference on Robotics and Automation, ICRA 2018, Brisbane, Australia, May \n437 21-25, 2018, pages 1134–1141, 2018. \n438 [30] Bradly C. Stadie, Pieter Abbeel, and Ilya Sutskever. Third person imitation learning. In 5th \n439 International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, \n440 2017, Conference Track Proceedings. OpenReview.net, 2017. \n441 [31] Jakub Sygnowski and Henryk Michalewski. Learning from the memory of atari 2600. In Tristan \n442 Cazenave, Mark H. M. Winands, Stefan Edelkamp, Stephan Schiffel, Michael Thielscher, and \n443 Julian Togelius, editors, Computer Games - 5th Workshop on Computer Games, CGW 2016, \n444 and 5th Workshop on General Intelligence in Game-Playing Agents, GIGA 2016, Held in \n445 Conjunction with the 25th International Conference on Artificial Intelligence, IJCAI 2016, New \n446 York City, NY, USA, July 9-10, 2016, Revised Selected Papers, volume 705 of Communications \n447 in Computer and Information Science, pages 71–85, 2016. \n48 [32] Andrea Tirinzoni, Andrea Sessa, Matteo Pirotta, and Marcello Restelli. Importance weighted \n449 transfer of samples in reinforcement learning. In Jennifer G. Dy and Andreas Krause, editors, \n450 Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stock \n451 holmsmässan, Stockholm, Sweden, July 10-15, 2018, volume 80 of Proceedings of Machine \n452 Learning Research, pages 4943–4952. PMLR, 2018. \n53 [33] Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based \n454 control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS \n455 2012, Vilamoura, Algarve, Portugal, October 7-12, 2012, pages 5026–5033, 2012. \n456 [34] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine \n457 Learning Research, 9:2579–2605, 2008. \n58 [35] Ruohan Wang, Carlo Ciliberto, Pierluigi Vito Amadori, and Yiannis Demiris. Random expert \n459 distillation: Imitation learning via expert policy support estimation. In Kamalika Chaudhuri and \n460 Ruslan Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine \n461 Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings \n462 of Machine Learning Research, pages 6536–6544. PMLR, 2019. \n463 [36] Andrew Warrington, Jonathan Wilder Lavington, Adam Scibior, Mark Schmidt, and Frank ´ \n464 Wood. Robust asymmetric learning in pomdps. In Marina Meila and Tong Zhang, editors, \n465 Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 \n466 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages \n467 11013–11023. PMLR, 2021. ",
|
| 1341 |
+
"bbox": [
|
| 1342 |
+
148,
|
| 1343 |
+
92,
|
| 1344 |
+
828,
|
| 1345 |
+
916
|
| 1346 |
+
],
|
| 1347 |
+
"page_idx": 10
|
| 1348 |
+
},
|
| 1349 |
+
{
|
| 1350 |
+
"type": "text",
|
| 1351 |
+
"text": "",
|
| 1352 |
+
"bbox": [
|
| 1353 |
+
150,
|
| 1354 |
+
87,
|
| 1355 |
+
826,
|
| 1356 |
+
635
|
| 1357 |
+
],
|
| 1358 |
+
"page_idx": 11
|
| 1359 |
+
},
|
| 1360 |
+
{
|
| 1361 |
+
"type": "text",
|
| 1362 |
+
"text": "468 Checklist ",
|
| 1363 |
+
"text_level": 1,
|
| 1364 |
+
"bbox": [
|
| 1365 |
+
142,
|
| 1366 |
+
656,
|
| 1367 |
+
254,
|
| 1368 |
+
672
|
| 1369 |
+
],
|
| 1370 |
+
"page_idx": 11
|
| 1371 |
+
},
|
| 1372 |
+
{
|
| 1373 |
+
"type": "text",
|
| 1374 |
+
"text": "1. For all authors... ",
|
| 1375 |
+
"bbox": [
|
| 1376 |
+
214,
|
| 1377 |
+
684,
|
| 1378 |
+
339,
|
| 1379 |
+
696
|
| 1380 |
+
],
|
| 1381 |
+
"page_idx": 11
|
| 1382 |
+
},
|
| 1383 |
+
{
|
| 1384 |
+
"type": "text",
|
| 1385 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] See Section 6. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] See supplementary material. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 1386 |
+
"bbox": [
|
| 1387 |
+
238,
|
| 1388 |
+
702,
|
| 1389 |
+
825,
|
| 1390 |
+
808
|
| 1391 |
+
],
|
| 1392 |
+
"page_idx": 11
|
| 1393 |
+
},
|
| 1394 |
+
{
|
| 1395 |
+
"type": "text",
|
| 1396 |
+
"text": "2. If you are including theoretical results... ",
|
| 1397 |
+
"bbox": [
|
| 1398 |
+
214,
|
| 1399 |
+
810,
|
| 1400 |
+
493,
|
| 1401 |
+
825
|
| 1402 |
+
],
|
| 1403 |
+
"page_idx": 11
|
| 1404 |
+
},
|
| 1405 |
+
{
|
| 1406 |
+
"type": "text",
|
| 1407 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
|
| 1408 |
+
"bbox": [
|
| 1409 |
+
236,
|
| 1410 |
+
829,
|
| 1411 |
+
738,
|
| 1412 |
+
861
|
| 1413 |
+
],
|
| 1414 |
+
"page_idx": 11
|
| 1415 |
+
},
|
| 1416 |
+
{
|
| 1417 |
+
"type": "text",
|
| 1418 |
+
"text": "3. If you ran experiments... ",
|
| 1419 |
+
"bbox": [
|
| 1420 |
+
214,
|
| 1421 |
+
864,
|
| 1422 |
+
393,
|
| 1423 |
+
878
|
| 1424 |
+
],
|
| 1425 |
+
"page_idx": 11
|
| 1426 |
+
},
|
| 1427 |
+
{
|
| 1428 |
+
"type": "text",
|
| 1429 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See Section 5. ",
|
| 1430 |
+
"bbox": [
|
| 1431 |
+
236,
|
| 1432 |
+
883,
|
| 1433 |
+
825,
|
| 1434 |
+
911
|
| 1435 |
+
],
|
| 1436 |
+
"page_idx": 11
|
| 1437 |
+
},
|
| 1438 |
+
{
|
| 1439 |
+
"type": "text",
|
| 1440 |
+
"text": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See supplementary material. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes] See Section 5. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Section 5. ",
|
| 1441 |
+
"bbox": [
|
| 1442 |
+
236,
|
| 1443 |
+
92,
|
| 1444 |
+
825,
|
| 1445 |
+
180
|
| 1446 |
+
],
|
| 1447 |
+
"page_idx": 12
|
| 1448 |
+
},
|
| 1449 |
+
{
|
| 1450 |
+
"type": "text",
|
| 1451 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1452 |
+
"bbox": [
|
| 1453 |
+
212,
|
| 1454 |
+
184,
|
| 1455 |
+
823,
|
| 1456 |
+
199
|
| 1457 |
+
],
|
| 1458 |
+
"page_idx": 12
|
| 1459 |
+
},
|
| 1460 |
+
{
|
| 1461 |
+
"type": "text",
|
| 1462 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [Yes] \n(c) Did you include any new assets either in the supplemental material or as a URL? [N/A] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [N/A] ",
|
| 1463 |
+
"bbox": [
|
| 1464 |
+
238,
|
| 1465 |
+
203,
|
| 1466 |
+
825,
|
| 1467 |
+
324
|
| 1468 |
+
],
|
| 1469 |
+
"page_idx": 12
|
| 1470 |
+
},
|
| 1471 |
+
{
|
| 1472 |
+
"type": "text",
|
| 1473 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1474 |
+
"bbox": [
|
| 1475 |
+
215,
|
| 1476 |
+
328,
|
| 1477 |
+
705,
|
| 1478 |
+
343
|
| 1479 |
+
],
|
| 1480 |
+
"page_idx": 12
|
| 1481 |
+
},
|
| 1482 |
+
{
|
| 1483 |
+
"type": "text",
|
| 1484 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
|
| 1485 |
+
"bbox": [
|
| 1486 |
+
238,
|
| 1487 |
+
347,
|
| 1488 |
+
825,
|
| 1489 |
+
436
|
| 1490 |
+
],
|
| 1491 |
+
"page_idx": 12
|
| 1492 |
+
}
|
| 1493 |
+
]
|
parse/dev/4WgqjmYacAf/4WgqjmYacAf_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6res1KC1Z3Z/6res1KC1Z3Z_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/6res1KC1Z3Z/6res1KC1Z3Z_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ.md
ADDED
|
@@ -0,0 +1,619 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PSEUDOINVERSE-GUIDED DIFFUSION MODELS FOR INVERSE PROBLEMS
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
Diffusion models have become competitive candidates for solving various inverse problems. Models trained for specific inverse problems work well but are limited to their particular use cases, whereas methods that use problem-agnostic models are general but often perform worse empirically. To address this dilemma, we introduce Pseudoinverse-guided Diffusion Models (ΠGDM), an approach that uses problem-agnostic models to close the gap in performance. ΠGDM directly estimates conditional scores from the measurement model of the inverse problem without additional training. It can address inverse problems with noisy, non-linear, or even non-differentiable measurements, in contrast to many existing approaches that are limited to noiseless linear ones. We illustrate the empirical effectiveness of ΠGDM on several image restoration tasks, including super-resolution, inpainting and JPEG restoration. On ImageNet, ΠGDM is competitive with state-of-the-art diffusion models trained on specific tasks, and is the first to achieve this with problem-agnostic diffusion models. ΠGDM can also solve a wider set of inverse problems where the measurement processes are composed of several simpler ones.
|
| 8 |
+
|
| 9 |
+
# 1 INTRODUCTION
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
Figure 1: High-level illustration of ΠGDM. $( T o p )$ Problem-agnostic diffusion models perform an iterative denoising operation to produce random samples. (Bottom) ΠGDM utilizes problem-agnostic diffusion models to solve inverse problems, a key component of which is pseudoinverse guidance (ΠG). ΠG converts the problem-agnostic score function into a problem-specific one, using information about the measurements y and measurement model, denoted as $h$ here ( $h$ is JPEG compression $^ +$ masking in this figure, best viewed zoomed in). The additional guidance term is a vector-Jacobian product (VJP) that encourages consistency between the denoising result and the measurements, after a pseudoinverse transformation $h ^ { \dagger }$ . ΠGDM applies the denoising process from ΠG in an iterative fashion to generate valid solutions to the inverse problem.
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
Figure 2: ΠGDM applies a single problem-agnostic diffusion model for various inverse problems, avoiding the cost of training multiple problem-specific ones. Best viewed zoomed in.
|
| 16 |
+
|
| 17 |
+
Diffusion models (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2021c) have been successfully applied to various applications such as text-to-image generation (Rombach et al., 2022; Saharia et al., 2022b), natural language generation (Li et al., 2022), audio synthesis (Kong et al., 2020), and time series modeling (Tashiro et al., 2021). The ability to model complex, high-dimensional distributions also makes diffusion models strong candidates for solving inverse problems, where the goal is to infer the underlying signal from measurements (Bora et al., 2017; Daras et al., 2021; Ongie et al., 2020; Kadkhodaie & Simoncelli, 2021).
|
| 18 |
+
|
| 19 |
+
Most methods that solve inverse problems with diffusion models fall into one of the two paradigms. In the first paradigm, one trains a problem-specific, conditional diffusion model that is limited to specific inverse problems, such as super-resolution (Saharia et al., 2021; Whang et al., 2021; Saharia et al., 2022a). In the second paradigm, one uses problem-agnostic diffusion models that are trained for generative modeling but not train on any specific inverse problem; solutions are obtained via a “plug-and-play” approach that combines the diffusion model and the measurement process, e.g., via Bayes’ rule (Venkatakrishnan et al., 2013a; Bardsley, 2012; Laumont et al., 2022; Choi et al., 2021; Song et al., 2021b; Jalal et al., 2021; Chung et al., 2021; Kawar et al., 2021; 2022a; Chung et al., 2022b; Daras et al., 2022a). These methods can easily adapt to different tasks without re-training the diffusion model but tend to perform worse than problem-specific diffusion models.
|
| 20 |
+
|
| 21 |
+
To achieve the best of both worlds, we introduce pseudoinverse guidance (ΠG), which uses problemagnostic diffusion models to reach the empirical performance of problem-specific ones. Conditioned on the measurements and an explicit measurement model, ΠG estimates the problem-specific score function via Bayes’ rule and uses these scores to draw samples. However, unlike classifier/classifierfree guidance (Dhariwal & Nichol, 2021; Ho & Salimans, 2022), ΠG obtains the problem-specific score directly via the known measurement model, without training additional models. Intuitively, ΠG guides the diffusion process by matching the one-step denoising solution and the ground-truth measurements, after transforming both via a “pseudoinverse” of the measurement model (see Fig. 1). This perspective allows ΠG to be the first guidance-based approach for inverse problem solving that handles measurements with Gaussian noise, as well as some non-linear, non-differentiable measurement models, such as JPEG compression (Kawar et al., 2022b).
|
| 22 |
+
|
| 23 |
+
We evaluate our method, termed Pseudoinverse-Guided Diffusion Models (ΠGDM), on various inverse problems, such as super-resolution, inpainting, and JPEG restoration over ImageNet validation images, and show that it achieves similar performance when compared against state-of-the-art taskspecific diffusion models (Saharia et al., 2021; Dhariwal & Nichol, 2021; Saharia et al., 2022a). To the best of our knowledge, ΠGDM is the first approach based on problem-agnostic models to achieve this quality on ImageNet. We further apply ΠGDM to a wider range of inverse problems, where the measurement process is composed of different types of measurements. This allows us to easily solve a much wider set of problems, including ones have never been solved with diffusion models (see Fig. 2), such as low-resolution $^ +$ JPEG compression $^ +$ masking.
|
| 24 |
+
|
| 25 |
+
# 2 PRELIMINARIES: DIFFUSION MODELS
|
| 26 |
+
|
| 27 |
+
Let us denote the data distribution as $p _ { 0 } ( \mathbf { x } _ { 0 } )$ and define a family of distributions $p _ { t } ( \mathbf { x } _ { t } )$ by injecting i.i.d. Gaussian noise of standard deviation $\sigma _ { t }$ to samples of $p _ { 0 } ( \mathbf { x } )$ , i.e., $p _ { t } ( \mathbf { x } _ { t } | \mathbf { x } _ { 0 } ) = \mathcal { N } ( \mathbf { x } _ { 0 } , \sigma _ { t } ^ { 2 } I )$ . The standard deviation $\sigma _ { t }$ is monotonically increasing with respect to time $t \in [ 0 , T ]$ , with $\sigma _ { 0 } = 0$ and $\sigma _ { T }$ being much larger than the standard deviation of the data1. Samples from $p _ { t } ( \mathbf { x } )$ can be simulated by the following family of stochastic differential equations (SDE), solving from $t = T$ to $t = 0$ (Grenander & Miller, 1994; Karras et al., 2022; Zhang et al., 2022):
|
| 28 |
+
|
| 29 |
+
$$
|
| 30 |
+
\begin{array} { r } { \mathrm { d } \mathbf { x } = - \underbrace { \dot { \sigma } _ { t } \sigma _ { t } \nabla _ { \mathbf { x } } \log p _ { t } ( \mathbf { x } ) \mathrm { d } t } _ { \mathrm { P r o b a b i l i s t i c } \mathrm { O D E } } - \underbrace { \beta _ { t } \sigma _ { t } ^ { 2 } \nabla _ { \mathbf { x } } \log p _ { t } ( \mathbf { x } ) \mathrm { d } t + \sqrt { 2 \beta _ { t } } \sigma _ { t } \mathrm { d } \omega _ { t } } _ { \mathrm { L a n g e v i n p r o c e s s } } , } \end{array}
|
| 31 |
+
$$
|
| 32 |
+
|
| 33 |
+
where $\nabla _ { \mathbf { x } } \log p _ { t } ( \mathbf { x } )$ is the score function, $\omega _ { t }$ is the standard Wiener process, and $\beta _ { t }$ is a function that describes the amount of stochastic noise injected in the process. If $\beta _ { t } = 0$ for all $t$ , then Eq. 1 becomes an ordinary differential equation (ODE) (Anderson, 1982). A common choice of $\beta _ { t }$ is $\eta \dot { \sigma } _ { t } / \sigma _ { t }$ , where $\eta = 1$ corresponds to the variance-exploding SDE (VE-SDE, Song et al. (2021c)) and $\eta \ : = \ : 0$ corresponds to a version of denoising diffusion implicit models (DDIM, Song et al. (2021a)). Various forms of SDEs used by diffusion models in the literature can be described with Eq. 1 with certain $\sigma _ { t }$ and $\beta _ { t }$ functions, up to a time-dependent scaling factor over x.
|
| 34 |
+
|
| 35 |
+
Diffusion models, a.k.a. score-based generative models (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2021c), solve Eq. 1 with two key approximations. The distribution with the highest noise level, $p _ { T } ( \mathbf { x } )$ , is approximated with $\mathcal { N } ( 0 , \sigma _ { T } ^ { 2 } \bar { I } )$ , and the score function is approximated with a neural network $\nabla _ { \mathbf { x } } \log p _ { t } ( \mathbf { x } ) \approx S _ { \theta } ( \mathbf { x } ; \sigma _ { t } )$ , trained with denoising score matching objectives (Vincent, 2011). Then, samples are drawn from diffusion models by solving the ODE or SDE in Eq. 1, such as with Euler’s method, Euler-Maruyama, and higher order ODE solvers (Lu et al., 2022; Karras et al., 2022; Zhang & Chen, 2022).
|
| 36 |
+
|
| 37 |
+
# 3 METHODS
|
| 38 |
+
|
| 39 |
+
Problem statement Suppose we have measurements $\mathbf { y } \in \mathbb { R } ^ { m }$ of some signal $\mathbf { x } _ { 0 } \in \mathbb { R } ^ { n }$ , such that
|
| 40 |
+
|
| 41 |
+
$$
|
| 42 |
+
{ \bf y } = { \cal H } { \bf x } _ { 0 } + { \bf z } ,
|
| 43 |
+
$$
|
| 44 |
+
|
| 45 |
+
where $\pmb { H } \in \mathbb { R } ^ { n \times m }$ is the known measurement matrix (model), and $\mathbf { z } \ \sim \ \mathcal { N } ( 0 , \sigma _ { \mathbf { y } } ^ { 2 } I )$ is an i.i.d. Gaussian noise vector with known dimension-wise standard deviation $\sigma _ { \mathbf { y } }$ . Our goal is to solve the inverse problem and recover $\mathbf { x } _ { 0 } \in \mathbb { R } ^ { n }$ from the measurements $\mathbf { y }$ . In later parts of the paper, we may consider inverse problems whose measurements are not linear, which we denote as $\mathbf { y } = h ( \mathbf { x } _ { 0 } )$ .
|
| 46 |
+
|
| 47 |
+
Diffusion models can solve such inverse problems via Eq. 1, assuming that the problem-specific scores for all noise levels, i.e., $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } | \mathbf { y } )$ , are available. While it is possible to train a conditional diffusion model for a specific $\pmb { H }$ , it is computationally expensive to do this for a large family of problems, such as sparse reconstruction in medical imaging (Chung & Ye, 2022). Therefore, we wish to utilize more commonly available problem-agnostic score models $S _ { \theta } ( \mathbf { x } ; \sigma _ { t } )$ that are not trained specifically for the target inverse problem. If $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } | \mathbf { y } )$ can be effectively approximated with $S _ { \theta } ( \mathbf { x } _ { t } ; \sigma _ { t } )$ , then we can directly plug it in Eq. 1 to solve the inverse problem.
|
| 48 |
+
|
| 49 |
+
# 3.1 APPROXIMATING THE PROBLEM-SPECIFIC SCORE FUNCTION
|
| 50 |
+
|
| 51 |
+
The problem-specific score can be decomposed via Bayes’ rule:
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
\begin{array} { r } { \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } | \mathbf { y } ) = \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } ) + \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } ) , } \end{array}
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
where the first term can be approximated with the score network $S _ { \theta } ( \mathbf { x } _ { t } ; \sigma _ { t } )$ (Vincent, 2011), and the second term is a guidance term which is the score of $p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ .
|
| 58 |
+
|
| 59 |
+
Unfortunately, the score $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ is intractable to compute, and we have to resort to approximations to efficiently estimate it. To see why this is true, we consider the underlying graphical model for $\mathbf { x } _ { 0 } , \mathbf { x } _ { t }$ , and $\mathbf { y }$ , which is $\mathbf { y } \left. \mathbf { x } _ { 0 } \right. \mathbf { x } _ { t } .$ $\mathbf { x } _ { t }$ is produced by adding independent Gaussian
|
| 60 |
+
|
| 61 |
+
noise to $\mathbf { x } _ { \mathrm { 0 } }$ , so it is independent of the measurement y when conditioned on $\mathbf { x } _ { 0 }$ . Therefore, we can write:
|
| 62 |
+
|
| 63 |
+
$$
|
| 64 |
+
p _ { t } ( \mathbf { y } \vert \mathbf { x } _ { t } ) = \int _ { \mathbf { x } _ { 0 } } p ( \mathbf { x } _ { 0 } \vert \mathbf { x } _ { t } ) p ( \mathbf { y } \vert \mathbf { x } _ { 0 } ) \mathrm { d } \mathbf { x } _ { 0 } ,
|
| 65 |
+
$$
|
| 66 |
+
|
| 67 |
+
which involves a marginalization over $\mathbf { x } _ { 0 }$ . The likelihood of $p ( \mathbf { y } \vert \mathbf { x } _ { 0 } )$ is tractable, yet samples from $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ can only be approximated by the diffusion model with high precision (using the variational inference argument by Sohl-Dickstein et al. (2015); Ho et al. (2020)); when $t x$ , sampling from $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ essentially becomes sampling from the entire diffusion model. Even using Monte Carlo methods, it is computationally infeasible to estimate $p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ , let alone its score (where the Monte Carlo estimate will also be biased).
|
| 68 |
+
|
| 69 |
+
Our solution to this issue is to use reasonable approximations to the true $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ , such that the resulting approximation to the score $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ is easy to compute. Intuitively, instead of representing $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ with the entire diffusion model from time $t$ to $0$ , we use a one-step denoising process. Specifically, we first approximate $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ with the following Gaussian:
|
| 70 |
+
|
| 71 |
+
$$
|
| 72 |
+
p _ { t } ( \mathbf x _ { 0 } | \mathbf x _ { t } ) \approx \mathcal N ( \hat { \mathbf x } _ { t } , r _ { t } ^ { 2 } \pmb I ) ,
|
| 73 |
+
$$
|
| 74 |
+
|
| 75 |
+
where the mean is obtained from Tweedie’s formula:
|
| 76 |
+
|
| 77 |
+
$$
|
| 78 |
+
\hat { \mathbf { x } } _ { t } = \mathbb { E } [ \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ] = \mathbf { x } _ { t } + \sigma _ { t } ^ { 2 } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } ) \approx \mathbf { x } _ { t } + \sigma _ { t } ^ { 2 } S _ { \theta } ( \mathbf { x } ; \sigma _ { t } ) .
|
| 79 |
+
$$
|
| 80 |
+
|
| 81 |
+
Eq. 5 represents the minimum mean squared error (MMSE) estimator of $\mathbf { x } _ { \mathrm { 0 } }$ given $\mathbf { x } _ { t }$ and the noise standard deviation $\sigma _ { t }$ (Stein, 1981; Efron, 2011; Saremi & Hyvarinen ¨ , 2019), and $r _ { t }$ is a timedependent standard deviation value that should depend on the data (see discussion in App. A.3). Our choice for the mean (MMSE) can be justified using an argument related to variational inference (App. A.6).
|
| 82 |
+
|
| 83 |
+
Our next step is to approximate the score of $p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ . Since the measurement model obtains $\mathbf { y }$ by performing a linear transform on $\mathbf { x } _ { \mathrm { 0 } }$ and adding independent Gaussian noise (Eq. 2), and $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ is Gaussian under our approximation (Eq. 4), the distribution of y conditioned on $\mathbf { x } _ { t }$ is also Gaussian under our approximation, as follows:
|
| 84 |
+
|
| 85 |
+
$$
|
| 86 |
+
p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } ) \approx \mathcal { N } ( H \hat { \mathbf { x } } _ { t } , r _ { t } ^ { 2 } H H ^ { \top } + \sigma _ { \mathbf { y } } ^ { 2 } I ) .
|
| 87 |
+
$$
|
| 88 |
+
|
| 89 |
+
Thus, we have the following approximation to the score2:
|
| 90 |
+
|
| 91 |
+
$$
|
| 92 |
+
\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } ) \approx \Big ( \underbrace { ( \mathbf { y } - H \hat { \mathbf { x } } _ { t } ) ^ { \top } \left( r _ { t } ^ { 2 } H H ^ { \top } + \sigma _ { \mathbf { y } } ^ { 2 } I \right) ^ { - 1 } H } _ { \mathrm { v e c t o r } } \underbrace { \frac { \partial \hat { \mathbf { x } } _ { t } } { \partial \mathbf { x } _ { t } } } _ { \mathrm { J a c o b i a n } } \Big ) ^ { \top } .
|
| 93 |
+
$$
|
| 94 |
+
|
| 95 |
+
This is a vector-Jacobian product and can be computed with backpropagation.
|
| 96 |
+
|
| 97 |
+
# 3.2 EXTENDING TO NON-LINEAR OPERATORS
|
| 98 |
+
|
| 99 |
+
In many cases, we have that $\sigma _ { \mathbf { y } } = 0$ , and thus, Eq. 7 can be simplified to:
|
| 100 |
+
|
| 101 |
+
$$
|
| 102 |
+
\nabla _ { \mathbf x _ { t } } \log p _ { t } ( \mathbf y | \mathbf x _ { t } ) \approx r _ { t } ^ { - 2 } \big ( ( H ^ { \dagger } \mathbf y - H ^ { \dagger } H \hat { \mathbf x } _ { t } ) ^ { \top } \frac { \partial \hat { \mathbf x } _ { t } } { \partial \mathbf x _ { t } } \big ) ^ { \top } ;
|
| 103 |
+
$$
|
| 104 |
+
|
| 105 |
+
where for a matrix with linearly independent rows, $\pmb { H } ^ { \dag } = \pmb { H } ^ { \top } ( \pmb { H } \pmb { H } ^ { \top } ) ^ { - 1 }$ is the Moore-Penrose pseudoinverse of $\pmb { H }$ . In this paper, we use the term pseudoinverse guidance (ΠG) to denote our guidance method, which uses Eq. 8 for noiseless measurements and Eq. 7 for noisy linear ones.
|
| 106 |
+
|
| 107 |
+
Notably, we only need to perform automatic differentiation explicitly through the score model, but not through the computational graph with $\pmb { H }$ or $H ^ { \dagger }$ (see Listing 1 in App. A.1). This allows us to extend ΠG to measurements that are not necessarily linear or even differentiable. We note that the matrix pseudoinverse satisfies $H H ^ { \dagger } H \mathbf { x } = H \mathbf { x }$ for all $\mathbf { x } \in \mathcal { X }$ . Analogously, for some non-linear measurement function $h : \mathbb { R } ^ { n } \mathbb { R } ^ { m }$ , we may find another function $\mathbf { \bar { \Sigma } } h ^ { \dagger } : \mathbf { \bar { \mathbb { R } } } ^ { m } \mathbb { R } ^ { n }$ such that $h ( h ^ { \dagger } ( h ( \mathbf { x } ) ) ) = h ( \mathbf { x } )$ for all $\mathbf { x } \in \mathbb { R } ^ { n }$ , similar to Kawar et al. (2022b). Two examples are as follows:
|
| 108 |
+
|
| 109 |
+
Quantization Let $h ( \mathbf { x } ) = \lfloor \mathbf { x } \rfloor$ be the element-wise floor function of $\mathbf { x } \in \mathbb { R } ^ { n }$ . Then we can define ${ \bar { h ^ { \dag } } } ( \mathbf { x } ) : = \mathbf { x }$ for all $\mathbf { x } \in \mathbb { Z }$ , and $\bar { h ( h ^ { \dag } ( h ( \mathbf x ) ) ) } )$ is still the floor function.
|
| 110 |
+
|
| 111 |
+
JPEG encoding Let $h ( \mathbf { x } )$ be the JPEG encoding function, where quantization occurs after a discrete cosine transform operation. The corresponding JPEG decoding algorithm does not modify the values produced after quantization, so we can simply define $h ^ { \dagger } ( \mathbf { x } )$ as the JPEG decoding algorithm.
|
| 112 |
+
|
| 113 |
+
This idea can also be applied to other measurement models, such as the formation of a low dynamic range image (details in App. A.5). The corresponding ΠG term would then become:
|
| 114 |
+
|
| 115 |
+
$$
|
| 116 |
+
\nabla _ { \mathbf { x } _ { t } } \log { p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } ) } \approx r _ { t } ^ { - 2 } \big ( ( h ^ { \dagger } ( \mathbf { y } ) - h ^ { \dagger } ( h ( \hat { \mathbf { x } } _ { t } ) ) ) ^ { \top } \frac { \partial \hat { \mathbf { x } } _ { t } } { \partial \mathbf { x } _ { t } } \big ) ^ { \top } ,
|
| 117 |
+
$$
|
| 118 |
+
|
| 119 |
+
which generalizes the linear case (Eq. 8) when $h ( \mathbf { x } ) = H \mathbf { x }$ and $h ^ { \dagger } ( \mathbf { x } ) = { \cal H } ^ { \dagger } \mathbf { x }$ for all $\mathbf { x } \in \mathbb { R } ^ { n }$
|
| 120 |
+
|
| 121 |
+
# 3.3 ADAPTIVE WEIGHTS IN GUIDED DIFFUSION MODELS
|
| 122 |
+
|
| 123 |
+
Similar to the guidance scalar in the classifier(-free) guidance literature (Dhariwal & Nichol, 2021; Ho & Salimans, 2022), we introduce a scalar weight in front of the guidance term $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ . However, unlike most existing methods that apply a fixed weight for different diffusion times, we introduce a heuristic that implicitly adapts the guidance weights according to the timestep. We use $f ( \mathbf { x } _ { t } ; s , t , \eta )$ to denote the one step update using the problem-agnostic score model from time $t$ to times $s$ (assuming $s < t ,$ ), using the sampler introduced in the DDIM paper (Song et al., 2021a), with $\eta \in [ 0 , 1 ]$ being a hyperparameter (details in App. A.4). Our one-step sampling update from time $t$ to time $s$ with pseudoinverse guidance is:
|
| 124 |
+
|
| 125 |
+
$$
|
| 126 |
+
\begin{array} { r } { \mathbf { x } _ { s } = f ( \mathbf { x } _ { t } ; s , t , \eta ) + r _ { t } ^ { 2 } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } ) . } \end{array}
|
| 127 |
+
$$
|
| 128 |
+
|
| 129 |
+
If we use the noiseless case in Eq. 9, this becomes:
|
| 130 |
+
|
| 131 |
+
$$
|
| 132 |
+
\mathbf { x } _ { s } = f ( \mathbf { x } _ { t } ; s , t , \eta ) + \left( \left( h ^ { \dagger } ( \mathbf { y } ) - h ^ { \dagger } ( h ( \hat { \mathbf { x } } _ { t } ) ) \right) ^ { \top } \frac { \partial \hat { \mathbf { x } } _ { t } } { \partial \mathbf { x } _ { t } } \right) ^ { \top } .
|
| 133 |
+
$$
|
| 134 |
+
|
| 135 |
+
We describe the algorithm in Algorithm 1 (App. A.1). As the coefficients for the problem-agnostic score $\nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } )$ depend on the step $t s$ , this is equivalent to using the original DDIM sampler but adapting the weights to the pseudoinverse guidance term at different timesteps. To illustrate this, we can compare the ratio between the weights of our approach and the ones with $w _ { r } = 1$ in Ho et al. (2022) (see Fig. 6). Intuitively, our approach increases the weights during the initial sampling phase and then decreases it to one towards the end. We also compare our weights with the ones used in Ho et al. (2022) on image restoration problems, both of which use the pseudoinverse guidance with 100 diffusion steps and $\eta = 0 . 2$ . On the super-resolution case (Fig. 7), our weights consistently produce sharp images. We further illustrate the advantages of our weights on JPEG restoration in App. A.4, where large, fixed weights that worked better in super-resolution could be unstable in another task.
|
| 136 |
+
|
| 137 |
+
# 3.4 DIFFERENCES FROM EXISTING GUIDANCE METHODS
|
| 138 |
+
|
| 139 |
+
Table 1: Comparison of different guidance methods.
|
| 140 |
+
|
| 141 |
+
<table><tr><td>Guidance</td><td>Expression</td><td>Xt →y differentiable</td><td>Train on (Xt,y)</td><td>Noisy y</td></tr><tr><td>Classifier</td><td>Vxt logq(y|xt）</td><td>Required</td><td>Yes</td><td>-</td></tr><tr><td>Reconstruction</td><td>Vxlly-Hxtll²</td><td>Required</td><td>No</td><td>No</td></tr><tr><td>Pseudoinverse</td><td>Eqs. 7 to 9</td><td>Not required</td><td>No</td><td>Yes</td></tr></table>
|
| 142 |
+
|
| 143 |
+
Our approach is notably different from prior guidance-based methods in terms of how the conditional score $\nabla _ { \mathbf { x } _ { t } } \log p ( \mathbf { y } | \mathbf { x } _ { t } )$ is approximated (see Tab. 1). Compared with classifier / classifier-free guidance (Dhariwal & Nichol, 2021; Ho & Salimans, 2022), we do not require training on pairs of $\left( \mathbf { x } _ { t } , \mathbf { y } \right)$ (noisy data and measurements). Compared with reconstruction guidance (Ho et al., 2022; Chung et al., 2022b; Ryu & Ye, 2022), ΠG has three advantages:
|
| 144 |
+
|
| 145 |
+
• Our approximation of $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ is consistent, i.e., it does not depend on the measurement model $\pmb { H }$ . The same cannot be said for reconstruction guidance, which makes isotropic Gaussian assumptions on $\mathbf { y }$ (see App. A.2).
|
| 146 |
+
• In reconstruction guidance, the pseudoinverse $H ^ { \dagger }$ is replaced with matrix transpose $\pmb { H } ^ { \top }$ (see App. A.2), which is different for linear $\pmb { H }$ whose singular values are not all 0 or 1.
|
| 147 |
+
• ΠG can be applied to noisy, non-linear, or non-differentiable measurement models, as discussed in Sec. 3.2. In cases like JPEG, it is easier to define a generalized notion of pseudoinverse than a generalized notion of transpose (or adjoint).
|
| 148 |
+
|
| 149 |
+
# 4 RELATED WORK
|
| 150 |
+
|
| 151 |
+
Deep neural networks have been extensively used as priors for solving inverse problems (Venkatakrishnan et al., 2013b). Here, we focus on the setting where we can train models based on clean data but not on the problem, which is only known at inference time. This is reasonable in many real-world applications, such as medical imaging (Jalal et al., 2021; Chung & Ye, 2022) and JPEG restoration (Ehrlich et al., 2020). These inverse problem solvers may use different types of neural networks, such as randomly initialized networks (Ulyanov et al., 2018), denoisers (Romano et al., 2016), robust classifiers (Santurkar et al., 2019), and generative models (Bora et al., 2017). Methods based on generative adversarial networks (GANs, Goodfellow et al. (2014)) search for the latent variables and/or the generator parameters that would produce images aligning with the measurements (Bora et al., 2017; Pan et al., 2021; Menon et al., 2020); these methods often require hundreds if not thousands of iterations, despite recent methods with improved efficiency (Daras et al., 2022a).
|
| 152 |
+
|
| 153 |
+
As another family of generative models, diffusion models are also used as inverse problem solvers, with two notable advantages over GANs: (i) it is trained with regression objectives over noisy data, so it can naturally deal with measurement noise without having to perform inversion like in GANs; $( i i )$ its close connections to SDE/ODE solvers allow the use of more efficient iterative updates. In particular, Denoising Diffusion Restoration Models (DDRM, Kawar et al. (2022a)) leverage both to derive efficient inverse problem solvers for both noisy and noiseless measurements.
|
| 154 |
+
|
| 155 |
+
Similar to DDRM, many works adopt a “replacement” approach, where consistency with the measurements are enforced by replacing parts of its intermediate predictions from the one-step denoiser with the measurements, sometimes in a transformed space (Song et al., 2021c; Choi et al., 2021; Song et al., 2021b; Chung et al., 2021; Kawar et al., 2021). Despite being successful in many tasks, they have trouble dealing with sparse measurements, where the replacements have weaker impact on the sampling process. By computing additional gradients through the diffusion model, ΠG allows the measurements to impact all the predicted values during the updates, regardless of sparsity. This is similar to reconstruction guidance, which also differentiates through the diffusion model during its updates (Ho et al., 2022; Ryu & Ye, 2022; Chung et al., 2022b). In fact, ΠG is identical in the noiseless, linear case if the transpose of the measurement matrix is equal to its pseudoinverse (App. A.2). Nevertheless, ΠG introduces principled ways of dealing with noisy, non-linear, or even non-differentiable measurements, as discussed in Sec. 3.4.
|
| 156 |
+
|
| 157 |
+
# 5 EXPERIMENTS
|
| 158 |
+
|
| 159 |
+
Our approach, named Pseudoinverse-guided Diffusion Models (ΠGDM), combines ΠG (Eqs. 7 to 9) and the adaptive weight schedule (Eq. 10). While we use a sampler based on DDIM here, we note that other samplers can be used as well. We evaluate quantitative results on the ImageNet dataset (Russakovsky et al., 2015) with publicly available diffusion models trained on images of size $2 5 6 \times 2 5 6$ , as there are extensive prior results with problem-specific diffusion models trained on ImageNet (Dhariwal & Nichol, 2021; Saharia et al., 2021; 2022a).
|
| 160 |
+
|
| 161 |
+
• First, we compare ΠGDM against problem-specific models on $4 \times$ super-resolution, inpainting, and JPEG restoration. Despite the “unfair” advantage held by problem-specific models, ΠGDM is on par with them in terms of performance. • Next, we perform an ablation study over the two components introduced in this paper. • Finally, we apply ΠGDM to inverse problems where the measurement process is composed of several steps, such as JPEG $^ +$ super-resolution $^ +$ inpainting, denoising $^ +$ inpainting, etc..
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
Figure 3: Results on JPEG restoration. From left to right: the compressed JPEG image, restoration results from Palette (task-specific) and ΠGDM (task-agnostic), and the reference image.
|
| 165 |
+
|
| 166 |
+
The compositional nature of these problems makes it infeasible to train diffusion models for each problem and highlights the strength of ΠGDM.
|
| 167 |
+
|
| 168 |
+
# 5.1 QUANTITATIVE RESULTS
|
| 169 |
+
|
| 170 |
+
We consider two popular metrics, Frechet Inception Distance (FID, (Heusel et al., 2017)) and Classifier Accuracy (CA) of a pre-trained ResNet50 model (He et al., 2015). Unless specified otherwise, we use the noiseless version for pseudoinverse guidance (Eq. 8). We report super-resolution results on the full ImageNet validation set, and to follow the earlier practice established in Saharia et al. (2022a), we report inpainting and JPEG restoration results on a subset that contains 10k images3. The ImageNet models that we use are trained with 1000 discrete timesteps, corresponding to 1000 discrete noise levels (Dhariwal & Nichol, 2021). For simplicity, we always use uniform spacing when we iterate the timesteps. Performance may further improve with better timestep scheduling, such as the one that iterates more frequently at lower noise levels (Karras et al., 2022). We use 100 iterations and $\eta = 1 . 0$ for ΠGDM, and include additional task-specific details in App. B.
|
| 171 |
+
|
| 172 |
+
# 5.1.1 SUPER-RESOLUTION
|
| 173 |
+
|
| 174 |
+
We apply average pooling $( P o o l )$ and bicubic interpolation (Bicubic, which applies a convolution to the image) to produce two sets of $6 4 \times 6 4$ images, and then apply our $4 \times$ super-resolution algorithms to each. We consider both class-conditional (denoted as $c c$ in Tab. 2) and class-unconditional models as the base generative model. In Tab. 2, we report results from ΠGDM and three other baselines: DDRM (Kawar et al. (2022a)), SR3 (Saharia et al. (2021)), and ADM-U (Dhariwal & Nichol (2021)). DDRM uses task-agnostic models, whereas SR3 and ADM-U use diffusion models specifically for the $6 4 2 5 6$ super-resolution problem. On Pool, ΠGDM significantly outperforms DDRM, while only being slightly worse than ADM-U; on Bicubic, ΠGDM outperforms all three baselines. Perhaps surprisingly, the ADM-U model performs much worse in Bicubic than Pool because it was trained on low-resolution images generated by average pooling4 (i.e., the Pool problem); when Bicubic images are used, the generated results become more blurry. This suggests that problem-specific diffusion models may fail to generalize beyond settings that they are trained on.
|
| 175 |
+
|
| 176 |
+
# 5.1.2 INPAINTING
|
| 177 |
+
|
| 178 |
+
We use the two types of inpainting masks used in (Saharia et al., 2022a): the center $1 2 8 \times 1 2 8$ pixels (Center), and freeform masks simulating brushstrokes that contain roughly $2 0 \% - 3 0 \%$ of the pixels in each image (Freeform). In addition, we report ΠGDM results over the noisy inpainting problem with i.i.d. Gaussian noise of $\sigma _ { \mathbf { y } } = 0 . 0 5$ (the pixel intensity range is $[ 0 , 1 ] )$ ; the problem becomes harder as the model has to perform denoising and inpainting at the same time. For the noisy setting, we use Eq. 7 for ΠG. In Tab. 3, we report quantitative results on the two inpainting tasks, mainly comparing with Palette (Saharia et al., 2022a), which trains a diffusion model specifically on the inpainting task. While ΠGDM achieves a slightly worse FID compared with Palette, it has a higher classifier accuracy in both cases. Moreover, ΠGDM suffers merely a small performance drop when applied to the more challenging denoising $^ +$ inpainting task, demonstrating its robustness to noisy measurements. Methods based on reconstruction guidance, however, fail to perform denoising effectively, as there are no mechanisms to address measurement noise (see Fig. 4).
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
Figure 4: Results on noisy inpainting problems. Reconstruction guidance (third column) does not handle measurement noise and will keep the noisy measurements, so only the masked regions are denoised.
|
| 182 |
+
|
| 183 |
+
Table 2: $4 \times$ super-resolution results. Dark-colored rows indicate methods using problem-specific models.
|
| 184 |
+
|
| 185 |
+
<table><tr><td>Filter</td><td>Method</td><td>FID↓</td><td>CA↑</td></tr><tr><td rowspan="5">Pool</td><td>ADM(cc,Dhariwal & Nichol (2021))</td><td>3.1</td><td>73.4%</td></tr><tr><td>DDRM (Kawar et al.,2022a)</td><td>14.8</td><td>64.6%</td></tr><tr><td>IIGDM(Ours)</td><td>3.8</td><td>72.3%</td></tr><tr><td>DDRM(cc, Kawar et al. (2022a))</td><td>14.1</td><td>65.2%</td></tr><tr><td>IIGDM(cc,Ours)</td><td>3.6</td><td>72.2%</td></tr><tr><td rowspan="6">Bicubic</td><td>SR3 (Saharia et al., 2021)</td><td>5.2</td><td>68.3%</td></tr><tr><td>ADM (cc, Dhariwal & Nichol (2021))</td><td>14.8</td><td>66.7%</td></tr><tr><td>DDRM (Kawar et al., 2022a)</td><td>21.3</td><td>63.2%</td></tr><tr><td>IIGDM(Ours)</td><td>3.6</td><td>72.1%</td></tr><tr><td>DDRM(cc, Kawar et al. (2022a))</td><td>19.6</td><td>65.3%</td></tr><tr><td>IIGDM (cc, Ours)</td><td>3.2</td><td>75.1%</td></tr></table>
|
| 186 |
+
|
| 187 |
+
Table 3: Inpainting results. Dark-colored rows indicate methods using problem-specific models.
|
| 188 |
+
|
| 189 |
+
<table><tr><td>Mask</td><td>Method</td><td>FID-10k↓</td><td>CA↑</td></tr><tr><td rowspan="5">Center</td><td>DeepFillv2 (Yu et al.,2019)</td><td>18.0</td><td>64.3%</td></tr><tr><td>Palette (Saharia et al., 2022a)</td><td>6.6</td><td>69.3%</td></tr><tr><td>DDRM (Kawar et al., 2022a)</td><td>24.4</td><td>62.1%</td></tr><tr><td>IIGDM(Ours)</td><td>7.3</td><td>72.6%</td></tr><tr><td>IIGDM (noisy, Ours)</td><td>9.5</td><td>72.2%</td></tr><tr><td rowspan="5">Freeform</td><td>DeepFillv2 (Yu et al.,2019)</td><td>9.4</td><td>68.8%</td></tr><tr><td>Palette (Saharia et al., 2022a)</td><td>5.2</td><td>72.3%</td></tr><tr><td>DDRM (Kawar et al., 2022a)</td><td>8.6</td><td>71.9%</td></tr><tr><td>IIGDM(Ours)</td><td>5.3</td><td>75.3%</td></tr><tr><td>IIGDM (noisy, Ours)</td><td>73</td><td>74.5%</td></tr></table>
|
| 190 |
+
|
| 191 |
+
# 5.1.3 JPEG RESTORATION
|
| 192 |
+
|
| 193 |
+
We consider the three JPEG quality factors (QFs) used in Saharia et al. (2022a), which are 5, 10, and 20. In Tab. 4, we report quantitative results on JPEG, where we compare against a regressionbased baseline and Palette, both of which are trained specifically for JPEG images with QFs ranging from 5 to 30. Compared with Palette, ΠGDM achieves a slightly worse FID (less than 0.6) but higher classifier accuracy on QFs 10 and 20. We note that the model used in ΠGDM has never seen any JPEG images compressed to these quality factors, demonstrating the strength of task-agnostic diffusion models.
|
| 194 |
+
|
| 195 |
+
# 5.2 ABLATION STUDIES
|
| 196 |
+
|
| 197 |
+
ΠGDM introduces two key components: pseudoinverse guidance (ΠG) for problem-specific score estimation and the adaptive guidance weight schedule for sampling (AW). To illustrate their effectiveness, we compare with alternative approaches. The alternative to ΠG is the reconstruction guidance, whereas the alternative to AW is the standard weight schedule set with $w _ { r } \in \{ 1 , 2 , 5 \}$ in Ho et al. (2022) (the $w _ { r }$ with best performance is reported).
|
| 198 |
+
|
| 199 |
+
Table 4: JPEG restoration results. Dark-colored rows indicate methods using problem-specific models.
|
| 200 |
+
|
| 201 |
+
<table><tr><td>QF</td><td>Method</td><td>FID-10k↓</td><td>CA个</td></tr><tr><td rowspan="3">5</td><td>Regression (Saharia et al., 2022a)</td><td>29.0</td><td>52.8%</td></tr><tr><td>Palette (Saharia et al., 2022a)</td><td>8.3</td><td>64.2%</td></tr><tr><td>IIGDM (Ours)</td><td>8.6</td><td>64.1%</td></tr><tr><td rowspan="3">10</td><td>Regression (Saharia et al.,2022a)</td><td>18.0</td><td>63.5%</td></tr><tr><td>Palette (Saharia et al., 2022a)</td><td>5.4</td><td>70.7%</td></tr><tr><td>IIGDM (Ours)</td><td>6.0</td><td>71.0%</td></tr><tr><td rowspan="3">20</td><td>Regression (Saharia et al., 2022a)</td><td>11.5</td><td>69.7%</td></tr><tr><td>Palette (Saharia et al., 2022a)</td><td>4.3</td><td>73.5%</td></tr><tr><td>IIGDM (Ours)</td><td>4.7</td><td>74.4%</td></tr></table>
|
| 202 |
+
|
| 203 |
+
Table 5: Ablation studies on pseudoinverse guidance (ΠG) and our adaptive weight schedule (AW).
|
| 204 |
+
|
| 205 |
+
<table><tr><td rowspan="2">IIG</td><td rowspan="2">AW</td><td colspan="2">Deblur</td><td colspan="2">Bicubic</td></tr><tr><td>PSNR↑</td><td>KID-1k ×10³↓</td><td>FID-10k↓</td><td>CA↑</td></tr><tr><td>×</td><td>X</td><td>21.98</td><td>41.03</td><td>18.6</td><td>40.6%</td></tr><tr><td>×</td><td>√</td><td>20.97</td><td>42.56</td><td>18.9</td><td>60.4%</td></tr><tr><td>√</td><td>×</td><td>31.95</td><td>0.98</td><td>15.4</td><td>67.6%</td></tr><tr><td>?</td><td></td><td>39.36</td><td>0.00</td><td>6.2</td><td>72.4%</td></tr></table>
|
| 206 |
+
|
| 207 |
+
We consider the uniform kernel deblurring (Deblur) and bicubic $4 \times$ super-resolution (Bicubic) tasks discussed in (Kawar et al., 2022a). The measurement matrix for both tasks have varying singular values (see Fig. 5 in App. A.2), so ΠG and reconstruction guidance updates are quite different since $\smash { H ^ { \top } \neq H ^ { \dagger } }$ . From Tab. 5, we can see that methods that use ΠG achieves a significant improvement over reconstruction guidance, and switching to the pseudoinverse in the guidance term is critical. ΠG itself achieves superior performance with AW, illustrating the importance of having a good sampling algorithm along with the guidance term. We provide additional experimental details and further ablation studies on the number of iterations per sample and $\eta$ in App. B.
|
| 208 |
+
|
| 209 |
+
# 5.3 INVERSE PROBLEMS WITH COMPOSED MEASUREMENTS
|
| 210 |
+
|
| 211 |
+
Finally, we discuss cases where the measurement process consists of several simpler measurements, leading to some applications such as JPEG restoration with super-resolution $^ +$ inpainting, etc., where the compositional nature of the measurements makes it too expensive to train problem-specific diffusion models individually. Specifically, let $h ( \mathbf { x } ) = h _ { 1 } \circ h _ { 2 } \ldots \circ h _ { k } ( \mathbf { x } )$ be a measurement model composed of $k$ smaller measurements. For certain measurements, such as low-resolution filtering, JPEG, and masking, we can approximate $h ^ { \dagger } ( \mathbf { x } )$ with $h _ { k } ^ { \dagger } \circ . . . \circ h _ { 2 } ^ { \dagger } \circ h _ { 1 } ^ { \dagger } ( { \bf x } )$ , and then use ΠGDM with Eq. 9 directly. We illustrate some examples in Fig. 2 and Fig. 13 (App. B.3). To the best of our knowledge, many of these problems have not been solved with problem-agnostic diffusion models before (such as super-resolution $+ \mathrm { J P E G } +$ inpainting).
|
| 212 |
+
|
| 213 |
+
# 6 DISCUSSIONS, LIMITATIONS, AND FUTURE WORK
|
| 214 |
+
|
| 215 |
+
In this paper, we introduced ΠGDM, an inverse problem solver using unconditional diffusion models. On various tasks, ΠGDM achieves competitive quality with conditional models while avoiding expensive problem-specific training. As a result, we can use problem-agnostic diffusion models to solve certain problems that would be cost-ineffective to address individually with conditional diffusion models, leading to a much wider set of applications. The ability to handle measurement noise also gives ΠGDM the potential to address certain real-world applications, such as MRI imaging with Gaussian noise (Sijbers & Den Dekker, 2004).
|
| 216 |
+
|
| 217 |
+
Despite having better restoration results than DDRM (Kawar et al., 2022a), ΠGDM is slower, as each iteration costs more memory and compute due to the vector-Jacobian product over the score model. Therefore, it would be helpful to explore more efficient sampling techniques. It is also interesting to investigate if similar ideas as ΠG can be used for diffusion models that do not directly operate on the data space (Vahdat et al., 2021; Rombach et al., 2022; Sinha et al., 2021), or are based on alternative forward diffusion models (Jing et al., 2022; Rissanen et al., 2022; Daras et al., 2022b; Hoogeboom & Salimans, 2022).
|
| 218 |
+
|
| 219 |
+
Reproducibility statement We have made the following efforts to facilitate reproducibility of our work. (i) Our experiments are conducted on publicly available datasets and model checkpoints5 (Sec. 5). (ii) We include a detailed description of our algorithm in Algorithm 1. (iii) We discuss all the key hyperparameters and evaluation metrics to reproduce our experiments in Sec. 3.3 and App. B. (iv) We provide more explanations to some statements in the main paper in App. A.2 to A.5.
|
| 220 |
+
|
| 221 |
+
# REFERENCES
|
| 222 |
+
|
| 223 |
+
Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3):313–326, 1982.
|
| 224 |
+
|
| 225 |
+
Johnathan M Bardsley. Mcmc-based image reconstruction with uncertainty quantification. SIAM Journal on Scientific Computing, 34(3):A1316–A1332, 2012.
|
| 226 |
+
|
| 227 |
+
Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd ´ gans. arXiv preprint arXiv:1801.01401, 2018.
|
| 228 |
+
|
| 229 |
+
Ashish Bora, Ajil Jalal, Eric Price, and Alexandros G Dimakis. Compressed sensing using generative models. arXiv preprint arXiv:1703.03208, March 2017.
|
| 230 |
+
|
| 231 |
+
Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, August 2021.
|
| 232 |
+
|
| 233 |
+
Hyungjin Chung and Jong Chul Ye. Score-based diffusion models for accelerated mri. Medical Image Analysis, pp. 102479, 2022.
|
| 234 |
+
|
| 235 |
+
Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-Closer-Diffuse-Faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction. arXiv preprint arXiv:2112.05146, December 2021.
|
| 236 |
+
|
| 237 |
+
Hyungjin Chung, Jeongsol Kim, Michael T Mccann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022a.
|
| 238 |
+
|
| 239 |
+
Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. arXiv preprint arXiv:2206.00941, 2022b.
|
| 240 |
+
|
| 241 |
+
Giannis Daras, Joseph Dean, Ajil Jalal, and Alexandros G Dimakis. Intermediate layer optimization for inverse problems using deep generative models. arXiv preprint arXiv:2102.07364, February 2021.
|
| 242 |
+
|
| 243 |
+
Giannis Daras, Yuval Dagan, Alexandros G Dimakis, and Constantinos Daskalakis. Score-guided intermediate layer optimization: Fast langevin mixing for inverse problem. arXiv preprint arXiv:2206.09104, 2022a.
|
| 244 |
+
|
| 245 |
+
Giannis Daras, Mauricio Delbracio, Hossein Talebi, Alexandros G Dimakis, and Peyman Milanfar. Soft diffusion: Score matching for general corruptions. arXiv preprint arXiv:2209.05442, 2022b.
|
| 246 |
+
|
| 247 |
+
Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. arXiv preprint arXiv:2105.05233, May 2021.
|
| 248 |
+
|
| 249 |
+
Bradley Efron. Tweedie’s formula and selection bias. Journal of the American Statistical Association, 106(496):1602–1614, 2011.
|
| 250 |
+
|
| 251 |
+
Max Ehrlich, Larry Davis, Ser-Nam Lim, and Abhinav Shrivastava. Quantization guided JPEG artifact correction. arXiv preprint arXiv:2004.09320, April 2020.
|
| 252 |
+
|
| 253 |
+
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pp. 2672–2680, 2014.
|
| 254 |
+
|
| 255 |
+
Ulf Grenander and Michael I Miller. Representations of knowledge in complex systems. Journal of the Royal Statistical Society: Series B (Methodological), 56(4):549–581, 1994.
|
| 256 |
+
|
| 257 |
+
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, December 2015.
|
| 258 |
+
|
| 259 |
+
Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two Time-Scale update rule converge to a local nash equilibrium. arXiv preprint arXiv:1706.08500, June 2017.
|
| 260 |
+
|
| 261 |
+
Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.
|
| 262 |
+
|
| 263 |
+
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. arXiv preprint arXiv:2006.11239, June 2020.
|
| 264 |
+
|
| 265 |
+
Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022.
|
| 266 |
+
|
| 267 |
+
Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. arXiv preprint arXiv:2209.05557, 2022.
|
| 268 |
+
|
| 269 |
+
Ajil Jalal, Marius Arvinte, Giannis Daras, Eric Price, Alexandros G Dimakis, and Jonathan I Tamir. Robust compressed sensing MRI with deep generative priors. arXiv preprint arXiv:2108.01368, August 2021.
|
| 270 |
+
|
| 271 |
+
Bowen Jing, Gabriele Corso, Renato Berlinghieri, and Tommi Jaakkola. Subspace diffusion generative models. arXiv preprint arXiv:2205.01490, 2022.
|
| 272 |
+
|
| 273 |
+
Zahra Kadkhodaie and Eero Simoncelli. Stochastic solutions for linear inverse problems using the prior implicit in a denoiser. Advances in Neural Information Processing Systems, 34:13242– 13254, 2021.
|
| 274 |
+
|
| 275 |
+
Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusionbased generative models. arXiv preprint arXiv:2206.00364, 2022.
|
| 276 |
+
|
| 277 |
+
Bahjat Kawar, Gregory Vaksman, and Michael Elad. SNIPS: Solving noisy inverse problems stochastically. arXiv preprint arXiv:2105.14951, May 2021.
|
| 278 |
+
|
| 279 |
+
Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. arXiv preprint arXiv:2201.11793, 2022a.
|
| 280 |
+
|
| 281 |
+
Bahjat Kawar, Jiaming Song, Stefano Ermon, and Michael Elad. Jpeg artifact correction using denoising diffusion restoration models. arXiv preprint arXiv:2209.11888, 2022b.
|
| 282 |
+
|
| 283 |
+
Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020.
|
| 284 |
+
|
| 285 |
+
Remi Laumont, Valentin De Bortoli, Andr ´ es Almansa, Julie Delon, Alain Durmus, and Marcelo ´ Pereyra. Bayesian imaging using plug & play priors: when langevin meets tweedie. SIAM Journal on Imaging Sciences, 15(2):701–737, 2022.
|
| 286 |
+
|
| 287 |
+
Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. arXiv preprint arXiv:2205.14217, 2022.
|
| 288 |
+
|
| 289 |
+
Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image hdr reconstruction by learning to reverse the camera pipeline. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1651–1660, 2020.
|
| 290 |
+
|
| 291 |
+
Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv preprint arXiv:2206.00927, 2022.
|
| 292 |
+
|
| 293 |
+
Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021.
|
| 294 |
+
|
| 295 |
+
Sachit Menon, Alexandru Damian, Shijia Hu, Nikhil Ravi, and Cynthia Rudin. PULSE: SelfSupervised photo upsampling via latent space exploration of generative models. arXiv preprint arXiv:2003.03808, March 2020.
|
| 296 |
+
|
| 297 |
+
Gregory Ongie, Ajil Jalal, Christopher A Metzler, Richard G Baraniuk, Alexandros G Dimakis, and Rebecca Willett. Deep learning techniques for inverse problems in imaging. IEEE Journal on Selected Areas in Information Theory, 1(1):39–56, 2020.
|
| 298 |
+
|
| 299 |
+
Xingang Pan, Xiaohang Zhan, Bo Dai, Dahua Lin, Chen Change Loy, and Ping Luo. Exploiting deep generative prior for versatile image restoration and manipulation. IEEE transactions on pattern analysis and machine intelligence, PP, September 2021. ISSN 0162-8828. doi: 10.1109/TPAMI. 2021.3115428.
|
| 300 |
+
|
| 301 |
+
Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022.
|
| 302 |
+
|
| 303 |
+
Yaniv Romano, Michael Elad, and Peyman Milanfar. The little engine that could: Regularization by denoising (RED). arXiv preprint arXiv:1611.02862, November 2016.
|
| 304 |
+
|
| 305 |
+
Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High- ¨ resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10684–10695, 2022.
|
| 306 |
+
|
| 307 |
+
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. International journal of computer vision, 115 (3):211–252, December 2015. ISSN 0920-5691, 1573-1405. doi: 10.1007/s11263-015-0816-y.
|
| 308 |
+
|
| 309 |
+
Dohoon Ryu and Jong Chul Ye. Pyramidal denoising diffusion probabilistic models. arXiv preprint arXiv:2208.01864, 2022.
|
| 310 |
+
|
| 311 |
+
Chitwan Saharia, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. arXiv preprint arXiv:2104.07636, 2021.
|
| 312 |
+
|
| 313 |
+
Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pp. 1–10, 2022a.
|
| 314 |
+
|
| 315 |
+
Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
|
| 316 |
+
|
| 317 |
+
Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. arXiv preprint arXiv:1906.09453, June 2019.
|
| 318 |
+
|
| 319 |
+
Saeed Saremi and Aapo Hyvarinen. Neural empirical bayes. ¨ Journal of machine learning research: JMLR, 20(181):1–23, 2019. ISSN 1532-4435, 1533-7928.
|
| 320 |
+
|
| 321 |
+
Jan Sijbers and AJ Den Dekker. Maximum likelihood estimation of signal amplitude and noise variance from mr data. Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, 51(3):586–594, 2004.
|
| 322 |
+
|
| 323 |
+
Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems, 34: 12533–12548, 2021.
|
| 324 |
+
|
| 325 |
+
Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, March 2015.
|
| 326 |
+
|
| 327 |
+
Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021a.
|
| 328 |
+
|
| 329 |
+
Yang Song, Liyue Shen, Lei Xing, and Stefano Ermon. Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005, 2021b.
|
| 330 |
+
|
| 331 |
+
Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021c.
|
| 332 |
+
|
| 333 |
+
Charles M Stein. Estimation of the mean of a multivariate normal distribution. The annals of Statistics, pp. 1135–1151, 1981.
|
| 334 |
+
|
| 335 |
+
Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems, 34:24804–24816, 2021.
|
| 336 |
+
|
| 337 |
+
Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Deep image prior. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 9446–9454, 2018.
|
| 338 |
+
|
| 339 |
+
Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. arXiv preprint arXiv:2106.05931, June 2021.
|
| 340 |
+
|
| 341 |
+
Singanallur V Venkatakrishnan, Charles A Bouman, and Brendt Wohlberg. Plug-and-play priors for model based reconstruction. In 2013 IEEE Global Conference on Signal and Information Processing, pp. 945–948. IEEE, 2013a.
|
| 342 |
+
|
| 343 |
+
Singanallur V Venkatakrishnan, Charles A Bouman, and Brendt Wohlberg. Plug-and-play priors for model based reconstruction. In 2013 IEEE Global Conference on Signal and Information Processing, pp. 945–948. IEEE, 2013b.
|
| 344 |
+
|
| 345 |
+
Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661–1674, July 2011.
|
| 346 |
+
|
| 347 |
+
Jay Whang, Mauricio Delbracio, Hossein Talebi, Chitwan Saharia, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. arXiv preprint arXiv:2112.02475, December 2021.
|
| 348 |
+
|
| 349 |
+
Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. LSUN: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, June 2015.
|
| 350 |
+
|
| 351 |
+
Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pp. 4471–4480, 2019.
|
| 352 |
+
|
| 353 |
+
Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, April 2022.
|
| 354 |
+
|
| 355 |
+
Qinsheng Zhang, Molei Tao, and Yongxin Chen. gddim: Generalized denoising diffusion implicit models. arXiv preprint arXiv:2206.05564, June 2022.
|
| 356 |
+
|
| 357 |
+
# A ADDITIONAL METHOD DETAILS
|
| 358 |
+
|
| 359 |
+
# A.1 ALGORITHM DETAILS
|
| 360 |
+
|
| 361 |
+
We illustrate a PyTorch-like implementation for computing the pseudoinverse for the noiseless case in Listing 1. For a different inverse problem, we only need to change the definitions for functions H and H pinv. In practice, many diffusion model architectures adopt the Variance-Preserving (VP) SDE, which scales the signal $\mathbf { x } _ { \mathrm { 0 } }$ down as the noise level increases, specifically:
|
| 362 |
+
|
| 363 |
+
$$
|
| 364 |
+
\mathbf { x } _ { t } = \mathbf { x } _ { 0 } + \sigma _ { t } \epsilon \quad ( \mathbf { V } \mathbf { E } ) \quad \Longleftrightarrow \quad \tilde { \mathbf { x } } _ { t } = \sqrt { \alpha _ { t } } \mathbf { x } _ { 0 } + \sqrt { 1 - \alpha _ { t } } \epsilon \quad ( \mathbf { V } \mathbf { P } ) .
|
| 365 |
+
$$
|
| 366 |
+
|
| 367 |
+
To adjust for this in ΠGDM, we need to scale the guidance term by $\sqrt { \alpha _ { t } }$ (Dhariwal & Nichol, 2021).
|
| 368 |
+
We list the full algorithm for ΠGDM for VP-SDE in Algorithm 1. # "H_pinv": (b, m) $- >$ (b, n), "H": (b, n) $- >$ (b, m) are functions over batches # "y" has shape (b, m); "x_t" has shape (b, n)
|
| 369 |
+
# "hatx_t" with shape (b, n) is the solution from the denoiser
|
| 370 |
+
hatx_t $=$ denoise(x_t, sigma_t)
|
| 371 |
+
# Compute the fixed coefficient; "mat" has shape (b, n)
|
| 372 |
+
mat $=$ H_pinv(y) - H_pinv(H(hatx_t))
|
| 373 |
+
# Compute the inner product between "hatx_t" and "mat", and then sum over batch mat_x $=$ (mat.detach() $^ *$ hatx_t).sum()
|
| 374 |
+
# Compute the guidance term (without r_t).
|
| 375 |
+
guidance $=$ torch.autograd.grad(mat_x, x_t)[0]
|
| 376 |
+
|
| 377 |
+
Listing 1: Pseudocode for computing the pseudoinverse guidance for the noiseless case.
|
| 378 |
+
|
| 379 |
+
# Algorithm 1 ΠGDM for VP-SDE.
|
| 380 |
+
|
| 381 |
+
<table><tr><td colspan="4">Inputs: y, h(x) (noiseless)or H,σy (noisy), Xt,η ∈ [0,1],ε-prediction diffusion model. Find a sequence of timesteps {Ui}=o, where Uo = O and UN = T.</td></tr><tr><td>Initialize x ~ N(0,I). fori= N,.,1do</td><td></td><td></td><td></td></tr><tr><td>t←Ui,s←Ui-1 1</td><td></td><td>Get start and end times for this iteration</td><td></td></tr><tr><td>αt↑ 1+0</td><td></td><td></td><td>Get α in VP-SDE</td></tr><tr><td>∈θ ← ∈-prediction(x;t)</td><td></td><td>Predict the (standardized) noise</td><td></td></tr><tr><td>Xt←x-√l-ateθ. Vat</td><td></td><td>Predict the one-step denoised result</td><td></td></tr><tr><td>C1←n√ (1- αt</td><td>1 -αs</td><td>Get coefficients C1, C2 in DDIM</td><td></td></tr><tr><td>αs C2 ←√1-as-c².</td><td>1-αt</td><td></td><td></td></tr><tr><td>if noiseless then</td><td></td><td></td><td></td></tr><tr><td></td><td>g←(h+(y）-h+(h(x）））)</td><td></td><td></td></tr><tr><td>else</td><td></td><td></td><td></td></tr><tr><td></td><td>9←((y-Hxt)(HHT+1)</td><td>T</td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td>end if</td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td>∈ ~ N(0,I)</td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td>Sample i.i.d. Gaussian</td></tr><tr><td></td><td>X←√αsxt+c1∈+c2∈θ+√αtg</td><td></td><td></td></tr><tr><td></td><td></td><td></td><td>>IIGDM update, first three terms are simply DDIM.</td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td>>Additional √αt in front of g comes from VP-SDE</td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td>end for</td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr><tr><td></td><td></td><td></td><td></td></tr></table>
|
| 382 |
+
|
| 383 |
+
# A.2 STATEMENTS ABOUT RECONSTRUCTION GUIDANCE
|
| 384 |
+
|
| 385 |
+
We note that in reconstruction guidance (Ho et al., 2022), the following approximation is made:
|
| 386 |
+
|
| 387 |
+
$$
|
| 388 |
+
p _ { t } ^ { ( \mathrm { R G } ) } ( \mathbf { y } | \mathbf { x } _ { t } ) \approx \mathcal { N } ( H \hat { \mathbf { x } } _ { t } , \sigma _ { t } ^ { 2 } \pmb { I } ) ,
|
| 389 |
+
$$
|
| 390 |
+
|
| 391 |
+
where we omit the $\alpha _ { t }$ term in (Ho et al., 2022) as we use the Variance Exploding (VE) parametrization throughout the paper6. While Ho et al. (2022) only listed super-resolution and inpainting as two examples, the general idea can be extended to any linear $\pmb { H }$ .
|
| 392 |
+
|
| 393 |
+
Transpose, not pseudoinverse. This approximation would lead to the following score function:
|
| 394 |
+
|
| 395 |
+
$$
|
| 396 |
+
\nabla _ { \mathbf x _ { t } } \log p _ { t } ^ { ( \mathrm { R G } ) } ( \mathbf y | \mathbf x _ { t } ) \approx \frac { 1 } { \sigma _ { t } ^ { 2 } } \left( ( H ^ { \top } \mathbf y - H ^ { \top } H \hat { \mathbf x } _ { t } ) ^ { \top } \frac { \partial \hat { \mathbf x } _ { t } } { \partial \mathbf x _ { t } } \right) ^ { \top } ;
|
| 397 |
+
$$
|
| 398 |
+
|
| 399 |
+
which essentially replaces the pseudoinverse term $H ^ { \dagger }$ in Eq. 8 with the transpose term $\pmb { H } ^ { \top }$ (ignoring the differences in the variance approximation).
|
| 400 |
+
|
| 401 |
+
Taking the singular value decomposition over $\pmb { H } = \pmb { U } \Sigma \pmb { V }$ , we have that:
|
| 402 |
+
|
| 403 |
+
$$
|
| 404 |
+
\begin{array} { r l } & { \pmb { H } ^ { \top } \pmb { H } = ( \pmb { U } \pmb { \Sigma } \pmb { V } ) ^ { \top } ( \pmb { U } \pmb { \Sigma } \pmb { V } ) = \pmb { V } ^ { \top } \pmb { \Sigma } ^ { 2 } \pmb { V } } \\ & { \pmb { H } ^ { \dag } \pmb { H } = ( \pmb { U } \pmb { \Sigma } ^ { - 1 } \pmb { V } ) ^ { \top } ( \pmb { U } \pmb { \Sigma } \pmb { V } ) = \pmb { V } ^ { \top } \mathbb { I } ( \pmb { \Sigma } ^ { 2 } ) \pmb { V } . } \end{array}
|
| 405 |
+
$$
|
| 406 |
+
|
| 407 |
+
where $\mathbb { I } ( \Sigma ^ { 2 } )$ is the diagonal matrix which take 1 if the corresponding entry in $\Sigma ^ { 2 }$ is non-zero, and 0 otherwise. Multiplying the former with a vector $\mathbf { x }$ (as in reconstruction guidance) will scale the singular vectors by $\textstyle \sum ^ { 2 }$ where as multiplying the latter (as in pseudoinverse guidance) keeps the scale for all singular vectors that correspond to nonzero singular values. When the measurement matrix has very different singular values (see Fig. 5), reconstruction guidance may improperly rescale the singular vectors, leading to reduced performance compared with pseudoinverse guidance (as in Table 5).
|
| 408 |
+
|
| 409 |
+
|
| 410 |
+
Figure 5: Singular values of the bicubic downsampling and uniform blurring measurement matrix.
|
| 411 |
+
|
| 412 |
+
Consistent approximation of $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ . In reconstruction guidance, the isotropic Gaussian approximation over the distribution $p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ means that the approximation of the distribution $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ will depend on the measurement model. For example, suppose we have two diagonal measurement matrices $D _ { 1 }$ and $D _ { 2 }$ with all positive values along the diagonal, and ${ \cal D } _ { 1 } \ne { \cal D } _ { 2 }$ . If we use $D _ { 1 }$ as the measurement, then ${ \bf x } _ { 0 } = { \cal D } _ { 1 } ^ { - 1 } { \bf y }$ and $p _ { t } ^ { ( \mathrm { R G } ) } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) \approx \mathcal { N } ( \hat { \mathbf { x } } _ { t } , \sigma _ { t } ^ { 2 } D _ { 1 } ^ { - 2 } )$ , but if we use $D _ { 2 }$ as the measurement, then $\mathbf { x } _ { 0 } = \pmb { { D } } _ { 2 } ^ { - 1 } \mathbf { y }$ and $p _ { t } ^ { ( \mathrm { R G } ) } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) \approx \mathcal { N } ( \hat { \mathbf { x } } _ { t } , \sigma _ { t } ^ { 2 } D _ { 2 } ^ { - 2 } )$ , which is different from the earlier approximation. However, conditioned on $\mathbf { x } _ { t }$ , the distribution of $\mathbf { x } _ { \mathrm { 0 } }$ can be inferred from the diffusion model alone, so it should not depend on the measurement model. Therefore, reconstruction guidance does not make a consistent approximation of the distribution $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ ; pseudoinverse guidance, on the other hand, approximates $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ directly from the diffusion model and then approximates $p _ { t } ( \mathbf { y } | \mathbf { x } _ { t } )$ by marginalization of Gaussians.
|
| 413 |
+
|
| 414 |
+
# A.3 ABOUT THE VARIANCE OF THE APPROXIMATION
|
| 415 |
+
|
| 416 |
+
Our approximation for $p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ depends on the variance term $r _ { t }$ , which should depend on the variance of the data distribution. For example, if the data distribution $p _ { 0 } ( \mathbf { x } _ { 0 } ) = \mathcal { N } ( \mathbf { 0 } , I )$ is the standard normal distribution, then from Bayes’ rule we have the following closed-form solution for
|
| 417 |
+
|
| 418 |
+
|
| 419 |
+
Figure 6: Left: $\sigma _ { t }$ as a function of $t$ . Right: the ratio between our guidance weights and the Video Diffusion Models (VDM) weight $w _ { r } = 1$ (Ho et al., 2022) under different $\eta$ values. We take 100 uniformly spaced timesteps (out of a possible of 1000 timesteps).
|
| 420 |
+
|
| 421 |
+
|
| 422 |
+
Figure 7: Two examples that compare our adaptive weight schedule with the different weights $w _ { r }$ in Video Diffusion Models (VDM, (Ho et al., 2022)) on $4 \times$ super-resolution (Bicubic). For fair comparison, ΠG is used for all cases.
|
| 423 |
+
|
| 424 |
+
the posterior:
|
| 425 |
+
|
| 426 |
+
$$
|
| 427 |
+
p _ { t } ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) \propto p _ { 0 } ( \mathbf { x } _ { 0 } ) p _ { t } ( \mathbf { x } _ { t } | \mathbf { x } _ { 0 } ) = \mathcal { N } \left( \frac { \mathbf { x } _ { t } } { \sigma _ { t } ^ { 2 } + 1 } , \frac { \sigma _ { t } ^ { 2 } } { \sigma _ { t } ^ { 2 } + 1 } I \right) ,
|
| 428 |
+
$$
|
| 429 |
+
|
| 430 |
+
so in this case, we have $r _ { 0 } = 0$ and $r _ { T } \approx 1$ . In Ho et al. (2022), $r _ { t }$ is set as $\sigma _ { t }$ ; this is reasonable in when $t 0$ (noise level is small), but unrealistic when $t \to T$ (noise level is much higher than the data variance). Nevertheless, in the noiseless case, we do not have to make the value of $r _ { t }$ explicit, as we can simply rescale the gradient terms using the guidance weights.
|
| 431 |
+
|
| 432 |
+
In the noisy case, the choice of $r _ { t }$ matters more as it interacts with $\sigma _ { \mathbf { y } }$ in Eq. 7. We simply use
|
| 433 |
+
|
| 434 |
+
$$
|
| 435 |
+
r _ { t } = \sqrt { \frac { \sigma _ { t } ^ { 2 } } { \sigma _ { t } ^ { 2 } + 1 } }
|
| 436 |
+
$$
|
| 437 |
+
|
| 438 |
+
from Eq. 17, which provides good empirical results for our noisy inpainting experiments.
|
| 439 |
+
|
| 440 |
+
To see why this is the case, let us take denoising as an example where $H = I$ . When $\sigma _ { t } \ll \sigma _ { \mathbf { y } }$ is small, then rt ≈ σt, and r2t (r2t + σ2y)−1 ≈ σ2t σ−2y becomes small, meaning that the noisy mea- surement provides little impact to the guidance term. Whereas when is large, then , and $r _ { t } ^ { 2 } ( r _ { t } ^ { 2 } + \bar { \sigma _ { \mathbf { y } } ^ { 2 } } ) ^ { - 1 } \approx 1$ , meaning that the guidance term will be impacted by the noisy measurements. The sampling procedure would first guide the unconditional samples towards the noisy measurements, and then perform denoising without overfitting them.
|
| 441 |
+
|
| 442 |
+
# A.4 ABOUT ADAPTIVE GUIDANCE WEIGHTS
|
| 443 |
+
|
| 444 |
+
The sampling updates for the original DDIM paper is derived from the VP-SDE, so we rewrite the updates in the form of VE-SDE used in this paper:
|
| 445 |
+
|
| 446 |
+
$$
|
| 447 |
+
\begin{array} { r l } & { f ( \mathbf { x } _ { t } ; s , t , \eta ) = \hat { \mathbf { x } } _ { t } + \eta c _ { t s } \epsilon + \sigma _ { t } ^ { - 1 } \sqrt { \sigma _ { s } ^ { 2 } - \eta ^ { 2 } c _ { t s } ^ { 2 } } ( \mathbf { x } _ { t } - \hat { \mathbf { x } } _ { t } ) , } \\ & { \qquad = \mathbf { x } _ { t } + \sigma _ { t } ^ { 2 } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } ) + \eta c _ { t s } \epsilon - \sigma _ { t } \sqrt { \sigma _ { s } ^ { 2 } - \eta ^ { 2 } c _ { t s } ^ { 2 } } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } ) } \end{array}
|
| 448 |
+
$$
|
| 449 |
+
|
| 450 |
+
|
| 451 |
+
Figure 8: Two examples that compare our adaptive weight schedule with the different weights $w _ { r }$ in Video Diffusion Models (VDM, (Ho et al., 2022)) on JPEG $\mathrm { Q F } { = } 1 0$ ) restoration.
|
| 452 |
+
|
| 453 |
+
where $\epsilon \sim \mathcal { N } ( 0 , I )$ and
|
| 454 |
+
|
| 455 |
+
$$
|
| 456 |
+
c _ { t s } = \sqrt { \frac { ( \sigma _ { t } ^ { 2 } - \sigma _ { s } ^ { 2 } ) \sigma _ { s } ^ { 2 } } { \sigma _ { t } ^ { 2 } } } ,
|
| 457 |
+
$$
|
| 458 |
+
|
| 459 |
+
corresponds to the $c _ { 1 }$ coefficient in the original DDIM sampler (using $\sigma _ { t }$ (VE) instead of $\alpha _ { t }$ (VP) formuation). Note that $\hat { \mathbf { x } } _ { t } = \mathbf { x } _ { t } + \sigma _ { t } ^ { 2 } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } )$ is the denoised result.
|
| 460 |
+
|
| 461 |
+
In Ho et al. (2022), the guidance is applied to $\hat { \mathbf { x } } _ { t }$ for some constant $w _ { r }$ , such that7:
|
| 462 |
+
|
| 463 |
+
$$
|
| 464 |
+
\hat { \mathbf { x } } _ { t } ^ { ( R G ) } = \mathbf { x } _ { t } + \sigma _ { t } ^ { 2 } \nabla _ { \mathbf { x } _ { t } } \log p _ { t } ( \mathbf { x } _ { t } ) - \frac { w _ { r } } { 2 } \nabla _ { \mathbf { x } _ { t } } \lVert \mathbf { y } - \pmb { H } \hat { \mathbf { x } } _ { t } \rVert _ { 2 } ^ { 2 } ,
|
| 465 |
+
$$
|
| 466 |
+
|
| 467 |
+
Using the DDIM sampler, the update for the next sample replaces $\hat { \mathbf { x } } _ { t }$ in Eq. 18 with $\hat { \mathbf { x } } _ { t } ^ { ( R G ) }$ which further multiples a factor to the guidance term. In our case, we directly add the guidance term to Eq. 19, which is more similar to the approach in Chung et al. (2022b). This would be equivalent to the weights in Ho et al. (2022) if $w _ { t }$ is different for different $t$ , i.e., applying time-dependent guidance weights during sampling (hence being “adaptive”). While it is possible to tune $w _ { r }$ to achieve decent results, we found that different tasks may require different $w _ { t }$ . For example, $w _ { r } = 5$ works well for super-resolution, but suffers from numerical overflow issues in JPEG restoration (see Fig. 8).
|
| 468 |
+
|
| 469 |
+
# A.5 ABOUT THE LOW DYNAMIC RANGE MEASUREMENT FUNCTION
|
| 470 |
+
|
| 471 |
+
Let $h ( \mathbf { x } )$ be a function that reduces the dynamic range of an image. Typically, this consists of a dynamic range clipping stage $h _ { 1 }$ , a non-linear mapping stage $h _ { 2 }$ , and a quantization stage $h _ { 3 }$ (Liu et al., 2020). The non-linear mapping is also known as the camera response function, and it is fair to assume it is invertible (its inverse denoted as $h _ { 2 } ^ { \dagger }$ ).
|
| 472 |
+
|
| 473 |
+
The dynamic range clipping function typically consists of the form $h _ { 1 } ( x ) = \mathrm { { c l i p } } ( x , a , b )$ where $a$ and $b$ are lower and upper clipping ranges; we can define $h _ { 1 } ^ { \dagger }$ as follows:
|
| 474 |
+
|
| 475 |
+
$$
|
| 476 |
+
h _ { 1 } ^ { \dagger } ( y ) = { \left\{ \begin{array} { l l } { y } & { { \mathrm { i f ~ } } y \in [ a , b ] } \\ { a } & { { \mathrm { i f ~ } } y < a } \\ { b } & { { \mathrm { o t h e r w i s e } } } \end{array} \right. } .
|
| 477 |
+
$$
|
| 478 |
+
|
| 479 |
+
Therefore, we can define $h ^ { \dagger } ( { \bf x } ) = h _ { 3 } ^ { \dagger } ( h _ { 2 } ^ { \dagger } ( h _ { 1 } ^ { \dagger } ( { \bf x } ) ) )$ for pseudoinverse guidance. For ease of illustration, we assume $h _ { 2 }$ to be the identity function in our qualitative results, and focus on the clipping and quantization functions; in these cases, we clip images of range $[ - 1 , 1 ]$ to $[ - 0 . 6 , 0 . 6 ]$ , and then quantize 8 bit representations into 4 bits.
|
| 480 |
+
|
| 481 |
+
# A.6 JUSTIFYING OUR APPROXIMATION
|
| 482 |
+
|
| 483 |
+
As discussed in Sec. 3.1, it is computationally infeasible to sample from more exact representations of $p ( x _ { 0 } | x _ { t } )$ (i.e., the diffusion model), so we need to approximate it. A straightforward way is to approximate via variational inference: instead of the multi-step diffusion process, we use a simple Gaussian to approximate it. Let us denote the Gaussian as $q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ ; we can minimize the KL divergence between $q ( \mathbf { x } _ { 0 } | \boldsymbol { x } _ { t } )$ and $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ , which gives us the following objective function:
|
| 484 |
+
|
| 485 |
+
$$
|
| 486 |
+
\operatorname* { m i n } _ { q } \mathbb { E } _ { p ( \mathbf { x } _ { 0 } ) p ( \mathbf { x } _ { t } | \mathbf { x } _ { 0 } ) } [ \mathrm { K L } ( p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) | | q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) ) ] = \operatorname* { m i n } _ { q } \mathbb { E } _ { p ( \mathbf { x } _ { 0 } , \mathbf { x } _ { t } ) } [ \log p ( \mathbf { x } _ { 0 } , \mathbf { x } _ { t } ) - \log q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) ] ,
|
| 487 |
+
$$
|
| 488 |
+
|
| 489 |
+
If we define $q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ as Gaussian with a fixed standard deviation and mean as a function $\mu$ of $\mathbf { x } _ { t }$ then the objective is equivalent to:
|
| 490 |
+
|
| 491 |
+
$$
|
| 492 |
+
\operatorname* { m i n } _ { \mu } \mathbb { E } _ { p ( \mathbf { x } _ { 0 } , \mathbf { x } _ { t } ) } [ \left\| \mu ( \mathbf { x } _ { t } ) - \mathbf { x } _ { 0 } \right\| _ { 2 } ^ { 2 } ] .
|
| 493 |
+
$$
|
| 494 |
+
|
| 495 |
+
which is exactly the denoising score matching $/$ denoising autoencoder objective (Vincent, 2011).
|
| 496 |
+
Therefore, we can use the single step denoiser result as the mean of $q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ .
|
| 497 |
+
|
| 498 |
+
Nevertheless, one might be interested in how “tight” is our approximation. While it is intractable to compare with ground truth $p ( \mathbf { y } | \mathbf { x } _ { t } )$ or even $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ , it is not hard to compare the score functions of $p ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } )$ and our approximation $q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) ^ { \mathrm { 8 } }$ . In fact, denoting the denoiser as $D$ , we have that:
|
| 499 |
+
|
| 500 |
+
$$
|
| 501 |
+
\begin{array} { r l } & { \nabla _ { { \mathbf x } _ { t } } \log p ( { \mathbf x } _ { 0 } | { \mathbf x } _ { t } ) = \nabla _ { { \mathbf x } _ { t } } \log p ( { \mathbf x } _ { t } | { \mathbf x } _ { 0 } ) - \nabla _ { { \mathbf x } _ { t } } \log p ( { \mathbf x } _ { t } ) } \\ & { \qquad = ( { \mathbf x } _ { 0 } - { \mathbf x } _ { t } ) / \sigma _ { t } ^ { 2 } - ( D ( { \mathbf x } _ { t } ) - { \mathbf x } _ { t } ) / \sigma _ { t } ^ { 2 } = ( { \mathbf x } _ { 0 } - D ( { \mathbf x } _ { t } ) ) / \sigma _ { t } ^ { 2 } , } \end{array}
|
| 502 |
+
$$
|
| 503 |
+
|
| 504 |
+
and (overloading “gradient” notation for derivatives)
|
| 505 |
+
|
| 506 |
+
$$
|
| 507 |
+
\nabla _ { \mathbf { x } _ { t } } \log q ( \mathbf { x } _ { 0 } | \mathbf { x } _ { t } ) \propto [ \nabla _ { \mathbf { x } _ { t } } D ( \mathbf { x } _ { t } ) ] ( \mathbf { x } _ { 0 } - D ( \mathbf { x } _ { t } ) ) .
|
| 508 |
+
$$
|
| 509 |
+
|
| 510 |
+
Therefore, the ground truth score is proportional to $\left( \mathbf { x } _ { 0 } - D ( \mathbf { x } _ { t } ) \right)$ whereas our score is proportional to $[ \nabla _ { \mathbf { x } _ { t } } D ( \mathbf { x } _ { t } ) ] ( \mathbf { x } _ { 0 } - D ( \mathbf { x } _ { t } ) )$ . The two terms are different by a left matrix multiply of the gradient $\nabla _ { x _ { t } } D ( x _ { t } )$ . In the literature of plug-and-play methods, a reasonable assumption for the denoiser is that it can be represented as a pseudolinear filter over the input (see Romano et al. (2016) for detailed explanations), so the gradient behaves roughly like a matrix. This suggests that our approximation is reasonably close, at least when the above score functions are concerned.
|
| 511 |
+
|
| 512 |
+
# B EXPERIMENTAL DETAILS
|
| 513 |
+
|
| 514 |
+
# B.1 ADDITIONAL EXPERIMENTAL SETUPS FOR QUANTITATIVE RESULTS
|
| 515 |
+
|
| 516 |
+
$4 \times$ super-resolution Following CCDF (Chung et al., 2021) and SDEdit (Meng et al., 2021), we initialize our sampler with a smaller noise level than the maximum one by adding Gaussian noise to the linearly upsampled image (of size $2 5 6 \times 2 5 6 )$ ), which is chosen to be the one at the 500-th discrete timestep (where the model is trained with a total of 1000 discrete timesteps). Here, we choose 100 iterations and $\eta = 1 . 0$ for ΠGDM. FID is evaluated over the restoration results on the entire ImageNet validation set, and compared against the statistics of the ImageNet training set.
|
| 517 |
+
|
| 518 |
+
The baselines are run as follows. For DDRM (Kawar et al. (2022a)), we used the default hyperparameter settings. For SR3 (Saharia et al. (2021)), we reported the official results from the paper. For ADM-U (Dhariwal & Nichol (2021)), we used their publicly available $6 4 2 5 6$ ImageNet checkpoint and run 100 iterations for each image with the default command.
|
| 519 |
+
|
| 520 |
+
Inpainting For ΠGDM, we use a class-conditional model, initialize our sampler from pure Gaussian noise at the maximum noise level $\sigma _ { T }$ , apply 100 iterations to each image, and set $\eta = 1 . 0$ . Following Saharia et al. (2022a), we evaluate FID over a $1 0 \mathrm { k }$ subset of the ImageNet validation set, and compare against the statistics of the ImageNet validation set.
|
| 521 |
+
|
| 522 |
+
Table 6: $4 \times$ super-resolution results $( P o o l )$ from ΠGDM using the class-unconditional model.
|
| 523 |
+
|
| 524 |
+
<table><tr><td></td><td colspan="2">FID↓</td><td colspan="2">CA 个</td></tr><tr><td>Steps n</td><td>20</td><td>50 100</td><td>20</td><td>50</td><td>100</td></tr><tr><td>0.0</td><td>6.5</td><td>4.4</td><td>4.3 70.0</td><td>70.3</td><td>69.2</td></tr><tr><td>0.5</td><td>7.2</td><td>4.5</td><td>3.9</td><td>71.4</td><td>70.6</td></tr><tr><td>1.0</td><td>10.9</td><td>6.1</td><td>3.8</td><td>70.0 68.2 71.3</td><td>72.3</td></tr></table>
|
| 525 |
+
|
| 526 |
+
Table 7: $4 \times$ super-resolution results (Bicubic) from ΠGDM using the class-conditional model.
|
| 527 |
+
|
| 528 |
+
<table><tr><td></td><td colspan="2">FID↓</td><td colspan="3">CA↑</td></tr><tr><td>Steps m</td><td rowspan="2">20</td><td rowspan="2">50 100</td><td rowspan="2">20</td><td rowspan="2">50</td><td rowspan="2">100</td></tr><tr><td></td></tr><tr><td>0.0</td><td>7.5</td><td>4.1 3.9</td><td>73.0</td><td>74.0</td><td>72.6</td></tr><tr><td>0.5</td><td>8.0</td><td>4.2 3.4</td><td>72.7</td><td>74.6</td><td>73.7</td></tr><tr><td>1.0</td><td>10.8</td><td>5.5</td><td>3.2</td><td>70.8</td><td>74.2 75.1</td></tr></table>
|
| 529 |
+
|
| 530 |
+
JPEG Restoration For each quality factor, we use the quantization matrix in Ehrlich et al. (2020) to compress the original $2 5 6 \times 2 5 6$ image, with $2 \times 2$ chroma subsampling. The quantization matrix is embedded in every JPEG file, so having this available to the algorithm is a natural and realistic setting. For ΠGDM, we use a class-unconditional model, initialize our sampler from pure Gaussian noise at the maximum timestep, apply 100 iterations to each image, and set $\eta = 1 . 0$ . The FID is evaluated as in the inpainting case, following Saharia et al. (2022a).
|
| 531 |
+
|
| 532 |
+
# B.2 ADDITIONAL ABLATION STUDIES AND DETAILS
|
| 533 |
+
|
| 534 |
+
Uniform deblurring We use the uniform $9 \times 9$ deblurring kernel used in Kawar et al. (2022a). The problem itself is relatively simple as it has few non-zero singular values, and simply taking the pseudoinverse over the observations already gives good results. For all methods, we use a classunconditional model, initialize our sampler from the 100-th discrete timestep using the CCDF / SDEdit approach. We use a total of 20 iterations for each image and $\eta = 0 . 5$ for the 1000 images used in the evaluation set of DGP $/$ DDRM (Pan et al., 2021; Kawar et al., 2022a). We compare PSNR metrics with images scaled to $[ 0 , 1 ]$ , and Kernel Inception Distance (KID, (Binkowski et al. ´ , 2018)) metrics against the 1000 reference images, following the practice in Kawar et al. (2022a).
|
| 535 |
+
|
| 536 |
+
$4 \times$ super-resolution The experiment setup is identical to the quantitative evaluation case, except that we evaluate the metrics over the $1 0 \mathrm { k }$ subset from Saharia et al. (2022a). We use a total of 100 iterations for each images and $\eta = 1 . 0$ for the 10000 images.
|
| 537 |
+
|
| 538 |
+
Ablation study over $\eta$ and number of iterations We report additional results over the hyperparameters in the DDIM sampler, which are $\eta$ (the amount of noise injected at each step) and the number of iterations (steps) per image on super-resolution tasks. We consider the Pool and Bicubic $4 \times$ super-resolution task over the entire ImageNet validation set. From the results in Tabs. 6 and 7, we can draw similar conclusions as the ones from the DDIM paper (Song et al., 2021a): more iterations generally lead to improved performance, whereas the effect of $\eta$ varies. When the number of iterations is small, smaller $\eta$ is better (as it injects less noise in the process); when the opposite is true, larger $\eta$ is better (due to the sampling process being more robust to errors in the score function).
|
| 539 |
+
|
| 540 |
+
# B.3 ADDITIONAL FIGURES
|
| 541 |
+
|
| 542 |
+
We list additional qualitative results in Figs. 9 to 15. All the results with ΠGDM and reconstruction guidance are generated with 100 steps and $\eta = 1 . 0$ . We use $w _ { r } = 1$ for reconstruction guidance.
|
| 543 |
+
|
| 544 |
+
|
| 545 |
+
Figure 9: Comparing methods for the Pool $4 \times$ super-resolution problem, including reconstruction guidance (Ho et al., 2022), ADM-U (Dhariwal & Nichol, 2021), and ΠGDM. Best viewed zoomed in.
|
| 546 |
+
|
| 547 |
+
|
| 548 |
+
Figure 10: Comparing methods for the Bicubic $4 \times$ super-resolution problem, including reconstruction guidance (Ho et al., 2022), ADM-U (Dhariwal & Nichol, 2021), and ΠGDM. Best viewed zoomed in.
|
| 549 |
+
|
| 550 |
+
|
| 551 |
+
Figure 11: Inpainting results using ΠGDM for the Freeform problem with multiple random samples.
|
| 552 |
+
|
| 553 |
+
|
| 554 |
+
Figure 12: Results for the JPEG restoration problem, including Palette (Saharia et al., 2022a) and ΠGDM. Palette results are obtained from the official website.
|
| 555 |
+
|
| 556 |
+
|
| 557 |
+
Figure 13: Restoration results with ΠGDM over various composed measurements. The same random seed is used for different problems.
|
| 558 |
+
|
| 559 |
+
|
| 560 |
+
Figure 14: Inpainting results with ΠGDM for the Center problem on the LSUN Bedroom dataset (Yu et al., 2015).
|
| 561 |
+
|
| 562 |
+
|
| 563 |
+
Figure 15: Super-resolution results with ΠGDM for the Pool case on the LSUN Bedroom dataset (Yu et al., 2015).
|
| 564 |
+
|
| 565 |
+
Table 8: NFEs of various algorithms. We list common baselines, such as Palette (Saharia et al., 2022a), ADM-U (Dhariwal & Nichol, 2021), SR3 (Saharia et al., 2021), DGP (Pan et al., 2021), SNIPS (Kawar et al., 2021), RED (Romano et al., 2016), DDRM (Kawar et al., 2022a), and DPS (Chung et al., 2022a).
|
| 566 |
+
|
| 567 |
+
<table><tr><td>IIGDM (Ours)</td><td></td><td>Palette</td><td>Regression</td><td>ADM-U</td><td>SR3</td><td>DPS</td><td>DGP</td><td>SNIPS</td><td>RED</td><td>DDRM</td></tr><tr><td>NFEs</td><td>20 to 100</td><td>1000</td><td>1</td><td>100</td><td>1000</td><td>1000</td><td>1500</td><td>1000</td><td>500</td><td>20</td></tr></table>
|
| 568 |
+
|
| 569 |
+
Table 9: $4 \times$ super-resolution (Pool) and deblurring results on ImageNet 1K $( 2 5 6 \times 2 5 6 )$ .
|
| 570 |
+
|
| 571 |
+
<table><tr><td rowspan="2">Method</td><td colspan="3">4× super-resolution (Pool)</td><td colspan="3">Deblurring</td></tr><tr><td>SSIM↑</td><td>KID↓</td><td>NFEs↓</td><td>PSNR↑</td><td>KID↓</td><td>NFEs↓</td></tr><tr><td>Regression</td><td>0.71</td><td>44.90</td><td>0</td><td>19.26</td><td>38.00</td><td>0</td></tr><tr><td>DGP</td><td>0.56</td><td>21.22</td><td>1500</td><td>22.70</td><td>27.60</td><td>1500</td></tr><tr><td>RED</td><td>0.73</td><td>53.55</td><td>100</td><td>26.16</td><td>21.21</td><td>500</td></tr><tr><td>SNIPS</td><td>0.22</td><td>35.17</td><td>1000</td><td>34.32</td><td>0.49</td><td>1000</td></tr><tr><td>DDRM</td><td>0.72</td><td>7.22</td><td>20</td><td>35.64</td><td>0.71</td><td>20</td></tr><tr><td>IIGDM (Ours)</td><td>0.73</td><td>1.24</td><td>100</td><td>39.36</td><td>0.00</td><td>20</td></tr></table>
|
| 572 |
+
|
| 573 |
+
# B.4 RUNTIME OF DIFFERENT ALGORITHMS
|
| 574 |
+
|
| 575 |
+
Since the algorithm requires iterations with the diffusion model, the actual runtime of the algorithm would depend heavily on the number of Neural Function Evaluations (NFEs). Kawar et al.Kawar et al. (2022a) found that the runtime for diffusion models would dominate the total runtime, as other computations (such as matrix operations on images) are negligible. Therefore, we use NFE as the unit for estimating the runtime of different algorithms. In Tab. 8, we report the NFEs used by each algorithm.
|
| 576 |
+
|
| 577 |
+
We further note that ΠGDM and DPS take additional backpropagation steps through the neural network, so each NFE is roughly $3 \times$ as expensive as others. Thus, ΠGDM is only beaten in terms of actual wall-clock time by regression and DDRM, and we have shown that it has superior restoration results than both in Tabs. 2 to 4.
|
| 578 |
+
|
| 579 |
+
# B.5 COMPARISON WITH VARIOUS BASELINES
|
| 580 |
+
|
| 581 |
+
In Tab. 9, we compare ΠGDM against various baselines, including ones that are based on Plugand-play (PnP) methods (Venkatakrishnan et al., 2013b), such as Deep Generative Prior (Pan et al. (2021)), Regularizing by Denoising (RED, Romano et al. (2016)), Solving Noisy Inverse Problems Stochastically (SNIPS, Kawar et al. (2021)), and Denoising Diffusion Restoration Models (DDRM, Kawar et al. (2022a)). Similar to the setting for deblurring in Tab. 5, the comparison is done on the 1000 validation examples listed in the DGP paper (Pan et al., 2021), and the KID are computed with the 1000 ground truth images as the reference set.
|
| 582 |
+
|
| 583 |
+
We find that ΠGDM achieves the best performance when compared with other PnP baselines; this is reasonable given that DDRM (Kawar et al., 2022a) is shown to outperform the remaining competitors, and ΠGDM outperforms DDRM from the results in 2.
|
| 584 |
+
|
| 585 |
+
# B.6 COMPARISON WITH DIFFUSION POSTERIOR SAMPLING
|
| 586 |
+
|
| 587 |
+
Diffusion Posterior Sampling (DPS, (Chung et al., 2022a)) is a concurrent method that similarly uses a gradient-based guidance method. They approximate $p ( \mathbf { y } | \mathbf { x } _ { t } )$ with $p ( \mathbf { y } \vert \mathbf { x } _ { 0 } ) \vert _ { \mathbf { x } _ { 0 } = \hat { \mathbf { x } } _ { t } }$ , which is similar to reconstruction guidance.
|
| 588 |
+
|
| 589 |
+
Findings. We investigate the performance of DPS over several tasks and over different learning rates and diffusion step hyperparameters. Similar to what Chung et al. has found (Chung et al.,
|
| 590 |
+
|
| 591 |
+
Table $1 0 \colon 4 \times$ super-resolution (Pool) comparisons with DPS, under different steps.
|
| 592 |
+
|
| 593 |
+
<table><tr><td>Steps</td><td></td><td>20</td><td>50</td><td>100</td><td>200</td><td>500</td><td>1000</td></tr><tr><td>LPIPS</td><td>DPS IIGDM</td><td>0.559 0.164</td><td>0.511 0.140</td><td>0.414 0.140</td><td>0.335</td><td>0.215</td><td>0.162</td></tr><tr><td>SSIM</td><td>DPS IIGDM</td><td>0.504 0.779</td><td>0.580 0.772</td><td>0.634 0.777</td><td>1 0.696</td><td>1 0.756</td><td>1 0.778 -</td></tr></table>
|
| 594 |
+
|
| 595 |
+
Table $1 1 \colon 4 \times$ super-resolution (Bicubic) comparisons with DPS, under different steps.
|
| 596 |
+
|
| 597 |
+
<table><tr><td>Steps</td><td></td><td>20</td><td>50</td><td>100</td><td>200</td><td>500</td><td>1000</td></tr><tr><td>LPIPS</td><td>DPS IIGDM</td><td>0.558 0.163</td><td>0.510 0.166</td><td>0.410 0.122</td><td>0.329 1</td><td>0.221 1</td><td>0.171 1</td></tr><tr><td>SSIM</td><td>DPS IIGDM</td><td>0.504 0.789</td><td>0.581 0.730</td><td>0.637 0.773</td><td>0.697 1</td><td>0.757 1</td><td>0.779 1</td></tr></table>
|
| 598 |
+
|
| 599 |
+
2022a), we find that while DPS has strong performance with 1000 diffusion steps, its performance becomes much worse with less diffusion steps. Moreover, even with the full 1000 steps, one would still need to tune for the learning rate hyperparameter to get good performance. In contrast, our method is $1 0 \times$ faster than DPS in the slowest case, and achieves decent performance even with fewer diffusion steps (such as 20 and 50 steps). In these settings, DPS does not produce reasonable results at all.
|
| 600 |
+
|
| 601 |
+
Image restoration quality. In our experiments, we evaluate performance of the models averaged over 5 validation images on ImageNet. For DPS, we consider different numbers of diffusion steps (from 20 to 1000), with the default learning rate being the one chosen in the DPS paper for 1000 steps. For ΠGDM, we use the same settings as discussed in the paper; we report for diffusion steps up to 100 steps, except for deblurring (where the task is simple enough to get good results in 20 steps).
|
| 602 |
+
|
| 603 |
+
We report LPIPS and SSIM metrics for Pool, Bicubic, and $9 \times 9$ uniform deblurring in Tabs. 10 to 12. From the tables, it is obvious that DPS performance drops rapidly once number of diffusion steps decreases under 1000, whereas ΠGDM performance remain competitive. We illustrate this trend visually in Fig. 16 for $4 \times$ super-resolution $( P o o l )$ . In Tab. 13, we report the results for DPS under different learning rate hyperparameters; we found that the optimal hyperparameter can be problem-dependent: the optimal one for super-resolution is around 1 and 2, whereas using that for deblurring will produce NaNs; the optimal one for deblurring is 0.2, where super-resolution results tend to become less optimal.
|
| 604 |
+
|
| 605 |
+
Loss curves. We can treat the guidance terms in both DPS and ΠGDM as a gradient step that optimizes the least squares loss function $\| \mathbf { y } - H \mathbf { x } _ { 0 } \| _ { 2 } ^ { 2 }$ , so it is natural to visualize the loss at each diffusion noise level. In Fig. 17, we visualize the loss curve of DPS and ΠGDM on a $4 \times$ superresolution (Pool) example as a function of the diffusion timestep level (so 1000 is highest noise level and 0 is lowest noise level). As expected, both methods start with the same loss. However, the loss function of DPS significantly increases around timesteps 1000 and 900, reaching $1 0 0 0 \times$ of the initial loss; this means that the DPS learning rate schedule is too large at this initial stage. In contrast, ΠGDM has a smooth loss curve over the entire process; in fact, the loss curves are quite consistent under 20, 50, or 100 steps. Moreover, the final loss for ΠGDM is also smaller than that of DPS, further illustrating its superiority.
|
| 606 |
+
|
| 607 |
+
Table 12: $9 \times 9$ uniform deblurring comparisons with DPS, under different steps.
|
| 608 |
+
|
| 609 |
+
<table><tr><td>Steps</td><td></td><td>20</td><td>50</td><td>100</td><td>200</td><td>500</td><td>1000</td></tr><tr><td>LPIPS</td><td>DPS IIGDM</td><td>0.539 0.004</td><td>0.472 1</td><td>0.412</td><td>0.340</td><td>0.260 1</td><td>0.245</td></tr><tr><td>SSIM</td><td>DPS IIGDM</td><td>0.516 0.974</td><td>0.575 1</td><td>0.604 1</td><td>0.662 1</td><td>0.698 1</td><td>1 0.719 1</td></tr></table>
|
| 610 |
+
|
| 611 |
+
Table 13: Performance of DPS under different learning rates and 1000 diffusion steps.
|
| 612 |
+
|
| 613 |
+
<table><tr><td rowspan="2">Task Learning rate</td><td colspan="4">Super-resolution (Pool)</td><td colspan="4">Deblurring</td></tr><tr><td>0.5</td><td>1.0</td><td>2.0</td><td>0.2</td><td></td><td>0.4</td><td>0.6</td></tr><tr><td>LPIPS</td><td>0.208</td><td>0.171</td><td>0.182</td><td>0.238</td><td></td><td>0.245</td><td>0.322</td></tr><tr><td>SSIM</td><td>0.755</td><td>0.779</td><td>0.775</td><td>0.700</td><td></td><td>0.719</td><td>0.608</td></tr></table>
|
| 614 |
+
|
| 615 |
+
|
| 616 |
+
Figure 16: Super-resolution results with DPS (Chung et al., 2022a) and ΠGDM for the $4 \times$ superresolution (Pool) under different number of diffusion steps.
|
| 617 |
+
|
| 618 |
+
|
| 619 |
+
Figure 17: Loss curves for $\| \mathbf { y } - H \mathbf { x } _ { 0 } \| _ { 2 } ^ { 2 }$ using DPS (Chung et al., 2022a) and ΠGDM for a $4 \times$ super-resolution $( P o o l )$ example.
|
parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/9_gsMA8MRKQ/9_gsMA8MRKQ_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/FQOC5u-1egI/FQOC5u-1egI.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/FQOC5u-1egI/FQOC5u-1egI_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/FQOC5u-1egI/FQOC5u-1egI_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/FQOC5u-1egI/FQOC5u-1egI_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/GQjaI9mLet/GQjaI9mLet.md
ADDED
|
@@ -0,0 +1,493 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# INDEPENDENT SE(3)-EQUIVARIANT MODELS FOREND-TO-END RIGID PROTEIN DOCKING
|
| 2 |
+
|
| 3 |
+
Octavian-Eugen Ganea†∗ MIT
|
| 4 |
+
|
| 5 |
+
Xinyuan Huang§∗ ETH Zurich
|
| 6 |
+
|
| 7 |
+
Charlotte Bunne ETH Zurich
|
| 8 |
+
|
| 9 |
+
Yatao Bian† Tencent AI Lab
|
| 10 |
+
|
| 11 |
+
Regina Barzilay MIT
|
| 12 |
+
|
| 13 |
+
Tommi Jaakkola MIT
|
| 14 |
+
|
| 15 |
+
Andreas Krause ETH Zurich
|
| 16 |
+
|
| 17 |
+
# ABSTRACT
|
| 18 |
+
|
| 19 |
+
Protein complex formation is a central problem in biology, being involved in most of the cell’s processes, and essential for applications, e.g. drug design or protein engineering. We tackle rigid body protein-protein docking, i.e., computationally predicting the 3D structure of a protein-protein complex from the individual unbound structures, assuming no conformational change within the proteins happens during binding. We design a novel pairwise-independent SE(3)-equivariant graph matching network to predict the rotation and translation to place one of the proteins at the right docked position relative to the second protein. We mathematically guarantee a basic principle: the predicted complex is always identical regardless of the initial locations and orientations of the two structures. Our model, named EQUIDOCK, approximates the binding pockets and predicts the docking poses using keypoint matching and alignment, achieved through optimal transport and a differentiable Kabsch algorithm. Empirically, we achieve significant running time improvements and often outperform existing docking software despite not relying on heavy candidate sampling, structure refinement, or templates.
|
| 20 |
+
|
| 21 |
+
# 1 INTRODUCTION
|
| 22 |
+
|
| 23 |
+
In a recent breakthrough, ALPHAFOLD 2 (Jumper et al., 2021; Senior et al., 2020) provides a solution to a grand challenge in biology—inferring a protein’s three-dimensional structure from its amino acid sequence (Baek et al., 2021), following the dogma sequence determines structure.
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
Figure 1: Different views of the 3D structure of a protein complex. a. Surface and b. cartoon view of protein Z and its inhibitor.
|
| 27 |
+
|
| 28 |
+
Besides their complex three-dimensional nature, proteins dynamically alter their function and structure in response to cellular signals, changes in the environment, or upon molecular docking. In par
|
| 29 |
+
|
| 30 |
+
ticular, protein interactions are involved in various biological processes including signal transduction, protein synthesis, DNA replication and repair. Molecular docking is key to understanding protein interactions’ mechanisms and effects, and, subsequently, to developing therapeutic interventions.
|
| 31 |
+
|
| 32 |
+
We here address the problem of rigid body protein-protein docking which refers to computationally predicting the 3D structure of a protein-protein complex given the 3D structures of the two proteins in unbound state. Rigid body means no deformations occur within any protein during binding, which is a realistic assumption in many biological settings.
|
| 33 |
+
|
| 34 |
+
Popular docking software (Chen et al., 2003; Venkatraman et al., 2009; De Vries et al., 2010; Torchala et al., 2013; Schindler et al., 2017; Sunny and Jayaraj, 2021) are typically computationally expensive, taking between minutes and hours to solve a single example pair, while not being guaranteed to find accurate complex structures. These methods largely follow the steps: i.) randomly sample a large number (e.g., millions) of candidate initial complex structures, ii.) employ a scoring function to rank the candidates, iii.) adjust and refine the top complex structures based on an energy model (e.g., force field). We here take a first step towards tackling these issues by using deep learning models for direct prediction of protein complex structures.
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
Figure 2: Same output guarantee of EQUIDOCK. We predict a rigid transformation to place the ligand in the binding location w.r.t the receptor. We mathematically guarantee to output the same complex structure — up to an SE(3) transformation — independently of the initial unbound positions, rotations, or roles of both constituents. (RMSD $=$ Root-mean-square deviation of atomic positions)
|
| 38 |
+
|
| 39 |
+
Contributions. We design EQUIDOCK, a fast, end-to-end method for rigid body docking that directly predicts the SE(3) transformation to place one of the proteins (ligand) at the right location and orientation with respect to the second protein (receptor). Our method is based on the principle that the exact same complex structure should be predicted irrespectively of the initial 3D placements and roles of both constituents (see Fig. 2). We achieve this desideratum by incorporating the inductive biases of pairwise SE(3)–equivariance and commutativity, and deriving novel theoretical results for necessary and sufficient model constraints (see Section 3). Next, we create EQUIDOCK to satisfy these properties by design, being a combination of: i) a novel type of pairwise independent SE(3)-equivariant graph matching networks, ii) an attention-based keypoint selection algorithm that discovers representative points and aligns them with the binding pocket residues using optimal transport, and iii) a differentiable superimposition model to recover the optimal global rigid transformation. Unlike prior work, our method does not use heavy candidate sampling or ranking, templates, task-specific geometric or chemical hand-crafted features, or pre-computed meshes. This enables us to achieve plausible structures with a speed-up of $8 0 - 5 0 0 \mathrm { x }$ compared to popular docking software, offering a promising competitive alternative to current solutions for this problem.
|
| 40 |
+
|
| 41 |
+
# 2 RELATED WORK
|
| 42 |
+
|
| 43 |
+
Geometric Deep Learning. Graph Neural Networks (GNNs) are becoming the de facto choice for learning with graph data (Bruna et al., 2013; Defferrard et al., 2016; Kipf and Welling, 2016; Gilmer et al., 2017; Xu et al., 2018; Li et al., 2019). Motivated by symmetries naturally occurring in different data types, architectures are tailored to explicitly incorporate such properties (Cohen and Welling, 2016a;b; Thomas et al., 2018; Fuchs et al., 2020; Finzi et al., 2020; Eismann et al., 2020; Satorras et al., 2021). GNNs are validated in a variety of tasks such as particle system dynamics or conformation-based energy estimation (Weiler and Cesa, 2019; Rezende et al., 2019).
|
| 44 |
+
|
| 45 |
+
Euclidean Neural Networks (E(3)-NNs). However, plain GNNs and other deep learning methods do not understand data naturally lying in the 3D Euclidean space. For example, how should the output deterministically change with the input, e.g. when it is rotated ? The recent Euclidean neural networks address this problem, being designed from geometric first-principles. They make use of SE(3)- equivariant and invariant neural layers, thus avoiding expensive data augmentation strategies. Such constrained models ease optimization and have shown important improvements in biology or chemistry – e.g. for molecular structures (Fuchs et al., 2020; Hutchinson et al., 2020; Wu et al., 2021;
|
| 46 |
+
|
| 47 |
+
Jumper et al., 2021; Ganea et al., 2021) and different types of 3D point clouds (Thomas et al., 2018). Different from prior work, we here derive constraints for pairs of 3D objects via pairwise independent SE(3)-equivariances, and design a principled approach for modeling rigid body docking.
|
| 48 |
+
|
| 49 |
+
Protein Folding. Deep neural networks have been used to predict inter-residue contacts, distance and/or orientations (Adhikari and Cheng, 2018; Yang et al., 2020; Senior et al., 2020; Ju et al., 2021), that are subsequently transformed into additional constraints or differentiable energy terms for protein structure optimization. ALPHAFOLD 2 (Jumper et al., 2021) and Rosetta Fold (Baek et al., 2021) are state-of-the-art approaches, and directly predict protein structures from co-evolution information embedded in homologous sequences, using geometric deep learning and E(3)-NNs.
|
| 50 |
+
|
| 51 |
+
Protein-Protein Docking and Interaction. Experimentally determining structures of protein complexes is often expensive and time-consuming, rendering a premium on computational methods (Vakser, 2014). Protein docking methods (Chen et al., 2003; Venkatraman et al., 2009; De Vries et al., 2010; Biesiada et al., 2011; Torchala et al., 2013; Schindler et al., 2017; Weng et al., 2019; Sunny and Jayaraj, 2021; Christoffer et al., 2021; Yan et al., 2020) typically run several steps: first, they sample thousands or millions of complex candidates; second, they use a scoring function for ranking (Moal et al., 2013; Basu and Wallner, 2016; Launay et al., 2020; Eismann et al., 2020); finally, top-ranked candidates undergo a structure refinement process using energy or geometric models (Verburgt and Kihara, 2021). Relevant to protein-protein interaction (PPI) is the task of protein interface prediction where GNNs have showed promise (Fout et al., 2017; Townshend et al., 2019; Liu et al., 2020; Xie and Xu, 2021; Dai and Bailey-Kellogg, 2021). Recently, ALPHAFOLD 2 and ROSETTAFOLD have been utilized as subroutines to improve PPIs from different aspects (Humphreys et al., 2021; Pei et al., 2021; Jovine), e.g., combining physics-based docking method CLUSPRO (Kozakov et al., 2017; Ghani et al., 2021), or using extended multiple-sequence alignments to predict the structure of heterodimeric protein complexes from the sequence information (Bryant et al., 2021). Concurrently to our work, Evans et al. (2021) extend ALPHAFOLD 2 to multiple chains during both training and inference.
|
| 52 |
+
|
| 53 |
+
Drug-Target Interaction (DTI). DTI aims to compute drug-target binding poses and affinity, playing an essential role in understanding drugs’ mechanism of action. Prior methods (Wallach et al., 2015; Li et al., 2021) predict binding affinity from protein-ligand co-crystal structures, but such data is expensive to obtain experimentally. These models are typically based on heavy candidate sampling and ranking (Trott and Olson, 2010; Koes et al., 2013; McNutt et al., 2021; Bao et al., 2021), being tailored for small drug-like ligands and often assuming known binding pocket. Thus, they are not immediately applicable to our use case. In contrast, our rigid docking approach is generic and could be extended to accelerate DTI research as part of future work.
|
| 54 |
+
|
| 55 |
+
# 3 MATHEMATICAL CONSTRAINTS FOR RIGID BODY DOCKING
|
| 56 |
+
|
| 57 |
+
We start by introducing the rigid body docking problem and derive the geometric constraints for enforcing same output complex prediction regardless of the initial unbound positions or roles (Fig. 2).
|
| 58 |
+
|
| 59 |
+
Rigid Protein-Protein Docking – Problem Setup. We are given as input a pair of proteins forming a complex. They are (arbitrarily) denoted as the ligand and receptor, consisting of $n$ and $m$ residues, respectively. These proteins are represented in their bound (docked) state as 3D point clouds $\mathbf { X } _ { 1 } ^ { * } \in \mathbb { R } ^ { 3 \times n }$ , $\mathbf { X } _ { 2 } ^ { \ast } \in \mathbb { R } ^ { 3 \times m }$ , where each residue’s location is given by the coordinates of its corresponding $\alpha$ -carbon atom. In the unbound state, the docked ligand is randomly rotated and translated in space, resulting in a modified point cloud $\mathbf { X } _ { 1 } \in \mathbb { R } ^ { 3 \times n }$ . For simplicity and w.l.o.g., the receptor remains in its bound location ${ \bf X } _ { 2 } = { \bf X } _ { 2 } ^ { * }$ .
|
| 60 |
+
|
| 61 |
+
The task is to predict a rotation $\mathbf { R } \in S O ( 3 )$ and a translation $\mathbf { t } \in \mathbb { R } ^ { 3 }$ such that $\mathbf { R } \mathbf { X } _ { 1 } + \mathbf { t } = \mathbf { X } _ { 1 } ^ { * }$ using as input the proteins and their unbound positions $\mathbf { X } _ { 1 }$ and $\mathbf { X } _ { 2 }$ .
|
| 62 |
+
|
| 63 |
+
Here, $\mathbf { R } = \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ and $\mathbf { t } = \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ are functions of the two proteins, where we omit residue identity or other protein information in this notation, for brevity.
|
| 64 |
+
|
| 65 |
+
Note that we assume rigid backbone and side-chains for both proteins. We therefore do not tackle the more challenging problem of flexible docking, but our approach offers an important step towards it.
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
Figure 3: Details on EQUIDOCK’s Architecture and Losses. a. The message passing operations in KÉÏIEGMN guarantee pairwise independent SE(3)-equivariance as in Eq. (4), b. We predict keypoints for HÉeach protein that are aligned with the binding pocket location using an additional optimal transport ËÏÊï PIRtb ÉLÉÏ r.tt ï ÉLÉÏITE ï(OT) loss, c. After predicting the docked position, we compute an MSE loss on the ligand, as well as a loss to discourage body intersections.
|
| 69 |
+
|
| 70 |
+
We desire that the predicted complex structure is independent of the initial locations and orientations of the two proteins, as well as of their roles – see Fig. 2. Formally, we wish to guarantee that:
|
| 71 |
+
|
| 72 |
+
$\begin{array} { r l r } & { } & { \left( { \bf R } ( { \bf Z } _ { 1 } | { \bf Z } _ { 2 } ) { \bf Z } _ { 1 } + { \bf t } ( { \bf Z } _ { 1 } | { \bf Z } _ { 2 } ) \right) \oplus { \bf Z } _ { 2 } \equiv \left( { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) \right) \oplus { \bf X } _ { 2 } , \quad ( { \bf t } ( { \bf X } _ { 2 } ) \mid { \bf X } _ { 2 } ) = { \bf 0 } , } \\ & { } & { \left( { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) \right) \oplus { \bf X } _ { 2 } \equiv { \bf X } _ { 1 } \oplus \left( { \bf R } ( { \bf X } _ { 2 } | { \bf X } _ { 1 } ) { \bf X } _ { 2 } + { \bf t } ( { \bf X } _ { 2 } | { \bf X } _ { 1 } ) \right) , } \\ & { } & { \forall { \bf Q } _ { 1 } , { \bf Q } _ { 2 } \in S O ( 3 ) , \forall { \bf g } _ { 1 } , { \bf g } _ { 2 } \in { \mathbb R } ^ { 3 } , \forall { \bf X } _ { 1 } \in { \mathbb R } ^ { 3 \times n } , { \bf X } _ { 2 } \in { \mathbb R } ^ { 3 \times m } , \ \mathrm { a n d } { \bf Z } _ { l } = { \bf G } _ { 3 } . } \end{array}$ (SE(3)-invariance) (commutativity) ${ \bf Z } _ { l } = { \bf Q } _ { l } { \bf X } _ { l } + { \bf g } _ { l } , l \in \{ 1 , 2 \}$ for any rotations $\mathbf { Q } _ { 1 } , \mathbf { Q } _ { 2 }$ and translations $\mathbf { g } _ { 1 } , \mathbf { g } _ { 2 }$ , where $\oplus$ is concatenation along columns, and $\equiv$ denotes identity after superimposition, i.e. zero Root-Mean-Square Deviation (RMSD) between the two 3D point sets after applying the Kabsch algorithm (Kabsch, 1976). An immediate question arises:
|
| 73 |
+
|
| 74 |
+
How do the constraints in Eq. (1) translate into constraints for $\mathbf { R } ( \cdot | \cdot )$ and $\mathbf { t } ( \cdot | \cdot )$ ?
|
| 75 |
+
|
| 76 |
+
The rotation $\mathbf { R }$ and translation t change in a systematic way when we apply $S E ( 3 )$ transformations or swap proteins’ roles. These properties restrict our class of functions as derived below.
|
| 77 |
+
|
| 78 |
+
SE(3)-equivariance Constraints. If we apply any distinct $S E ( 3 )$ transformations on the unbound ligand $\mathbf { X } _ { 1 }$ and receptor $\mathbf { X } _ { 2 }$ , i.e. we dock $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ onto $\mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 }$ , then the rotation matrix $\mathbf { R } ( \mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 } | \mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 } )$ and translation vector $\mathbf { t } ( \mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 } | \mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 } )$ can be derived from the original $\mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ and $\mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ assuming that we always do rotations first. In this case, $\mathbf { R } ( \mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 } | \mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 } )$ can be decomposed into three rotations: i.) apply $\mathbf { Q } _ { 1 } ^ { \top }$ to undo the rotation $\mathbf { Q } _ { 1 }$ applied on $\mathbf { X } _ { 1 }$ , ii.) apply ${ \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } )$ , iii.) apply $\mathbf { Q } _ { 2 }$ to rotate the docked ligand together with the receptor. This gives ${ \bf R } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } | { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } ) = { \bf Q } _ { 2 } { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf Q } _ { 1 } ^ { \top }$ , which in turn constraints the translation vector. We provide a formal statement and prove it in Appendix B.1:
|
| 79 |
+
|
| 80 |
+
Proposition 1. For any $\mathbf { Q } _ { 1 } , \mathbf { Q } _ { 2 } \in S O ( 3 ) , \mathbf { g } _ { 1 } , \mathbf { g } _ { 2 } \in \mathbb { R } ^ { 3 }$ , $S E ( 3 )$ -invariance of the predicted docked complex defined by Eq. (1) is guaranteed iff
|
| 81 |
+
|
| 82 |
+
$$
|
| 83 |
+
\begin{array} { r l } & { { \bf R } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } | { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } ) = { \bf Q } _ { 2 } { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf Q } _ { 1 } ^ { \top } } \\ & { { \bf t } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } | { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } ) = { \bf Q } _ { 2 } { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) - { \bf Q } _ { 2 } { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf Q } _ { 1 } ^ { \top } { \bf g } _ { 1 } + { \bf g } _ { 2 } . } \end{array}
|
| 84 |
+
$$
|
| 85 |
+
|
| 86 |
+
As a direct consequence of this proposition, we have the following statement.
|
| 87 |
+
|
| 88 |
+
Proposition 2. Any model satisfying Proposition $I$ guarantees invariance of the predicted complex w.r.t. any $S E ( 3 )$ transformation on $\mathbf { X } _ { 1 }$ , and equivariance w.r.t. any $S E ( 3 )$ transformation on $\mathbf { X } _ { 2 }$ :
|
| 89 |
+
|
| 90 |
+
$$
|
| 91 |
+
\begin{array} { r l } & { { \bf R } ( { \bf Z } _ { 1 } | { \bf X } _ { 2 } ) { \bf Z } _ { 1 } + { \bf t } ( { \bf Z } _ { 1 } | { \bf X } _ { 2 } ) = { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) , \quad w h e r e { \bf Z } _ { 1 } = { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } } \\ & { { \bf R } ( { \bf X } _ { 1 } | { \bf Z } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf Z } _ { 2 } ) = { \bf Q } _ { 2 } \left[ { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) \right] + { \bf g } _ { 2 } , \quad w h e r e { \bf Z } _ { 2 } = { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } } \\ & { \quad \forall { \bf Q } _ { 1 } , { \bf Q } _ { 2 } \in S O ( 3 ) , \forall { \bf g } _ { 1 } , { \bf g } _ { 2 } \in { \mathbb { R } } ^ { 3 } , \forall { \bf X } _ { 1 } \in { \mathbb { R } } ^ { 3 \times n } , \forall { \bf X } _ { 2 } \in { \mathbb { R } } ^ { 3 \times m } . \quad } \end{array}
|
| 92 |
+
$$
|
| 93 |
+
|
| 94 |
+
Commutativity. Instead of docking $\mathbf { X } _ { 1 }$ with respect to $\mathbf { X } _ { 2 }$ , we can also dock $\mathbf { X } _ { 2 }$ with respect to $\mathbf { X } _ { 1 }$ . In this case, we require the final complex structures to be identical after superimposition, i.e., zero RMSD. This property is named commutativity and it is satisfied as follows (proof in Appendix B.2).
|
| 95 |
+
|
| 96 |
+
Proposition 3. Commutativity as defined by Eq. (1) is guaranteed iff
|
| 97 |
+
|
| 98 |
+
$$
|
| 99 |
+
\mathbf { R } ( \mathbf { X } _ { 2 } | \mathbf { X } _ { 1 } ) = \mathbf { R } ^ { \top } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) ; \quad \mathbf { t } ( \mathbf { X } _ { 2 } | \mathbf { X } _ { 1 } ) = - \mathbf { R } ^ { \top } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) ,
|
| 100 |
+
$$
|
| 101 |
+
|
| 102 |
+
Point Permutation Invariance. We also enforce residue permutation invariance. Formally, both $\mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ and $\mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ should not depend on the order or columns of $\mathbf { X } _ { 1 }$ and, resp., of $\mathbf { X } _ { 2 }$ .
|
| 103 |
+
|
| 104 |
+
# 4 EQUIDOCK MODEL
|
| 105 |
+
|
| 106 |
+
Protein Representation. A protein is a sequence of amino acid residues that folds in a 3D structure. Each residue has a general structure with a side-chain specifying its type, allowing us to define a local frame and derive SE(3)-invariant features for any pair of residues —see Appendix A.
|
| 107 |
+
|
| 108 |
+
We represent a protein as a graph $\mathcal { G } = ( \nu , \mathcal { E } )$ , similar to Fout et al. (2017); Townshend et al. (2019); Liu et al. (2020). Each node $i \in \mathcal V$ represents one residue and has 3D coordinates $\mathbf { x } _ { i } \in \mathbb { R } ^ { 3 }$ corresponding to the $\alpha$ -carbon atom’s location. Edges are given by a $\mathbf { k }$ -nearest-neighbor (k-NN) graph using Euclidean distance of the original 3D node coordinates.
|
| 109 |
+
|
| 110 |
+
Overview of Our Approach. Our model is depicted in Fig. 3. We first build $\mathbf { k }$ -NN protein graphs $\mathcal { G } _ { 1 _ { - } } = ( \nu _ { 1 } , \mathcal { E } _ { 1 } )$ and $\bar { \mathcal { G } _ { 2 } } ^ { - } \bar { = } \left( \mathcal { V } _ { 2 } , \mathcal { E } _ { 2 } \right)$ . We then design SE(3)-invariant node features $\mathbf { F } _ { 1 } \dot { \in } \mathbb { R } ^ { d \times n } , \mathbf { F } _ { 2 } \in$ $\mathbb { R } ^ { d \times m }$ and edge features $\{ \mathbf { f } _ { j i } : \forall ( i , j ) \in \mathcal { E } _ { 1 } \cup \mathcal { E } _ { 2 } \}$ (see Appendix A).
|
| 111 |
+
|
| 112 |
+
Next, we apply several layers consisting of functions $\Phi$ that jointly transform node coordinates and features. Crucially, we guarantee, by design, pairwise independent $S E ( 3 )$ -equivariance for coordinate embeddings, and invariance for feature embeddings. This double constraint is formally defined:
|
| 113 |
+
|
| 114 |
+
Given $\mathbf { Z } _ { 1 } , \mathbf { H } _ { 1 } , \mathbf { Z } _ { 2 } , \mathbf { H } _ { 2 } = \Phi ( \mathbf { X } _ { 1 } , \mathbf { F } _ { 1 } , \mathbf { X } _ { 2 } , \mathbf { F } _ { 2 } )$
|
| 115 |
+
|
| 116 |
+
$$
|
| 117 |
+
\begin{array} { r l } & { \mathrm { 1 , } \mathbf { 1 } _ { 1 } \mathrm { , } \mathbf { 1 } _ { 2 } \mathrm { , } \mathbf { , n } _ { 2 } = \Psi \mathrm { ( A _ { 1 } , } \mathbf { r } _ { 1 } \mathrm { , } \mathbf { r } _ { 2 } \mathrm { , } \mathbf { r } _ { 2 } \mathrm { ) } } \\ & { \mathbf { Q } _ { 1 } \mathbf { Z } _ { 1 } + \mathbf { g } _ { 1 } \mathrm { , } \mathbf { H } _ { 1 } \mathrm { , } \mathbf { Q } _ { 2 } \mathbf { Z } _ { 2 } + \mathbf { g } _ { 2 } \mathrm { , } \mathbf { H } _ { 2 } = \Phi \big ( \mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 } \mathrm { , } \mathbf { F } _ { 1 } \mathrm { , } \mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 } \mathrm { , } \mathbf { F } _ { 2 } \big ) , } \end{array}
|
| 118 |
+
$$
|
| 119 |
+
|
| 120 |
+
$$
|
| 121 |
+
\forall \mathbf { Q } _ { 1 } , \mathbf { Q } _ { 2 } \in S O ( 3 ) , \forall \mathbf { g } _ { 1 } , \mathbf { g } _ { 2 } \in \mathbb { R } ^ { 3 } .
|
| 122 |
+
$$
|
| 123 |
+
|
| 124 |
+
We implement $\Phi$ as a novel type of message-passing neural network (MPNN). We then use the output node coordinate and feature embeddings to compute ${ \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } )$ and $\mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ . These functions depend on pairwise interactions between the two proteins modeled as cross-messages, but also incorporate the 3D structure in a pairwise-independent SE(3)-equivariant way to satisfy Eq. (1), Proposition 1 and Proposition 3. We discover keypoints from each protein based on a neural attention mechanism and softly guide them to represent the respective binding pocket locations via an optimal transport based auxiliary loss. Finally, we obtain the SE(3) transformation by superimposing the two keypoint sets via a differentiable version of the Kabsch algorithm. An additional soft-constraint discourages point cloud intersections. We now detail each of these model components.
|
| 125 |
+
|
| 126 |
+
Independent E(3)-Equivariant Graph Matching Networks (IEGMNs). Our architecture for $\Phi$ satisfying Eq. (4) is called Independent $E ( 3 )$ -Equivariant Graph Matching Network (IEGMN) – see Fig. 3. It extends both Graph Matching Networks (GMN) (Li et al., 2019) and E(3)-Equivariant Graph Neural Networks (E(3)-GNN) (Satorras et al., 2021). IEGMNs perform node coordinate and feature embedding updates for an input pair of graphs $\mathcal { G } _ { 1 } = ( \nu _ { 1 } , \mathcal { E } _ { 1 } )$ , $\mathcal { G } _ { 2 } = ( \nu _ { 2 } , \mathcal { E } _ { 2 } )$ , and use inter- and intranode messages, as well as E(3)-equivariant coordinate updates. The $l .$ -th layer of IEGMNs transforms node latent/feature embeddings $\{ \mathbf { h } _ { i } ^ { ( l ) } \} _ { i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } }$ and node coordinate embeddings $\{ \mathbf { x } _ { i } ^ { ( l ) } \} _ { i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } }$ as
|
| 127 |
+
|
| 128 |
+
$$
|
| 129 |
+
\begin{array} { r l } & { \mathbf { m } _ { j i } = \varphi ^ { c } ( \mathbf { h } _ { i } ^ { ( l ) } , \mathbf { h } _ { j } ^ { ( l ) } , \exp ( - \vert \mathbf { x } _ { i } ^ { ( l ) } - \mathbf { x } _ { j } ^ { ( l ) } \vert ^ { 2 } / \sigma ) , \mathbf { f } _ { j i } ) , \forall e _ { j i } \in \mathcal { E } _ { 1 } \cup \mathcal { E } _ { 2 } } \\ & { \mu _ { j i } = a _ { j i } \} \mathbf { W } \mathbf { h } _ { j } ^ { ( l ) } , \forall i \in \mathcal { V } _ { 1 } , j \in \mathcal { V } _ { 2 } \mathrm { o r } i \in \mathcal { V } _ { 2 , j } \mathrm { ~ } \forall \mathcal { E } _ { 1 } } \\ & { \mathbf { m } _ { i } = \displaystyle \frac { 1 } { \vert \mathbf { W } ( i ) \vert } \displaystyle \sum _ { j \in \mathcal { N } ( i ) } \mathbf { m } _ { j i } , \forall i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } } \\ & { \mu _ { i } = \displaystyle \sum _ { j \in \mathcal { V } _ { 2 } } \mu _ { j i , \forall i } \in \mathcal { V } _ { 1 } , \quad \mathrm { a n d } \quad \mu _ { i } = \displaystyle \sum _ { j \in \mathcal { V } _ { 1 } } \mu _ { j i , \forall i } , \forall i \in \mathcal { V } _ { 2 } } \\ & { \mathbf { x } _ { i } ^ { ( l + 1 ) } = \eta \mathbf { x } _ { i } ^ { ( 0 ) } + ( 1 - \eta ) \mathbf { x } _ { i } ^ { ( l ) } + \displaystyle \sum _ { j \in \mathcal { N } ( i ) } \big ( \mathbf { x } _ { i } ^ { ( l ) } - \mathbf { x } _ { j } ^ { ( l ) } \big ) \varphi ^ { x } ( \mathbf { m } _ { j i } ) , \forall i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } } \\ & { \mathbf { h } _ { i } ^ { ( l + 1 ) } = ( 1 - \beta ) \cdot \mathbf { h } _ { i } ^ { ( l ) } + \beta \cdot \nabla ^ { h } \big ( \mathbf { h } _ { i } ^ { ( l ) } , \mathbf { m } _ { i } , \mu _ { i } \big ) , \forall i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } , } \end{array}
|
| 130 |
+
$$
|
| 131 |
+
|
| 132 |
+
where $\mathcal { N } ( i )$ are the neighbors of node $i$ ; $\varphi ^ { x }$ is a real-valued (scalar) parametric function; W is a learnable matrix; $\varphi ^ { h } , \varphi ^ { e }$ are parametric functions (MLPs) outputting a vector $\mathbb { R } ^ { d }$ ; $\mathbf { f } _ { j i }$ and $\mathbf { f } _ { i }$ are the original edge and node features (extracted SE(3)-invariantly from the residues). $a _ { j \to i }$ is an attention based coefficient with trainable shallow neural networks $\psi ^ { q }$ and $\psi ^ { k }$ :
|
| 133 |
+
|
| 134 |
+
$$
|
| 135 |
+
a _ { j i } = \frac { \exp ( \langle \psi ^ { q } ( \mathbf { h } _ { i } ^ { ( l ) } ) , \psi ^ { k } ( \mathbf { h } _ { j } ^ { ( l ) } ) \rangle ) } { \sum _ { j ^ { \prime } } \exp ( \langle \psi ^ { q } ( \mathbf { h } _ { i } ^ { ( l ) } ) , \psi ^ { k } ( \mathbf { h } _ { j ^ { \prime } } ^ { ( l ) } ) \rangle ) } ,
|
| 136 |
+
$$
|
| 137 |
+
|
| 138 |
+
Note that all parameters of $\mathbf { W } , \varphi ^ { x } , \varphi ^ { h } , \varphi ^ { e } , \psi ^ { q } , \psi ^ { k }$ can be shared or different for different IEGMN layers . The output of several IEGMN layers is then denoted as:
|
| 139 |
+
|
| 140 |
+
$$
|
| 141 |
+
\mathbf { Z } _ { 1 } \in \mathbb { R } ^ { 3 \times n } , \mathbf { H } _ { 1 } \in \mathbb { R } ^ { d \times n } , \mathbf { Z } _ { 2 } \in \mathbb { R } ^ { 3 \times m } , \mathbf { H } _ { 2 } \in \mathbb { R } ^ { d \times m } = I E G M N ( \mathbf { X } _ { 1 } , \mathbf { F } _ { 1 } , \mathbf { X } _ { 2 } , \mathbf { F } _ { 2 } ) .
|
| 142 |
+
$$
|
| 143 |
+
|
| 144 |
+
It is then straightforward to prove the following (see Appendix B.3):
|
| 145 |
+
|
| 146 |
+
Proposition 4. IEGMNs satisfy the pairwise independent $S E ( 3 )$ -equivariance property in Eq. (4).
|
| 147 |
+
|
| 148 |
+
Keypoints for Differentiable Protein Superimposition. Next, we use multi-head attention to obtain $K$ points for each protein, $\mathbf { Y } _ { 1 } , \mathbf { Y } _ { 2 } \mathbf { \bar { \Pi } } \in \mathbb { R } ^ { 3 \times K }$ , which we name keypoints. We train them to become representative points for the binding pocket of the respective protein pair (softly-enforced by an additional loss described later). If this would holds perfectly, then the superimposition of $\mathbf { Y } _ { 1 }$ and $\mathbf { Y } _ { 2 }$ would give the corresponding ground truth superimposition of $\mathbf { X } _ { 1 }$ and $\mathbf { X } _ { 2 }$ . Our model is :
|
| 149 |
+
|
| 150 |
+
$$
|
| 151 |
+
\mathbf { y } _ { 1 k } : = \sum _ { i = 1 } ^ { n } \alpha _ { i } ^ { k } \mathbf { z } _ { 1 i } ; \quad \mathbf { y } _ { 2 k } : = \sum _ { j = 1 } ^ { m } \beta _ { j } ^ { k } \mathbf { z } _ { 2 j } ,
|
| 152 |
+
$$
|
| 153 |
+
|
| 154 |
+
where $\mathbf { z } _ { 1 i }$ denotes the i-th column of matrix $\mathbf { Z } _ { 1 }$ , and $\alpha _ { i } ^ { k } = s o f t m a x _ { i } ( \textstyle \frac { 1 } { \sqrt { d } } \mathbf { h } _ { 1 i } ^ { \top } \mathbf { W } _ { k } ^ { \prime } \mu ( \varphi ( \mathbf { H } _ { 2 } ) ) )$ are attention scores (similarly defined for $\beta _ { j } ^ { k \cdot }$ ), with $\mathbf { W } _ { k } ^ { \prime } \in \mathbb { R } ^ { d \times d }$ a parametric matrix (different for each attention head), $\varphi$ a linear layer plus a LeakyReLU non-linearity, and $\mu ( \cdot )$ is the mean vector.
|
| 155 |
+
|
| 156 |
+
Differentiable Kabsch Model. We design the rotation and translation that docks protein 1 into protein 2 to be the same transformation used to superimpose ${ \bf Y } _ { 1 }$ and $\mathbf { Y } _ { 2 }$ — see Fig. 3. For this, we compute a differentiable version of the Kabsch algorithm (Kabsch, 1976) as follows. Let $\mathbf { A } = \overline { { \mathbf { Y } } } _ { 2 } \overline { { \mathbf { Y } } } _ { 1 } ^ { \top } \in \mathbb { R } ^ { 3 \times 3 } .$ computed using zero-mean keypoints. The singular value decomposition (SVD) is $\mathbf { A } { \dot { \mathbf { \eta } } } = \mathbf { U } _ { 2 } \mathbf { S } \mathbf { U } _ { 1 } ^ { \top }$ , where $\mathbf { U } _ { 2 } , \mathbf { U } _ { 1 } ^ { - } \in O ( 3 )$ . Finally, we define the differentiable functions
|
| 157 |
+
|
| 158 |
+
$$
|
| 159 |
+
\begin{array} { r l } & { \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ; \theta ) = \mathbf { U } _ { 2 } \left( \begingroup _ { 0 } ^ { 1 } \ \right. \left( \begin{array} { l l l } { 0 } & { 0 } \\ { 0 } & { 1 } & { 0 } \\ { 0 } & { 0 } & { d } \end{array} \right) \mathbf { U } _ { 1 } ^ { \top } , \quad \mathrm { w h e r e \ } d = \mathrm { s i g n } ( \operatorname* { d e t } ( \mathbf { U } _ { 2 } \mathbf { U } _ { 1 } ^ { \top } ) ) } \\ & { \left. \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ; \theta ) = \mu ( \mathbf { Y } _ { 2 } ) - \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ; \theta ) \mu ( \mathbf { Y } _ { 1 } ) , \right. } \end{array}
|
| 160 |
+
$$
|
| 161 |
+
|
| 162 |
+
where $\mu ( \cdot )$ is the mean vector of a point cloud. It is straightforward to prove that this model satisfies all the equivariance properties in Eqs. (1) to (3). From a practical perspective, the gradient and backpropagation through the SVD operation was analyzed by (Ionescu et al., 2015; Papadopoulo and Lourakis, 2000) and implemented in the automatic differentiation frameworks such as PyTorch.
|
| 163 |
+
|
| 164 |
+
MSE Loss. During training, we randomly decide which protein is the receptor (say protein 2), keep it in the docked position (i.e., ${ \bf X } _ { 2 } = { \bf X } _ { 2 } ^ { * }$ ), predict the SE(3) transformation using Eq. (13) and use it to compute the final position of the ligand as $\tilde { \mathbf { X } } _ { 1 } = \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) \mathbf { X } _ { 1 } + \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } )$ . The mean squared error (MSE) loss is then $\begin{array} { r } { \mathcal { L } _ { \mathrm { M S E } } = \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \| \mathbf { x } _ { i } ^ { * } - \tilde { \mathbf { x } } _ { i } \| ^ { 2 } } \end{array}$ .
|
| 165 |
+
|
| 166 |
+
Optimal Transport and Binding Pocket Keypoint Alignment. As stated before, we desire that $\mathbf { Y } _ { 1 }$ and $\mathbf { Y } _ { 2 }$ are representative points for the binding pocket location of the respective protein pair. However, this needs to be encouraged explicitly, which we achieve using an additional loss.
|
| 167 |
+
|
| 168 |
+
We first define the binding pocket point sets, inspiring from previous PPI work (Section 2). Given the residues’ $\alpha$ -carbon locations of the bound (docked) structures, $\mathbf { X } _ { 1 } ^ { * }$ and $\mathbf { X } _ { 2 } ^ { * }$ , we select all pairs of residues at less than $\tau$ Euclidean distance $( \tau = 8 \mathring \mathbf { A }$ in our experiments). We assume these are all interacting residues. Denote these pairs as $\{ ( \mathbf { x } _ { 1 s } ^ { * } , \mathbf { x } _ { 2 s } ^ { * } ) , s \in { 1 , . . . , S } \}$ , where $S$ is variable across data pairs. We compute midpoints of these segments, denoted as $\mathbf { P } _ { 1 } ^ { \ast } , \mathbf { P } _ { 2 } ^ { \ast } \in \mathbb { R } ^ { 3 \times S }$ , where $\mathbf { p } _ { 1 s } ^ { * } = \mathbf { p } _ { 2 s } ^ { * } = 0 . 5 \cdot ( \mathbf { x } _ { 1 s } ^ { * } + \mathbf { x } _ { 2 s } ^ { * } )$ . We view $\mathbf { P } _ { 1 } ^ { * }$ and $\mathbf { P } _ { 2 } ^ { * }$ as binding pocket points. In the unbound state, these sets are randomly moved in space together with the respective protein residue coordinates $\mathbf { X } _ { 1 }$ and $\mathbf { X } _ { 2 }$ . We denote them as $\mathbf { P } _ { 1 } , \mathbf { P } _ { 2 } ^ { \bullet } \in \mathbb { R } ^ { 3 \times S }$ . For clarity, if $\mathbf { X } _ { 1 } = \mathbf { Q } \mathbf { X } _ { 1 } ^ { * } + \mathbf { g }$ , then $\mathbf { P } _ { 1 } = \mathbf { Q } \mathbf { P } _ { 1 } ^ { * } + \mathbf { g }$ .
|
| 169 |
+
|
| 170 |
+
We desire that $\mathbf { Y } _ { 1 }$ is a representative set for the 3D set ${ \bf P } _ { 1 }$ (and, similarly, $\mathbf { Y } _ { 2 }$ for $\mathbf { P } _ { 2 }$ ). However, while at training time we know that every point $\mathbf { p } _ { 1 s }$ corresponds to the point $\mathbf { p } _ { 2 s }$ (and, similarly, $\mathbf { y } _ { 1 k }$ aligns with $\mathbf { y } _ { 2 k }$ , by assumption), we unfortunately do not know the actual alignment between points in $\mathbf { Y } _ { l }$ and $\mathbf { P } _ { l }$ , for every $l \in \{ 1 , 2 \}$ . This can be recovered using an additional optimal transport loss:
|
| 171 |
+
|
| 172 |
+
$$
|
| 173 |
+
{ \mathcal { L } } _ { \mathrm { O T } } = \operatorname* { m i n } _ { \mathbf { T } \in \mathcal { U } ( S , K ) } \langle \mathbf { T } , \mathbf { C } \rangle , \quad \mathrm { w h e r e ~ } \mathbf { C } _ { s , k } = \| \mathbf { y } _ { 1 k } - \mathbf { p } _ { 1 s } \| ^ { 2 } + \| \mathbf { y } _ { 2 k } - \mathbf { p } _ { 2 s } \| ^ { 2 } ,
|
| 174 |
+
$$
|
| 175 |
+
|
| 176 |
+
where $\mathcal { U } ( S , K )$ is the set of $S \times K$ transport plans with uniform marginals. The optimal transport plan is computed using an Earth Mover’s Distance and the POT library (Flamary et al., 2021), while being kept fixed during back-propagation and optimization when only the cost matrix is trained.
|
| 177 |
+
|
| 178 |
+
Note that our approach assumes that $\mathbf { y } _ { 1 k }$ corresponds to $\mathbf { y } _ { 2 k }$ , for every $k \in \{ 1 , \ldots , K \}$ . Intuitively, each attention head $k$ will identify a specific geometric/chemical local surface feature of protein 1 by $\mathbf { y } _ { 1 k }$ , and match its complementary feature of protein 2 by $\mathbf { y } _ { 2 k }$ .
|
| 179 |
+
|
| 180 |
+
Avoiding Point Cloud Intersection. In practice, our model does not enforce a useful inductive bias, namely that proteins forming complexes are never "intersecting" with each other. To address this issue, we first state a notion of the "interior" of a protein point cloud. Following previous work cloud $\mathbf { X } \in \mathbb { R } ^ { 3 \times n }$ et as $\{ \mathbf { x } \in \mathbb { R } ^ { 3 } : G ( \mathbf { x } ) = \gamma \}$ et al., 2, where $\begin{array} { r } { G ( \mathbf { x } ) = - \sigma \ln ( \sum _ { i = 1 } ^ { n } \exp ( - | | \mathbf { \hat { x } } - \mathbf { x } _ { i } | | ^ { 2 } / \sigma ) ) } \end{array}$ The parameters $\sigma$ and $\gamma$ are chosen such that there exist no "holes" inside a protein (we found $\gamma = 1 0 , \sigma = 2 5$ to work well, see Appendix E). As a consequence, the interior of the protein is given by $\{ \mathbf { x } \in \mathbb { R } ^ { 3 } : G ( \mathbf { x } ) < \gamma \}$ . Then, the condition for non-intersecting ligand and receptor can be written as $G _ { 1 } ( \mathbf { x } _ { 2 j } ) > \gamma , \forall j \in { 1 , \ldots , m }$ and $G _ { 2 } ( \mathbf { x } _ { 1 i } ) > \gamma , \forall i \in { 1 , . . . , n }$ . As a loss function, this becomes
|
| 181 |
+
|
| 182 |
+
$$
|
| 183 |
+
\mathcal { L } _ { \mathrm { { N I } } } = \frac { 1 } { n } \sum _ { i = 1 } ^ { n } \operatorname* { m a x } ( 0 , \gamma - G _ { 2 } ( \mathbf { x } _ { 1 i } ) ) + \frac { 1 } { m } \sum _ { j = 1 } ^ { m } \operatorname* { m a x } ( 0 , \gamma - G _ { 1 } ( \mathbf { x } _ { 2 j } ) ) .
|
| 184 |
+
$$
|
| 185 |
+
|
| 186 |
+
Surface Aware Node Features. Surface contact modeling is important for protein docking. We here design a novel surface feature type that differentiates residues closer to the surface of the protein from those in the interior. Similar to Sverrisson et al. (2021), we prioritize efficiency and avoid pre-computing meshes, but show that our new feature is a good proxy for residue’s depth (i.e. distance to the protein surface). Intuitively, residues in the core of the protein are locally surrounded in all directions by other residues. This is not true for residues on the surface, e.g., neighbors are in a half-space if the surface is locally flat. Building on this intuition, for each node (residue) $i$ in the $k$ -NN protein graph, we compute the norm of the weighted average of its neighbor forces, which can be interpreted as the normalized gradient of the $G ( \mathbf { x } )$ surface function. This SE(3)-invariant feature is
|
| 187 |
+
|
| 188 |
+
$$
|
| 189 |
+
\rho _ { i } ( \mathbf { x } _ { i } ; \lambda ) = \frac { \big \| \sum _ { i ^ { \prime } \in \mathcal { N } _ { i } } w _ { i , i ^ { \prime } , \lambda } ( \mathbf { x } _ { i } - \mathbf { x } _ { i ^ { \prime } } ) \big \| } { \sum _ { i ^ { \prime } \in \mathcal { N } _ { i } } w _ { i , i ^ { \prime } , \lambda } \| \mathbf { x } _ { i } - \mathbf { x } _ { i ^ { \prime } } \| } , \quad \mathrm { ~ w h e r e ~ } w _ { i , i ^ { \prime } , \lambda } = \frac { \exp ( - | | \mathbf { x } _ { i } - \mathbf { x } _ { i ^ { \prime } } | | ^ { 2 } / \lambda ) } { \sum _ { j \in \mathcal { N } _ { i } } \exp ( - | | \mathbf { x } _ { i } - \mathbf { x } _ { j } | | ^ { 2 } / \lambda ) } .
|
| 190 |
+
$$
|
| 191 |
+
|
| 192 |
+
Intuitively, as depicted in Fig. 8, residues in the interior of the protein have values close to 0 since they are surrounded by vectors from all directions that cancel out, while residues near the surface have neighbors only in a narrower cone, with aperture depending on the local curvature of the surface. We show in Appendix C that this feature correlates well with more expensive residue depth estimation methods, e.g. based on MSMS, thus offering a computationally appealing alternative. We also compute an estimation of this feature for large dense point clouds based on the local surface angle.
|
| 193 |
+
|
| 194 |
+
# 5 EXPERIMENTS
|
| 195 |
+
|
| 196 |
+
Datasets. We leverage the following datasets: Docking Benchmark 5.5 (DB5.5) (Vreven et al., 2015) and Database of Interacting Protein Structures (DIPS) (Townshend et al., 2019). DB5.5 is a gold standard dataset in terms of data quality, but contains only 253 structures. DIPS is a larger protein complex structures dataset mined from the Protein Data Bank (Berman et al., 2000) and tailored for rigid body docking. Datasets information is given in Appendix D. We filter DIPS to only keep proteins with at most 10K atoms. Datasets are then randomly partitioned in train/val/test splits of sizes 203/25/25 (DB5.5) and 39,937/974/965 (DIPS). For DIPS, the split is based on protein family to separate similar proteins. For the final evaluation in Table 1, we use the full DB5.5 test set, and randomly sample 100 pairs from different protein families from the DIPS test set.
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
Figure 4: a. Complex-RMSD distributions (DIPS test set); b. Interface-RMSD distributions (DIPS test set); c. scatter plot for C-RMSD vs I-RMSD (DIPS test set).
|
| 200 |
+
|
| 201 |
+
Baselines. We compare our EQUIDOCK method with popular state-of-the-art docking software 2 CLUSPRO (PIPER) (Desta et al., 2020; Kozakov et al., 2017),ATTRACT (Schindler et al., 2017; de Vries et al., 2015), PATCHDOCK (Mashiach et al., 2010; SchneidmanDuhovny et al., 2005), and HDOCK (Yan et al., 2020; 2017b;a; Huang and Zou, 2014; 2008). These baselines provide user-friendly local packages suitable for automatic experiments or webservers for manual submissions.
|
| 202 |
+
|
| 203 |
+
Evaluation Metrics. To measure prediction’s quality, we report Complex Root Mean Square Deviation (CRMSD) and Interface Root Mean Square Deviation (IRMSD), defined below. Given the ground truth and predicted complex structures, $\mathbf { Z ^ { * } } \in \mathbb { R } ^ { 3 \times ( n + m ) }$ and $\mathbf { Z } \in \mathbb { R } ^ { 3 \times ( n + m ) }$ , we first superimpose them using the Kabsch algorithm (Kabsch, 1976), and then compute $\begin{array} { r } { \mathrm { C - R M S D } = \sqrt { \frac { 1 } { n + m } \| { \bf Z } ^ { * } - { \bf Z } \| _ { F } ^ { 2 } } } \end{array}$ . We compute I-RMSD similarly, but using only the coordinates of the interface residues with distance less than $8 \mathring \mathrm { A }$ to the other protein’s residues. For a fair comparison among baselines, we use only the $\alpha$ -carbon coordinates to compute both metrics.
|
| 204 |
+
|
| 205 |
+
|
| 206 |
+
Figure 5: Inference running time distributions (log10 scale).
|
| 207 |
+
|
| 208 |
+
Training Details. We train our models on the train part of DIPS first, using Adam (Kingma and Ba, 2014) with learning rate 2e-4 and early stopping with patience of 30 epochs. We update the best validation model only when it achieves a score of less than $98 \%$ of the previous best validation score, where the score is the median of Ligand RMSD on the full DIPS validation set. The best DIPS validated model is then tested on the DIPS test set. For DB5.5, we fine tune the DIPS pre-trained model on the DB5.5 training set using learning rate 1e-4 and early stopping with 150 epochs patience. The best DB5.5 validated model is finally tested on DB5.5 test set. During training, we randomly assign the roles of ligand and receptor. Also, during both training and testing, we randomly rotate and translate the ligand in space (even though our model is invariant to this operation) for all baselines.
|
| 209 |
+
|
| 210 |
+
|
| 211 |
+
Figure 6: Visualization of a protein complex successfully predicted by EQUIDOCK. Note that all other methods find the binding interface on the wrong side of the black protein.
|
| 212 |
+
|
| 213 |
+
Complex Prediction Results. Results are shown in Table 1, Fig. 4 and Appendix E. We note that our method is competitive and often outperforms the baselines. However, we do not use heavy candidate sampling and re-ranking, we do not rely on task-specific hand-crafted features, and we currently do not perform structure fine-tuning, aiming to predict the SE(3) ligand transformation in a direct shot. Moreover, we note that some of the baselines might have used part of our test set in validating their models, for example to learn surface templates, thus, their reported scores might be optimistic. Notably, HDOCK score function was validated on DB4 which overlaps with DB5.5. A more appropriate comparison would require us to re-build these baselines without information from our test sets, a task that is currently not possible without open-source implementations.
|
| 214 |
+
|
| 215 |
+
Computational Efficiency. We show inference times in Fig. 5 and Table 4. Note that EQUIDOCK is between 80-500 times faster than the baselines. This is especially important for intensive screening applications that aim to scan over vast search spaces, e.g. for drug discovery. In addition, it is also relevant for de novo design of binding proteins (e.g. antibodies (Jin et al., 2021)) or for use cases when protein docking models are just a component of significantly larger end-to-end architectures targeting more involved biological scenarios, e.g., representing a drug’s mechanism of action or modeling cellular processes with a single model as opposed to a multi-pipeline architecture.
|
| 216 |
+
|
| 217 |
+
Visualization. We show in Fig. 6 a successful example of a test DIPS protein pair for which our model significantly outperforms all baselines.
|
| 218 |
+
|
| 219 |
+
# 6 CONCLUSION
|
| 220 |
+
|
| 221 |
+
We have presented an extremely fast, end-to-end rigid protein docking approach that does not rely on candidate sampling, templates, task-specific features or pre-computed meshes. Our method smartly incorporates useful rigid protein docking priors including commutativity and pairwise independent SE(3)-equivariances, thus avoiding the computational burden of data augmentation.
|
| 222 |
+
|
| 223 |
+
We look forward to incorporating more domain knowledge in EQUIDOCK and extend it for flexible docking and docking molecular dynamics, as well as adapt it to other related tasks such as drug binding prediction. On the long term, we envision that fast and accurate deep learning models would allow us to tackle more complex and involved biological scenarios, for example to model the mechanism of action of various drugs or to design de novo binding proteins and drugs to specific targets (e.g. for antibody generation). Last, we hope that our architecture can inspire the design of other types of biological 3D interactions.
|
| 224 |
+
|
| 225 |
+
Limitations. First, our presented model does not incorporate protein flexibility which is necessary for various protein families, e.g., antibodies. Unfortunately, both DB5 and DIPS datasets are biased towards rigid body docking . Second, we only prevent steric clashes using a soft constraint (Eq. (15)) which has limitations (see Table 6). Future extensions would hard-constrain the model to prevent such artifacts.
|
| 226 |
+
|
| 227 |
+
# ACKNOWLEDGEMENTS
|
| 228 |
+
|
| 229 |
+
The authors thank Hannes Stärk, Gabriele Corso, Patrick Walters, Tian Xie, Xiang Fu, Jacob Stern, Jason Yim, Lewis Martin, Jeremy Wohlwend, Jiaxiang Wu, Wei Liu, and Ding Xue for insightful and helpful discussions. OEG is funded by the Machine Learning for Pharmaceutical Discovery and Synthesis (MLPDS) consortium, the Abdul Latif Jameel Clinic for Machine Learning in Health, the DTRA Discovery of Medical Countermeasures Against New and Emerging (DOMANE) threats program, and the DARPA Accelerated Molecular Discovery program. This publication was created as part of NCCR Catalysis (grant number 180544), a National Centres of Competence in Research funded by the Swiss National Science Foundation. RB and TJ also acknowledge support from NSF Expeditions grant (award 1918839): Collaborative Research: Understanding the World Through Code.
|
| 230 |
+
|
| 231 |
+
# REFERENCES
|
| 232 |
+
|
| 233 |
+
B. Adhikari and J. Cheng. Confold2: Improved contact-driven ab initio protein structure modeling. BMC bioinformatics, 19(1):1–5, 2018. 3
|
| 234 |
+
M. Baek, F. DiMaio, I. Anishchenko, J. Dauparas, S. Ovchinnikov, G. R. Lee, J. Wang, Q. Cong, L. N. Kinch, R. D. Schaeffer, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science, 373(6557):871–876, 2021. 1, 3
|
| 235 |
+
J. Bao, X. He, and J. Z. Zhang. Deepbsp—a machine learning method for accurate prediction of protein–ligand docking structures. Journal of Chemical Information and Modeling, 2021. 3
|
| 236 |
+
S. Basu and B. Wallner. Dockq: a quality measure for protein-protein docking models. PloS one, 11 (8):e0161879, 2016. 3
|
| 237 |
+
H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne. The protein data bank. Nucleic acids research, 28(1):235–242, 2000. 7
|
| 238 |
+
J. Biesiada, A. Porollo, P. Velayutham, M. Kouril, and J. Meller. Survey of public domain software for docking simulations and virtual screening. Human genomics, 5(5):1–9, 2011. 3
|
| 239 |
+
J. Bruna, W. Zaremba, A. Szlam, and Y. LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. 2
|
| 240 |
+
P. Bryant, G. Pozzati, and A. Elofsson. Improved prediction of protein-protein interactions using alphafold2 and extended multiple-sequence alignments. bioRxiv, 2021. 3
|
| 241 |
+
R. Chen, L. Li, and Z. Weng. Zdock: an initial-stage protein-docking algorithm. Proteins: Structure, Function, and Bioinformatics, 52(1):80–87, 2003. 1, 3
|
| 242 |
+
C. Christoffer, S. Chen, V. Bharadwaj, T. Aderinwale, V. Kumar, M. Hormati, and D. Kihara. Lzerd webserver for pairwise and multiple protein–protein docking. Nucleic Acids Research, 2021. 3
|
| 243 |
+
T. Cohen and M. Welling. Group equivariant convolutional networks. In International conference on machine learning, pages 2990–2999. PMLR, 2016a. 2
|
| 244 |
+
T. S. Cohen and M. Welling. Steerable cnns. arXiv preprint arXiv:1612.08498, 2016b. 2
|
| 245 |
+
B. Dai and C. Bailey-Kellogg. Protein interaction interface region prediction by geometric deep learning. Bioinformatics, 2021. 3
|
| 246 |
+
S. J. De Vries, M. Van Dijk, and A. M. Bonvin. The haddock web server for data-driven biomolecular docking. Nature protocols, 5(5):883–897, 2010. 1, 3
|
| 247 |
+
S. J. de Vries, C. E. Schindler, I. C. de Beauchêne, and M. Zacharias. A web interface for easy flexible protein-protein docking with attract. Biophysical journal, 108(3):462–465, 2015. 8
|
| 248 |
+
M. Defferrard, X. Bresson, and P. Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. arXiv preprint arXiv:1606.09375, 2016. 2
|
| 249 |
+
|
| 250 |
+
I. T. Desta, K. A. Porter, B. Xia, D. Kozakov, and S. Vajda. Performance and its limits in rigid body protein-protein docking. Structure, 28(9):1071–1081, 2020. 8
|
| 251 |
+
|
| 252 |
+
S. Eismann, R. J. Townshend, N. Thomas, M. Jagota, B. Jing, and R. O. Dror. Hierarchical, rotationequivariant neural networks to select structural models of protein complexes. Proteins: Structure, Function, and Bioinformatics, 2020. 2, 3
|
| 253 |
+
|
| 254 |
+
R. Evans, M. O’Neill, A. Pritzel, N. Antropova, A. W. Senior, T. Green, A. Žídek, R. Bates, S. Blackwell, J. Yim, O. Ronneberger, S. Bodenstein, M. Zielinski, A. Bridgland, A. Potapenko, A. Cowie, K. Tunyasuvunakool, R. Jain, E. Clancy, P. Kohli, J. Jumper, and D. Hassabis. Protein complex prediction with alphafold-multimer. bioRxiv, 2021. doi: 10.1101/2021.10.04.463034. 3
|
| 255 |
+
|
| 256 |
+
M. Finzi, S. Stanton, P. Izmailov, and A. G. Wilson. Generalizing convolutional neural networks for equivariance to lie groups on arbitrary continuous data. In International Conference on Machine Learning, pages 3165–3176. PMLR, 2020. 2
|
| 257 |
+
|
| 258 |
+
R. Flamary, N. Courty, A. Gramfort, M. Z. Alaya, A. Boisbunon, S. Chambon, L. Chapel, A. Corenflos, K. Fatras, N. Fournier, L. Gautheron, N. T. Gayraud, H. Janati, A. Rakotomamonjy, I. Redko, A. Rolet, A. Schutz, V. Seguy, D. J. Sutherland, R. Tavenard, A. Tong, and T. Vayer. Pot: Python optimal transport. Journal of Machine Learning Research, 22(78):1–8, 2021. URL http://jmlr.org/papers/v22/20-451.html. 7
|
| 259 |
+
|
| 260 |
+
A. Fout, J. Byrd, B. Shariat, and A. Ben-Hur. Protein interface prediction using graph convolutional networks. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6533–6542, 2017. 3, 5
|
| 261 |
+
|
| 262 |
+
F. B. Fuchs, D. E. Worrall, V. Fischer, and M. Welling. Se (3)-transformers: 3d roto-translation equivariant attention networks. arXiv preprint arXiv:2006.10503, 2020. 2
|
| 263 |
+
|
| 264 |
+
O.-E. Ganea, L. Pattanaik, C. W. Coley, R. Barzilay, K. F. Jensen, W. H. Green, and T. S. Jaakkola. Geomol: Torsional geometric generation of molecular 3d conformer ensembles. arXiv preprint arXiv:2106.07802, 2021. 3
|
| 265 |
+
|
| 266 |
+
U. Ghani, I. Desta, A. Jindal, O. Khan, G. Jones, S. Kotelnikov, D. Padhorny, S. Vajda, and D. Kozakov. Improved docking of protein models by a combination of alphafold2 and cluspro. bioRxiv, 2021. 3
|
| 267 |
+
|
| 268 |
+
J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl. Neural message passing for quantum chemistry. In International Conference on Machine Learning, pages 1263–1272. PMLR, 2017. 2
|
| 269 |
+
|
| 270 |
+
S.-Y. Huang and X. Zou. An iterative knowledge-based scoring function for protein–protein recognition. Proteins: Structure, Function, and Bioinformatics, 72(2):557–579, 2008. 8
|
| 271 |
+
|
| 272 |
+
S.-Y. Huang and X. Zou. A knowledge-based scoring function for protein-rna interactions derived from a statistical mechanics-based iterative method. Nucleic acids research, 42(7):e55–e55, 2014. 8
|
| 273 |
+
|
| 274 |
+
I. R. Humphreys, J. Pei, M. Baek, A. Krishnakumar, I. Anishchenko, S. R. Ovchinnikov, J. Zheng, T. Ness, S. Banjade, S. R. Bagde, V. Stancheva, X. Li, K. Liu, Z. Zheng, D. Barerro, U. Roy, I. S. Fernandez, B. Szakal, D. Branzei, E. C. Greene, S. Biggins, S. Keeney, E. A. Miller, J. C. Fromme, T. Hendrickson, Q. Cong, and D. Baker. Structures of core eukaryotic protein complexes. 2021. doi: 10.1101/2021.09.30.462231. 3
|
| 275 |
+
|
| 276 |
+
M. Hutchinson, C. L. Lan, S. Zaidi, E. Dupont, Y. W. Teh, and H. Kim. Lietransformer: Equivariant self-attention for lie groups. arXiv Preprint, 2012.10885, 2020. 2
|
| 277 |
+
|
| 278 |
+
J. Ingraham, V. K. Garg, R. Barzilay, and T. Jaakkola. Generative models for graph-based protein design. 2019. 15
|
| 279 |
+
|
| 280 |
+
C. Ionescu, O. Vantzos, and C. Sminchisescu. Matrix backpropagation for deep networks with structured layers. In Proceedings of the IEEE International Conference on Computer Vision, pages 2965–2973, 2015. 6
|
| 281 |
+
|
| 282 |
+
W. Jin, J. Wohlwend, R. Barzilay, and T. Jaakkola. Iterative refinement graph neural network for antibody sequence-structure co-design. arXiv preprint arXiv:2110.04624, 2021. 9
|
| 283 |
+
L. Jovine. Using machine learning to study protein–protein interactions: From the uromodulin polymer to egg zona pellucida filaments. Molecular Reproduction and Development, $\mathrm { { n } ( \mathrm { { n } ( \mathrm { { n } / \mathrm { { a } ) } } } }$ . doi: https://doi.org/10.1002/mrd.23538. URL https://onlinelibrary.wiley.com/doi/ abs/10.1002/mrd.23538. 3
|
| 284 |
+
F. Ju, J. Zhu, B. Shao, L. Kong, T.-Y. Liu, W.-M. Zheng, and D. Bu. Copulanet: Learning residue co-evolution directly from multiple sequence alignment for protein structure prediction. Nature Communications, 12(1):2535, May 2021. 3
|
| 285 |
+
J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Žídek, A. Potapenko, et al. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583–589, 2021. 1, 3, 15
|
| 286 |
+
W. Kabsch. A solution for the best rotation to relate two sets of vectors. Acta Crystallographica Section A: Crystal Physics, Diffraction, Theoretical and General Crystallography, 32(5):922–923, 1976. 4, 6, 8
|
| 287 |
+
D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 8
|
| 288 |
+
T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2
|
| 289 |
+
D. R. Koes, M. P. Baumgartner, and C. J. Camacho. Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise. Journal of chemical information and modeling, 53(8): 1893–1904, 2013. 3
|
| 290 |
+
D. Kozakov, D. R. Hall, B. Xia, K. A. Porter, D. Padhorny, C. Yueh, D. Beglov, and S. Vajda. The cluspro web server for protein–protein docking. Nature protocols, 12(2):255–278, 2017. 3, 8
|
| 291 |
+
G. Launay, M. Ohue, J. Prieto Santero, Y. Matsuzaki, C. Hilpert, N. Uchikoga, T. Hayashi, and J. Martin. Evaluation of consrank-like scoring functions for rescoring ensembles of protein–protein docking poses. Frontiers in molecular biosciences, 7:308, 2020. 3
|
| 292 |
+
S. Li, J. Zhou, T. Xu, L. Huang, F. Wang, H. Xiong, W. Huang, D. Dou, and H. Xiong. Structureaware interactive graph neural networks for the prediction of protein-ligand binding affinity. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 975–985, 2021. 3
|
| 293 |
+
Y. Li, C. Gu, T. Dullien, O. Vinyals, and P. Kohli. Graph matching networks for learning the similarity of graph structured objects. In International Conference on Machine Learning, pages 3835–3845. PMLR, 2019. 2, 5
|
| 294 |
+
Y. Liu, H. Yuan, L. Cai, and S. Ji. Deep learning of high-order interactions for protein interface prediction. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 679–687, 2020. 3, 5
|
| 295 |
+
E. Mashiach, D. Schneidman-Duhovny, A. Peri, Y. Shavit, R. Nussinov, and H. J. Wolfson. An integrated suite of fast docking algorithms. Proteins: Structure, Function, and Bioinformatics, 78 (15):3197–3204, 2010. 8
|
| 296 |
+
A. T. McNutt, P. Francoeur, R. Aggarwal, T. Masuda, R. Meli, M. Ragoza, J. Sunseri, and D. R. Koes. Gnina 1.0: molecular docking with deep learning. Journal of cheminformatics, 13(1):1–20, 2021. 3
|
| 297 |
+
I. H. Moal, M. Torchala, P. A. Bates, and J. Fernández-Recio. The scoring of poses in protein-protein docking: current capabilities and future directions. BMC bioinformatics, 14(1):1–15, 2013. 3
|
| 298 |
+
T. Papadopoulo and M. I. Lourakis. Estimating the jacobian of the singular value decomposition: Theory and applications. In European Conference on Computer Vision, pages 554–570. Springer, 2000. 6
|
| 299 |
+
J. Pei, J. Zhang, and Q. Cong. Human mitochondrial protein complexes revealed by large-scale coevolution analysis and deep learning-based structure modeling. bioRxiv, 2021. 3
|
| 300 |
+
D. J. Rezende, S. Racanière, I. Higgins, and P. Toth. Equivariant hamiltonian flows. arXiv preprint arXiv:1909.13739, 2019. 2
|
| 301 |
+
M. F. Sanner, A. J. Olson, and J.-C. Spehner. Reduced surface: an efficient way to compute molecular surfaces. Biopolymers, 38(3):305–320, 1996. 17
|
| 302 |
+
V. G. Satorras, E. Hoogeboom, and M. Welling. E(n)-equivariant graph neural networks. arXiv preprint arXiv:2102.09844, 2021. 2, 5
|
| 303 |
+
C. E. Schindler, I. Chauvot de Beauchêne, S. J. de Vries, and M. Zacharias. Protein-protein and peptide-protein docking and refinement using attract in capri. Proteins: Structure, Function, and Bioinformatics, 85(3):391–398, 2017. 1, 3, 8
|
| 304 |
+
D. Schneidman-Duhovny, Y. Inbar, R. Nussinov, and H. J. Wolfson. Patchdock and symmdock: servers for rigid and symmetric docking. Nucleic acids research, 33(suppl_2):W363–W367, 2005. 8
|
| 305 |
+
A. W. Senior, R. Evans, J. Jumper, J. Kirkpatrick, L. Sifre, T. Green, C. Qin, A. Žídek, A. W. Nelson, A. Bridgland, et al. Improved protein structure prediction using potentials from deep learning. Nature, 577(7792):706–710, 2020. 1, 3
|
| 306 |
+
S. Sunny and P. Jayaraj. Fpdock: Protein–protein docking using flower pollination algorithm. Computational Biology and Chemistry, 93:107518, 2021. 1, 3
|
| 307 |
+
F. Sverrisson, J. Feydy, B. E. Correia, and M. M. Bronstein. Fast end-to-end learning on protein surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15272–15281, 2021. 7
|
| 308 |
+
N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. arXiv preprint arXiv:1802.08219, 2018. 2, 3
|
| 309 |
+
M. Torchala, I. H. Moal, R. A. Chaleil, J. Fernandez-Recio, and P. A. Bates. Swarmdock: a server for flexible protein–protein docking. Bioinformatics, 29(6):807–809, 2013. 1, 3
|
| 310 |
+
R. Townshend, R. Bedi, P. Suriana, and R. Dror. End-to-end learning on 3d protein structure for interface prediction. Advances in Neural Information Processing Systems, 32:15642–15651, 2019. 3, 5, 7
|
| 311 |
+
O. Trott and A. J. Olson. Autodock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry, 31(2):455–461, 2010. 3
|
| 312 |
+
I. A. Vakser. Protein-protein docking: From interaction to interactome. Biophysical journal, 107(8): 1785–1793, 2014. 3
|
| 313 |
+
V. Venkatraman, Y. D. Yang, L. Sael, and D. Kihara. Protein-protein docking using region-based 3d zernike descriptors. BMC bioinformatics, 10(1):1–21, 2009. 1, 3, 7
|
| 314 |
+
J. Verburgt and D. Kihara. Benchmarking of structure refinement methods for protein complex models. Proteins: Structure, Function, and Bioinformatics, 2021. 3
|
| 315 |
+
T. Vreven, I. H. Moal, A. Vangone, B. G. Pierce, P. L. Kastritis, M. Torchala, R. Chaleil, B. JiménezGarcía, P. A. Bates, J. Fernandez-Recio, et al. Updates to the integrated protein–protein interaction benchmarks: docking benchmark version 5 and affinity benchmark version 2. Journal of molecular biology, 427(19):3031–3041, 2015. 7
|
| 316 |
+
I. Wallach, M. Dzamba, and A. Heifets. Atomnet: a deep convolutional neural network for bioactivity prediction in structure-based drug discovery. arXiv preprint arXiv:1510.02855, 2015. 3
|
| 317 |
+
M. Weiler and G. Cesa. General $e ( 2 )$ -equivariant steerable cnns. arXiv preprint arXiv:1911.08251, 2019. 2
|
| 318 |
+
|
| 319 |
+
G. Weng, E. Wang, Z. Wang, H. Liu, F. Zhu, D. Li, and T. Hou. Hawkdock: a web server to predict and analyze the protein–protein complex based on computational docking and mm/gbsa. Nucleic acids research, 47(W1):W322–W330, 2019. 3
|
| 320 |
+
|
| 321 |
+
J. Wu, T. Shen, H. Lan, Y. Bian, and J. Huang. Se (3)-equivariant energy-based models for end-to-end protein folding. bioRxiv, 2021. 2
|
| 322 |
+
|
| 323 |
+
Z. Xie and J. Xu. Deep graph learning of inter-protein contacts. bioRxiv, 2021. 3
|
| 324 |
+
|
| 325 |
+
K. Xu, W. Hu, J. Leskovec, and S. Jegelka. How powerful are graph neural networks? In International Conference on Learning Representations, 2018. 2
|
| 326 |
+
|
| 327 |
+
Y. Yan, Z. Wen, X. Wang, and S.-Y. Huang. Addressing recent docking challenges: A hybrid strategy to integrate template-based and free protein-protein docking. Proteins: Structure, Function, and Bioinformatics, 85(3):497–512, 2017a. 8
|
| 328 |
+
|
| 329 |
+
Y. Yan, D. Zhang, P. Zhou, B. Li, and S.-Y. Huang. Hdock: a web server for protein–protein and protein–dna/rna docking based on a hybrid strategy. Nucleic acids research, 45(W1):W365–W373, 2017b. 8
|
| 330 |
+
|
| 331 |
+
Y. Yan, H. Tao, J. He, and S.-Y. Huang. The hdock server for integrated protein–protein docking. Nature protocols, 15(5):1829–1852, 2020. 3, 8
|
| 332 |
+
|
| 333 |
+
J. Yang, I. Anishchenko, H. Park, Z. Peng, S. Ovchinnikov, and D. Baker. Improved protein structure prediction using predicted interresidue orientations. Proceedings of the National Academy of Sciences, 117(3):1496–1503, 2020. 3
|
| 334 |
+
|
| 335 |
+
# Appendix
|
| 336 |
+
|
| 337 |
+
# CONTENTS
|
| 338 |
+
|
| 339 |
+
A Representing Proteins as Graphs 15
|
| 340 |
+
B Proofs of the Main Propositions 16
|
| 341 |
+
B.1 Proof of Proposition 1. 16
|
| 342 |
+
B.2 Proof of Proposition 3. 17
|
| 343 |
+
B.3 Proof of Proposition 4. 17
|
| 344 |
+
C Surface Features 17
|
| 345 |
+
D Datasets 18
|
| 346 |
+
E More Experimental Details and Results 20
|
| 347 |
+
|
| 348 |
+
# A REPRESENTING PROTEINS AS GRAPHS
|
| 349 |
+
|
| 350 |
+
A protein is comprised of amino acid residues. The structure of an amino acid residue is shown in Figure Fig. 7. Generally, an amino acid residue contains amino (-NH-), $\alpha$ -carbon atom and carboxyl (-CO-), along with a side chain (R) connected with the $\alpha$ -carbon atom. The side chain (R) is specific to each type of amino acid residues.
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
Figure 7: Representation of an amino acid residue and its local coordinate system.
|
| 354 |
+
|
| 355 |
+
We work on residue level (our approaches can be extended to atom level as well). A protein is represented by a set of nodes where each node is an amino acid residue in the protein. Each node $i$ has a 3D coordinate $\mathbf { x } _ { i } \in \mathbb { R } ^ { 3 }$ which is the 3D coordinate of $\alpha$ -carbon atom of the residue.
|
| 356 |
+
|
| 357 |
+
The neighborhood of a node is the set of $k$ ( $k = 1 0$ in our experiments) nearest nodes where the distance is the Euclidean distance between 3D coordinates.
|
| 358 |
+
|
| 359 |
+
Node feature is a one dimension indicator (one-hot encoding) of the type of amino acid residue. This one dimension indicator will be passed into an embedding layer.
|
| 360 |
+
|
| 361 |
+
Local Coordinate System. Similar to Ingraham et al. (2019) and Jumper et al. (2021), we introduce a local coordinate system for each residue which denotes the orientation of a residue. Based on this, we can further design SE(3)-invariant edge features. As shown in Figure 7, for a residue $i$ , we denote the unit vector pointing from $\alpha$ -carbon atom to nitrogen atom as $\mathbf { u } _ { i }$ . We denote the unit vector pointing from $\alpha$ -carbon atom to carbon atom of the carboxyl (-CO-) as $\mathbf { t } _ { i }$ . $\mathbf { u } _ { i }$ and $\mathbf { t } _ { i }$ together define a plane, and the normal of this plane is $\begin{array} { r } { \mathbf { n } _ { i } = \frac { \mathbf { u } _ { i } \times \mathbf { t } _ { i } } { \left\| \mathbf { u } _ { i } \times \mathbf { t } _ { i } \right\| } } \end{array}$ . Finally, we define $\mathbf { v } _ { i } = \mathbf { n } _ { i } \times \mathbf { u } _ { i }$ . Then ${ \bf n } _ { i }$ $\mathbf { u } _ { i }$ and $\mathbf { v } _ { i }$ together form the basis of residue $i$ ’s local coordinate system. They together encode the orientation of residue $i$ .
|
| 362 |
+
|
| 363 |
+
Then we introduce the edge features of an edge $j \to i \in \mathcal { E }$ . These features describe the relative position of $j$ with respect to $i$ , the relative orientation of $j$ with respect to $i$ and the distance between $j$ and $i$ .
|
| 364 |
+
|
| 365 |
+
Relative Position Edge Features First we introduce the edge features $\mathbf { p } _ { j i }$ which describe relative position of $j$ with respect to $i$ :
|
| 366 |
+
|
| 367 |
+
$$
|
| 368 |
+
\mathbf { p } _ { j i } = [ \mathbf { \overline { { u } } } _ { i } ^ { \top } ] [ \mathbf { x } _ { j } - \mathbf { x } _ { i } ]
|
| 369 |
+
$$
|
| 370 |
+
|
| 371 |
+
Relative Orientation Edge Features As we mention above, each residue has orientation which carries information. Here we introduce the edge features $\mathbf { q } _ { j i } , \mathbf { k } _ { j i }$ and $\mathbf { t } _ { j i }$ which describe relative orientation of $j$ with respect to $i$ :
|
| 372 |
+
|
| 373 |
+
$$
|
| 374 |
+
{ \bf q } _ { j i } = [ \bf { u } _ { i } ^ { \top } ] [ { \bf n } _ { j } ] , \quad { \bf k } _ { { j } i } = [ \bf { u } _ { i } ^ { \top } ] [ { \bf { u } } _ { j } ] , \quad { \bf t } _ { { j } i } = [ \bf { u } _ { i } ^ { \top } ] [ { \bf { v } } _ { j } ]
|
| 375 |
+
$$
|
| 376 |
+
|
| 377 |
+
Distance-Based Edge Features Distance also carries information. Here we use radial basis function of distance as edge features:
|
| 378 |
+
|
| 379 |
+
$$
|
| 380 |
+
\mathbf { f } _ { j i , r } = e ^ { - \frac { ( \| \mathbf { x } _ { j } - \mathbf { x } _ { i } \| ) ^ { 2 } } { 2 \sigma _ { r } ^ { 2 } } } , r = 1 , 2 , . . . , R
|
| 381 |
+
$$
|
| 382 |
+
|
| 383 |
+
Where $R$ and scale parameters $\{ \sigma _ { r } \} _ { 1 \le r \le R }$ are hyperparameters. In experiments, the set of scale parameters we used is $\{ 1 . 5 ^ { x } | x = 0 , 1 , \bar { 2 } , \bar { . . . } , 1 4 \}$ . So for each edge, there are 15 distance-based edge features.
|
| 384 |
+
|
| 385 |
+
Surface Aware Node Features We additionally compute 5 surface aware node features defined in Eq. (16) using $\lambda \in \{ 1 . , 2 . , 5 . , 1 0 . , 3 0 . \}$ .
|
| 386 |
+
|
| 387 |
+
# B PROOFS OF THE MAIN PROPOSITIONS
|
| 388 |
+
|
| 389 |
+
B.1 PROOF OF PROPOSITION 1.
|
| 390 |
+
|
| 391 |
+
Proof. Denote the predicted ligand position by $\mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) \mathbf { X } _ { 1 } + \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) = \tilde { \mathbf { X } } _ { 1 } .$
|
| 392 |
+
|
| 393 |
+
Assume first that SE(3)-invariance of the predicted docked complex defined by Eq. (1) is satisfied. Then the transformation to dock $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ with respect to $\mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 }$ is the same as the transformation to change $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ into $\mathbf { Q } _ { 2 } \tilde { \mathbf { X } } _ { 1 } + \mathbf { g } _ { 2 }$ . We use the notation: $\mathbf { R } ^ { \top } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) = ( \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) ) ^ { \top }$ . Then, we have the following derivation steps:
|
| 394 |
+
|
| 395 |
+
$$
|
| 396 |
+
\begin{array} { r l } & { \mathbf { R } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { X } _ { 1 } + \mathbf { t } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) = \tilde { \mathbf { X } } _ { 1 } } \\ & { \mathbf { X } _ { 1 } + \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { t } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) = \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \tilde { \mathbf { X } } _ { 1 } } \\ & { \mathbf { X } _ { 1 } + \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { t } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) = \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { Q } _ { 2 } ^ { \top } ( \mathbf { Q } _ { 2 } \tilde { \mathbf { X } } _ { 1 } + \mathbf { g } _ { 2 } - \mathbf { g } _ { 2 } ) } \\ & { \mathbf { X } _ { 1 } + \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { t } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) = \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { Q } _ { 2 } ^ { \top } ( \mathbf { Q } _ { 2 } \tilde { \mathbf { X } } _ { 1 } + \mathbf { g } _ { 2 } ) - \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { Q } _ { 2 } ^ { \top } \mathbf { g } _ { 2 } } \\ & \mathbf { X } _ { 1 } + \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { t } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) + \mathbf { R } ^ { \top } ( { \mathbf { X } } _ { 1 } | { \mathbf { X } } _ { 2 } ) \mathbf { Q } _ { 2 } ^ { \top } \end{array}
|
| 397 |
+
$$
|
| 398 |
+
|
| 399 |
+
From the last equation above, one derives the transformation of $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ into $\mathbf { Q } _ { 2 } \tilde { \mathbf { X } } _ { 1 } + \mathbf { g } _ { 2 }$ , which is, by definition of the functions $\mathbf { R }$ and $\mathbf { t }$ , the same as the transformation to dock $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ with respect to $\mathbf { Q } _ { 2 } \mathbf { X } _ { 2 } + \mathbf { g } _ { 2 }$ . This transformation is
|
| 400 |
+
|
| 401 |
+
$$
|
| 402 |
+
\begin{array} { r l } & { { \bf R } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } | { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } ) = { \bf Q } _ { 2 } { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf Q } _ { 1 } ^ { \top } } \\ & { { \bf t } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } + { \bf g } _ { 1 } | { \bf Q } _ { 2 } { \bf X } _ { 2 } + { \bf g } _ { 2 } ) = { \bf Q } _ { 2 } { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) - { \bf Q } _ { 2 } { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf Q } _ { 1 } ^ { \top } { \bf g } _ { 1 } + { \bf g } _ { 2 } } \end{array}
|
| 403 |
+
$$
|
| 404 |
+
|
| 405 |
+
which concludes the proof.
|
| 406 |
+
|
| 407 |
+
Conversely, assuming constraints in Eq. (2) hold, we derive that $\mathbf { Q } _ { 1 } \mathbf { X } _ { 1 } + \mathbf { g } _ { 1 }$ is transformed into $\mathbf { Q } _ { 2 } \tilde { \mathbf { X } } _ { 1 } + \dot { \mathbf { g } } _ { 2 }$ , which then is trivial to check that it satisfies SE(3)-invariance of the predicted docked complex defined by Eq. (1).
|
| 408 |
+
|
| 409 |
+
# B.2 PROOF OF PROPOSITION 3.
|
| 410 |
+
|
| 411 |
+
Proof. We use the notation $\mathbf { R } ^ { \top } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) : = \mathbf { \Gamma } ( \mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) ) ^ { \top }$ . As in Appendix B.1, we denote $\mathbf { R } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) \mathbf { X } _ { 1 } + \mathbf { t } ( \mathbf { X } _ { 1 } | \mathbf { X } _ { 2 } ) = \tilde { \mathbf { X } } _ { 1 }$ . Then the transformation to dock $\mathbf { X } _ { 2 }$ with respect to $\mathbf { X } _ { 1 }$ is the same as the transformation to change $\tilde { \mathbf { X } } _ { 1 }$ back to $\mathbf { X } _ { 1 }$ , which is
|
| 412 |
+
|
| 413 |
+
$$
|
| 414 |
+
\begin{array} { r l } & { { \bf R } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf X } _ { 1 } + { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) = \tilde { \bf X } _ { 1 } } \\ & { { \bf X } _ { 1 } + { \bf R } ^ { \top } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) = { \bf R } ^ { \top } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) \tilde { \bf X } _ { 1 } } \\ & { { \bf X } _ { 1 } = { \bf R } ^ { \top } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) \tilde { \bf X } _ { 1 } - { \bf R } ^ { \top } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) { \bf t } ( { \bf X } _ { 1 } | { \bf X } _ { 2 } ) } \end{array}
|
| 415 |
+
$$
|
| 416 |
+
|
| 417 |
+
From the last equation above, we derive the transformation to change $\tilde { \mathbf { X } } _ { 1 }$ back to $\mathbf { X } _ { 1 }$ , which is the same as the transformation to dock $\mathbf { X } _ { 2 }$ with respect to $\mathbf { X } _ { 1 }$ . □
|
| 418 |
+
|
| 419 |
+
# B.3 PROOF OF PROPOSITION 4.
|
| 420 |
+
|
| 421 |
+
Proof. Let IEGMN la $\mathbf { X } _ { 1 } ^ { ( l + 1 ) } , \mathbf { H } _ { 1 } ^ { ( l + 1 ) } , \mathbf { X } _ { 2 } ^ { ( l + 1 ) } , \mathbf { H } _ { 2 } ^ { ( l + 1 ) } = \operatorname { I E G M N } ( \mathbf { X } _ { 1 } ^ { ( l ) } , \mathbf { H } _ { 1 } ^ { ( l ) } , \mathbf { X } _ { 2 } ^ { ( l ) } , \mathbf { H } _ { 2 } ^ { ( l ) } )$ betors f an , we $\bar { \mathbf { Q } } _ { 1 } , \mathbf { Q } _ { 2 } \in S O ( 3 )$ $\mathbf { g } _ { 1 } , \mathbf { g } _ { 2 } \bar { \in } \mathbb { R } ^ { 3 }$ want to prove that IEGMN satisfy the pairwise independent SE(3)-equivariance property:
|
| 422 |
+
|
| 423 |
+
${ \bf 2 } _ { 1 } { \bf X } _ { 1 } ^ { ( l + 1 ) } + { \bf g } _ { 1 } , { \bf H } _ { 1 } ^ { ( l + 1 ) } , { \bf Q } _ { 2 } { \bf X } _ { 2 } ^ { ( l + 1 ) } + { \bf g } _ { 2 } , { \bf H } _ { 2 } ^ { ( l + 1 ) } = \mathrm { I E G M N } ( { \bf Q } _ { 1 } { \bf X } _ { 1 } ^ { ( l ) } + { \bf g } _ { 1 } , { \bf H } _ { 1 } ^ { ( l ) } , { \bf Q } _ { 2 } { \bf X } _ { 2 } ^ { ( l ) } + { \bf g } _ { 2 } , { \bf H } _ { 2 } ^ { ( l ) } )$ where each column of $\mathbf { X } _ { 1 } ^ { ( l ) } \in \mathbb { R } ^ { 3 \times n } , \mathbf { H } _ { 1 } ^ { ( l ) } \in \mathbb { R } ^ { d \times n } , \mathbf { X } _ { 2 } ^ { ( l ) } \in \mathbb { R } ^ { 3 \times m }$ and $\mathbf { H } _ { 2 } ^ { ( l ) } \in \mathbb { R } ^ { d \times m }$ represent an individual node’s coordinate embedding or feature embedding.
|
| 424 |
+
|
| 425 |
+
We first note that the equations of our proposed IEGMN layer that compute messages $\mathbf { m } _ { j i }$ , $\mu _ { j \to i }$ , $\mathbf { m } _ { i }$ and $\mu _ { i }$ are SE(3)-invariant. Indeed, they depend on the initial features which are SE(3)-invariant by design, the current latent node embeddings $\mathbf { \bar { \{ h } } _ { i } ^ { ( l ) } \} _ { i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } }$ , as well as the Euclidean distances on the current node coordinates $\{ \mathbf { x } _ { i } ^ { ( l ) } \} _ { i \in \mathcal { V } _ { 1 } \cup \mathcal { V } _ { 2 } }$ . Thus, we also derive that the equation that computes the new latent node embeddings $\mathbf { h } _ { i } ^ { ( l + 1 ) }$ is SE(3)-invariant. Last, the equation that updates the coordinates x(l+1)i is SE(3)-equivariant with respect to the 3D coordinates of nodes from the same graph as i, but SE(3)-invariant with respect to the 3D coordinates of nodes from the other graph since it only uses invariant transformations of the latter.
|
| 426 |
+
|
| 427 |
+
# C SURFACE FEATURES
|
| 428 |
+
|
| 429 |
+
Visualization. We further discuss our new surface features introduced in Eq. (16). We first visualize their design intuition in Fig. 8. A synthetic experiment is shown in Fig. 9.
|
| 430 |
+
|
| 431 |
+
Correlation with MSMS features. Next, we analyze how accurate are these features compared to established residue depth estimation methods, e.g. based on the MSMS software (Sanner et al., 1996). We plot the Spearman rank-order correlation of the two methods in Fig. 10. We observe a concentrated distribution with a mean of 0.68 and a median of 0.70, suggesting a strong correlation with the MSMS depth estimation.
|
| 432 |
+
|
| 433 |
+
Closed form expression. Finally, we prove that for points close to the protein surface and surrounded by (infinitely) many equally-distanced and equally-spaced points, one can derive a closed form expression of the surface features defined in Eq. (16). See Fig. 11. We work in 2 dimensions, but extensions to 3 dimensions are straightforward. Assume that the local surface at point $\mathbf { x } _ { i }$ has angle $\alpha$ . Further, assume that $\mathbf { x } _ { i }$ is surrounded by $N$ equally-distanced and equally-spaced points denoted by $\mathbf { x } _ { i } ^ { \prime }$ . Then, all $w _ { i , i ^ { \prime } , \lambda }$ will be identical. Then, the summation vector in the numerator of Eq. (16) will only have non-zero components on the direction that bisects the surface angle, as the other components will cancel-out. Then, under the limit $N \infty$ , we derive the closed form expression:
|
| 434 |
+
|
| 435 |
+
$$
|
| 436 |
+
\rho _ { i } ( \mathbf { x } _ { i } ; \lambda ) = { \frac { 1 } { N } } \left\| \sum _ { i ^ { \prime } \in N _ { i } } { \frac { \mathbf { x } _ { i } - \mathbf { x } _ { i ^ { \prime } } } { \| \mathbf { x } _ { i } - \mathbf { x } _ { i ^ { \prime } } \| } } \right\| = { \frac { 2 } { N } } \sum _ { j = 0 } ^ { \frac { N } { 2 } } \cos ( { \frac { j \alpha } { N } } ) \approx _ { N \to \infty } { \frac { 2 } { \alpha } } \int _ { 0 } ^ { \alpha / 2 } \cos ( \theta ) d \theta = 2 { \frac { \sin ( \alpha / 2 ) } { \alpha } }
|
| 437 |
+
$$
|
| 438 |
+
|
| 439 |
+
|
| 440 |
+
Figure 8: Intuition behind surface features defined in Eq. (16). a. Residues in the core (interior) of a protein are likely to have a small weighted average of directionally spread neighboring forces, while b. residues close to the surface receive vector contributions from a narrower space subset and, thus, have larger $\rho$ feature values.
|
| 441 |
+
|
| 442 |
+
|
| 443 |
+
Figure 9: Distribution of our surface feature values defined in Eq. (16) for 500 points uniformly distributed in the unit circle. One can notice a strong correlation with the depth (i.e. distance to surface) which is further quantified in Fig. 10. Note that the scale for $\lambda$ in this synthetic experiment differs from that of real proteins.
|
| 444 |
+
|
| 445 |
+
# D DATASETS
|
| 446 |
+
|
| 447 |
+
The overview of datasets is in Table 2. DB5.5 is obtained from https://zlab.umassmed.edu/ benchmark/, while DIPS is downloaded from https://github.com/drorlab/DIPS. While DIPS contains only the bound structures, thus currently being only suitable for rigid docking, DB5.5 also includes unbound protein structures, however, mostly showing rigid structures - see Fig. 12.
|
| 448 |
+
|
| 449 |
+
|
| 450 |
+
Figure 10: Distribution of the Spearman rank-order coefficient computed per each protein as the correlation between MSMS residues’ depths and our surface features defined in Eq. (16) (for $\lambda = 3 0$ ). Histogram computed over the ligands in the DIPS test set (100 proteins).
|
| 451 |
+
|
| 452 |
+
|
| 453 |
+
Figure 11: For points close to the protein surface where the local surface angle is $\alpha$ we can derive a closed form expression for the surface feature defined in Eq. (16) under the assumption of being surrounded by infinitely many points at approximately equal distances and equally-spaced . A similar derivation is possible in 3D.
|
| 454 |
+
|
| 455 |
+
|
| 456 |
+
Figure 12: Distance (RMSD) between unbound and bound structures of the DB5.5 dataset reveals that most of the proteins are relatively rigid. Thus, better datasets are needed to tackle the docking conformational change problem.
|
| 457 |
+
|
| 458 |
+
Table 2: Overview of Datasets. For DIPS, the statistics of number of residues and atoms per protein is based on a subset consisting of 200 proteins.
|
| 459 |
+
|
| 460 |
+
<table><tr><td colspan="4">Dataset # Pairs of Proteins # Residues per Protein # Atoms per Protein</td></tr><tr><td>DIPS</td><td>41876</td><td>276 (±189)</td><td>2159 (±1495)</td></tr><tr><td>DB5.5</td><td>253</td><td>268 (±215)</td><td>2089 (±1694)</td></tr></table>
|
| 461 |
+
|
| 462 |
+
# E MORE EXPERIMENTAL DETAILS AND RESULTS
|
| 463 |
+
|
| 464 |
+
Baseline Failures. On the test sets, ATTRACT fails for ’1N2C’ in DB5.5, ’oi_4oip.pdb1_8’, ’oi_4oip.pdb1 $_ { - 3 } ,$ and $\mathrm { ^ { , } p 7 } _ { - } 4 \mathrm { p 7 s . p d b 1 } _ { - } 2 \mathrm { ^ { , } }$ in DIPS. For such failure cases, we use the unbound input structure as the prediction for metrics calculation.
|
| 465 |
+
|
| 466 |
+
Hyperparameters. We perform hyperparameter search over the choices listed in Table 3 and select the best hyperparameters for DB5.5 and DIPS respectively based on their corresponding validation sets.
|
| 467 |
+
|
| 468 |
+
Table 3: Hyperparameter choices. LN stands for layer normalization, BN stands for batch normalization.
|
| 469 |
+
|
| 470 |
+
<table><tr><td>Hyperparameter</td><td>Choice</td></tr><tr><td>Node degree (for k-NN)</td><td>10</td></tr><tr><td>Weight of the intersection loss</td><td>0.0, 1.0</td></tr><tr><td>Normalization for hi of IEGMN layers</td><td>No,LN</td></tr><tr><td>Normalization for others</td><td>No, BN, LN</td></tr><tr><td>Number of attention heads (K)</td><td>25,50,100</td></tr><tr><td>Slope of leaky relu</td><td>0.1, 0.01</td></tr><tr><td>Dimension of hi of IEGMN layers</td><td>32,64</td></tr><tr><td>Dimension of residue type embedding</td><td>32,64</td></tr><tr><td>Number of IEGMN layers</td><td>5,8</td></tr><tr><td>If IEGMN layers except the first one share parameters T</td><td>True,False</td></tr><tr><td>η of coordinates skip connection</td><td>0.0, 0.25</td></tr><tr><td>Weight decay</td><td>0, 1e-5, 1e-4, 1e-3</td></tr></table>
|
| 471 |
+
|
| 472 |
+
Detailed Running Times. In addition to the main text, we show in Table 4 detailed running times of all methods. Hardware specifications are as follows: ATTRACT was run on a 6-Core Intel Core i7 $2 . 2 \operatorname { G H z }$ CPU; HDOCK was run on a single Intel Xeon Gold 6230 2.1 GHz CPU; EQUIDOCK was run on a single Intel Core i9-9880H 2.3 GHz CPU. CLUSPRO and PATCHDOCK have been manually run using their respective web servers.
|
| 473 |
+
|
| 474 |
+
Plots for DB5.5. We show the corresponding plots for DB5.5 results in Fig. 13.
|
| 475 |
+
|
| 476 |
+
Table 4: Inference time comparison (in seconds). Note: ClusPro and PatchDock were run manually using the respective public webservers, thus their runtimes are influenced by their cluster load.
|
| 477 |
+
|
| 478 |
+
<table><tr><td rowspan="2">Methods</td><td colspan="5">Runtime on DIPS Test Set</td><td colspan="5">Runtime on DB5.5 Test Set</td></tr><tr><td>Mean</td><td>Median</td><td>Min</td><td>Max</td><td>Std</td><td>Mean</td><td>Median</td><td>Min</td><td>Max</td><td>Std</td></tr><tr><td>ATTRACT (LOCAL)</td><td>1285</td><td>793</td><td>62</td><td>8192</td><td>793</td><td>570</td><td>524</td><td>180</td><td>1708</td><td>373</td></tr><tr><td>HDOCK (LOCAL)</td><td>778</td><td>635</td><td>145</td><td>3177</td><td>570</td><td>615</td><td>461</td><td>210</td><td>2593</td><td>459</td></tr><tr><td>CLUSPRO (WEB)</td><td>10475</td><td>9831</td><td>2632</td><td>22654</td><td>4512</td><td>15507</td><td>14393</td><td>9207</td><td>28528</td><td>4126</td></tr><tr><td>PATCHDOCK(WEB)</td><td>7378</td><td>6900</td><td>600</td><td>16560</td><td>3979</td><td>3290</td><td>2820</td><td>1080</td><td>14520</td><td>2459</td></tr><tr><td>EQUIDOCK (LOCAL)</td><td>5</td><td>3</td><td>1</td><td>22</td><td>5</td><td>5</td><td>3</td><td>1</td><td>53</td><td>10</td></tr></table>
|
| 479 |
+
|
| 480 |
+
Ablation Studies. To highlight contributions of different model components, we provide ablation studies in Table 5. One can note that, as expected, removing the pocket loss results in lower interface RMSD scores compared to removing other components.
|
| 481 |
+
|
| 482 |
+
Analysis of the Intersection Loss. We further analyze the intersection loss introduced in Eq. (15) with parameters $\gamma = 1 0$ and $\sigma = 2 5$ (chosen on DB5 validation set). We show in Table 6 that this loss achieves almost perfect values for the ground truth structures, being important to softly constrain non-intersecting predicted proteins.
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
Figure 13: DB5.5 test results: a. Complex-RMSD distributions; b. Interface-RMSD distributions; c. scatter plot for C-RMSD vs I-RMSD.
|
| 486 |
+
|
| 487 |
+
Table 5: Ablation studies. We show DIPS test median C-RMSD and I-RMSD values for the corresponding best validation models. Abbreviations: “intersection $\mathbf { l o s s ^ { \prime \prime } = }$ intersection loss in Eq. (15), “pocket loss” $=$ pocket loss in Eq. (14), “surface feas” $=$ surface features in Eq. (16).
|
| 488 |
+
|
| 489 |
+
<table><tr><td>Model</td><td>C-RMSD I-RMSD</td></tr><tr><td>Full model</td><td>13.29</td><td>10.18</td></tr><tr><td>without pocket loss</td><td>15.91</td><td>12.01</td></tr><tr><td>without pocket loss,intersection loss</td><td>16.43</td><td>12.92</td></tr><tr><td>without pocket loss,surface feas</td><td>14.80</td><td>13.10</td></tr><tr><td>without pocket loss,intersection loss,surface feas</td><td>15.19</td><td>11.38</td></tr><tr><td>without surface feas</td><td>13.73</td><td>10.65</td></tr><tr><td>without intersection loss</td><td>15.49</td><td>11.09</td></tr><tr><td>without intersection loss, surface feas</td><td>15.04</td><td>10.94</td></tr></table>
|
| 490 |
+
|
| 491 |
+
Table 6: Values of the intersection loss defined in Eq. (15) and evaluated on the DIPS validation set in different scenarios. “Centered structures” means that both ground truth ligand and receptor point clouds have been centered (0-mean), without any other modifications.
|
| 492 |
+
|
| 493 |
+
<table><tr><td rowspan=1 colspan=1>Centeredstructures</td><td rowspan=1 colspan=1>EquiDock trainedwith intersection loss</td><td rowspan=1 colspan=1>EquiDock trainedwithout intersection loss</td><td rowspan=1 colspan=1>Ground truthcomplexes</td></tr><tr><td rowspan=1 colspan=1>56.42</td><td rowspan=1 colspan=1>12.68</td><td rowspan=1 colspan=1>21.03</td><td rowspan=1 colspan=1>1.16</td></tr></table>
|
parse/dev/GQjaI9mLet/GQjaI9mLet_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/GQjaI9mLet/GQjaI9mLet_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/GQjaI9mLet/GQjaI9mLet_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Jep2ykGUdS/Jep2ykGUdS.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Jep2ykGUdS/Jep2ykGUdS_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Jep2ykGUdS/Jep2ykGUdS_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Jep2ykGUdS/Jep2ykGUdS_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/QwejPcX96r/QwejPcX96r.md
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/QwejPcX96r/QwejPcX96r_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/S7Evzt9uit3/S7Evzt9uit3.md
ADDED
|
@@ -0,0 +1,279 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning
|
| 2 |
+
|
| 3 |
+
Weixin Liang⇤ Stanford University wxliang@stanford.edu
|
| 4 |
+
|
| 5 |
+
Yuhui Zhang ⇤ Stanford University yuhuiz@stanford.edu
|
| 6 |
+
|
| 7 |
+
Yongchan Kwon ⇤ Columbia University yk3012@columbia.edu
|
| 8 |
+
|
| 9 |
+
Serena Yeung Stanford University syyeung@stanford.edu
|
| 10 |
+
|
| 11 |
+
James Zou Stanford University jamesz@stanford.edu
|
| 12 |
+
|
| 13 |
+
# Abstract
|
| 14 |
+
|
| 15 |
+
We present modality gap, an intriguing geometric phenomenon of the representation space of multi-modal models. Specifically, we show that different data modalities (e.g. images and text) are embedded at arm’s length in their shared representation in multi-modal models such as CLIP. Our systematic analysis demonstrates that this gap is caused by a combination of model initialization and contrastive learning optimization. In model initialization, we show empirically and theoretically that the representation of a common deep neural network is restricted to a narrow cone. As a consequence, in a multi-modal model with two encoders, the representations of the two modalities are clearly apart when the model is initialized. During optimization, contrastive learning keeps the different modalities separated by a certain distance, which is influenced by the temperature parameter in the loss function. Our experiments further demonstrate that varying the modality gap distance has a significant impact in improving the model’s downstream zeroshot classification performance and fairness. Our code and data are available at https://modalitygap.readthedocs.io/
|
| 16 |
+
|
| 17 |
+
# 1 Introduction
|
| 18 |
+
|
| 19 |
+
Multi-modal models map inputs from different data modalities (e.g. image and text) into a shared representation space (Figure 1 (a)). It has garnered tremendous interest and excitement as a framework for data integration. As a prominent example pre-trained on a web-scale collection of images and natural language, OpenAI’s CLIP model [39], has learned diverse visual concepts that can readily be transferred to downstream tasks through prompting: one can perform “zero-shot” visual classification by simply providing the names of the visual categories to be recognized.
|
| 20 |
+
|
| 21 |
+
In this work, we present the modality gap phenomenon: As shown in Figure 1 (b), CLIP’s image embeddings and text embeddings are located in two completely separate regions of the embedding space. We find this phenomenon consistently across various multi-modal models, covering texts, natural images [39], videos [50], medical images [53], and amino-acid sequences [11]. Interestingly, this phenomenon still holds even when we embed using multi-modal models with random weights (Figure 1 (c)). While it might seem reasonable to attribute the gap to differences in data distributions or to the different encoder architectures, we showed that these factors are not the fundamental cause.
|
| 22 |
+
|
| 23 |
+
This paper provides a three-part explanation for the modality gap phenomenon. (1) The general inductive bias of deep neural architecture creates a cone effect: The effective embedding space is restricted to a narrow cone for pre-trained models or models with random weights. (2) Different random initializations create different embedding cones. Since a multi-modal model consists of two encoders, which create different cones at random initialization, this explains how the modality gap is present at initialization. (3) The contrastive learning objective commonly used by multi-modal models preserves the gap. We support our explanations with theory and experiments. Our theoretical analysis shows that under mild assumptions, each neural network layer shrinks the angle between any pair of embedding vectors with high probability, thereby creating more narrow cones in deeper architectures. We further prove that different random initializations of model weights result in different cones. Interestingly, increasing the modality gap in models like CLIP can improve its downstream performance on several zero-shot learning and fairness tasks. The main objective of our paper is to i) empirically demonstrate the modality gap phenomenon across different data modalities and NN architectures; ii) explain how the gap arises and iii) show that the size of the gap can affect downstream applications. It is not our goal to propose a method to close the gap, since it’s not clear that it’s desirable to have no modality gap. Together, this paper makes the following contributions:
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
Figure 1: The pervasive modality gap in multi-modal contrastive representation learning. (a) Overview of multi-modal contrastive learning. Paired inputs from two modalities (e.g., image-caption) are sampled from the dataset and embedded into the hypersphere using two different encoders. The loss function is to maximize the cosine similarity between matched pairs given all the pairs within the same batch. (b) UMAP visualization of generated embeddings from pre-trained models. Paired inputs are fed into the pre-trained models and the embeddings are visualized in 2D using UMAP (lines indicate pairs). We observe a clear modality gap for various models trained on different modalities. (c) UMAP visualization of generated embeddings from same architectures with random weights. Modality gap exists in the initialization stage without any training.
|
| 27 |
+
|
| 28 |
+
1. To the best of our knowledge, we demonstrate a general modality gap phenomenon for the first time. We show that this phenomenon holds across a wide spectrum of multi-modal models, covering texts, natural images, videos, medical images, and amino-acid sequences. 2. We demonstrate the significant implications of modifying the gap in downstream applications. By simply modifying the gap’s distance, we can improve CLIP’s zero-shot performance and fairness. 3. To explain modality gap, we provide a three-part explanation supported by extensive theoretical and empirical analyses. Our analyses also provide new insights on the cone effect, which we show is a general phenomenon for deep neural networks. Existing work focuses on trained language models and attributes the cone effect to the optimization under unbalanced word frequencies distribution. We demonstrate that this effect holds not only across various modalities and network architectures, but also on random noise inputs and random weights, which is not captured in previous work.
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
(c) UMAP visualization of embeddings of 25 randomly initialized models on real data (color indicates random seed)
|
| 32 |
+
Figure 2: The cone effect phenomenon. (a) Histograms of the cosine similarity between all pairs of embeddings across various settings. The average cosine similarity is substantially larger than 0, indicating that the embedding space is a narrow cone. The cone effect also holds on randomly initialized models, and on random noise inputs. (b) Effects of nonlinear activation and depth. Inputs are 512-dim standard normal random vector. All MLP linear layers are $5 1 2 \times 5 1 2$ , with both weight and bias randomly initialized from $\textstyle { \mathcal { N } } ( 0 , { \frac { 1 } { 5 1 2 } } )$ . Y axis is the average cosine similarity between pairs of embeddings. (c) UMAP visualization of embeddings of 25 randomly initialized models (without training) on real data. Each random initialization forms a distinctively different cone. Real Data: 5,000 image-caption pairs from the validation set of MSCOCO Caption. Random Noise: Gaussian noise from the standard normal distribution as images, uniformly random integer sequences as texts.
|
| 33 |
+
|
| 34 |
+
4. We mathematically characterize the contraction mapping induced by linear layers with ReLU non-linearities to explain the cone effect. Our theory matches well with experiments and provides insights for understanding the general inductive biases of deep neural networks.
|
| 35 |
+
|
| 36 |
+
# 2 The Cone Effect Induces A Modality Gap
|
| 37 |
+
|
| 38 |
+
# 2.1 The Narrow Cone of Embeddings
|
| 39 |
+
|
| 40 |
+
In order for modality gap to exist, the embeddings from a encoder should be concentrated around a subregion of the full embedding space—otherwise, the embeddings from different encoders would overlap. Motivated by this, we begin our investigation by showing that the modality gap already arises at random model initialization due to the cone effect: The effective embedding space is restricted to a narrow cone for trained models and models with random weights. To demonstrate this, we extract 5,000 embeddings from the final layer of 3 pre-trained models respectively (ResNet, Vision Transformer, Text Transformer)2 on MSCOCO Caption [8]. We then compute the cosine similarity between all possible pairs of the 5,000 embeddings within each model (Figure 2 (a)). We found that both the average cosine similarity (0.56, 0.47, 0.51 respectively for the 3 models) and the minimum cosine similarity (0.23, 0.05, 0.01) are positive. These results indicate that the embedding space is a narrow cone.
|
| 41 |
+
|
| 42 |
+
In the literature, the cone effect has been observed in the language representations from language models (e.g., BERT) [12]. A common explanation is that the unbalanced distribution of word frequencies biased the optimization [15, 33]. However, we found that the cone effect still exists in models with random weights (Figure 2 (c)). In fact, the average cosine similarity there is even higher than in trained models. For example, any two embeddings from a randomly initialized ResNet have on average an almost perfect (0.99) cosine similarity. Interestingly, the cone effect still holds when the input data is random noise3, indicating that unbalanced data distribution suggested in previous works is not necessary for the cone effect. Together these experiments suggest that the cone effect reflects a more general inductive bias of deep networks than might be previously appreciated.
|
| 43 |
+
|
| 44 |
+
How narrow is the cone in 512-dim representation space? We clarify that a cosine similarity with 0.56 already indicates that the embedding space is actually an extremely narrow cone in the 512-dimensional feature space. Consider the fraction of surface area in a unit hypersphere: In 2D, arccos $( 0 . 5 6 ) { = } 5 5 . 9 4 ^ { \circ }$ , indicating that a cosine similarity of 0.56 can “occupy” $5 5 . 9 4 ^ { \circ } / 3 6 0 ^ { \circ } = 1 5 . 5 3 \%$ of the 2D unit circle. In 3D, a cosine similarity of 0.56 can “occupy” 2⇡r2(1 cos 55.94°2 )4⇡r2 of the 3D unit sphere. In 512D, a cosine similarity of 0.56 can “occupy” less than $\frac { 1 } { 2 ^ { 5 1 2 } }$ fraction of the surface area in a unit 512D hypersphere. These evidences show that the effective embedding space is restricted to an extremely narrow cone.
|
| 45 |
+
|
| 46 |
+
# 2.2 The effects of non-linear activation on cone effect
|
| 47 |
+
|
| 48 |
+
Design To study the effects of non-linear activation functions on the cone effect, we randomly initialized various MLPs with different non-linearities or without non-linearities. The inputs of the MLPs are 512-dim standard normal random vectors. All MLP linear layers are $5 1 2 \times 5 1 2$ , with both weight and bias randomly initialized from $\begin{array} { r } { \mathcal { N } ( 0 , \frac { 1 } { 5 1 2 } ) } \end{array}$ , here we denote a Gaussian distribution with mean $\mu$ and variance $\sigma ^ { 2 }$ by ${ \mathcal { N } } ( \mu , \sigma ^ { 2 } )$ .
|
| 49 |
+
|
| 50 |
+
Results As shown in Figure 2 (b), MLPs without non-linear activation shows little cone effect. However, with non-linearity, the average cosine similarity increases rapidly as the number of layers increases. For example, the average cosine similarity reaches 0.99 for a 2-layer MLP with Sigmoid. These results indicate that the non-linear activation functions play a crucial role in the cone effect.
|
| 51 |
+
|
| 52 |
+
Although it is easy to see that ReLU makes every coordinate non-negative, and thus cosine similarity after ReLU is guaranteed to be non-negative, we highlight that none of the 3 models in Figure 2 (a) has ReLU as the final layer before embedding extraction4. In addition, although all 3 models incorporate normalization layers such as batch norm [23] and layer norm [4] in their architectures, we still observe the cone effect. Further analyzing the connection between normalization and the cone effect is an interesting direction of future work.
|
| 53 |
+
|
| 54 |
+
# 2.3 Different random initializations create different cones
|
| 55 |
+
|
| 56 |
+
Next, we study the effect of different random initialization on the cone effect. In Figure 2 (c), we randomly initialized a model 25 times, and plotted its extracted embeddings on the same real data (i.e., MSCOCO Caption) via UMAP visualization [41]. We found that each random initialization forms a distinctively different cone. This phenomenon holds across various neural network architectures and input modalities (ResNet, Vision Transformer or Text Transformer), on ImageNet-pretrained models (Supp. Figure 13), on PCA visualization (Supp. Figure 7), or with random noise inputs (Supp. Figure 5). Since a multi-modal model consists of two encoders, which creates different cones at random initialization, this explains how the modality gap is present at initialization. While it might seem reasonable to attribute the modality gap to differences in data modalities [21], Figure 2 (c) shows the gap still exists even if the two encoders operate on the exact same data in the exact same modality. Therefore, the gap can exist without different modalities, and we emphasize that the modality gap phenomenon is non-trivial to understand.
|
| 57 |
+
|
| 58 |
+
# 3 Theoretical analysis
|
| 59 |
+
|
| 60 |
+
Here, we theoretically investigate the cone effect phenomenon. We show that (i) the cosine similarity increases as the layer gets deeper and (ii) the variance of an intermediate output mostly come from the model’s random initialization.
|
| 61 |
+
|
| 62 |
+
We first define some notations. We denote the ReLU activation by $\phi ( x ) \ : = \ \operatorname* { m a x } ( x , 0 )$ for $x \in \mathbb { R }$ , and we extend it by considering element-wise operation $\phi ( \mathbf { x } ) : = ( \phi ( x _ { 1 } ) , \ldots , \phi ( x _ { k } ) ) ^ { T } =$ $( \operatorname* { m a x } ( x _ { 1 } , 0 ) , \dots , \operatorname* { m a x } ( x _ { k } , 0 ) ) ^ { T }$ for a multivariate input $\mathbf { x } \ = \ ( x _ { 1 } , \ldots , x _ { k } ) ^ { T } \ \in \ \mathbb { R } ^ { k }$ and $k \in \mathbb N$ . The cosine similarity between two vectors $u , v \in \mathbb { R } ^ { k }$ is defined as $\begin{array} { r } { \cos ( u , v ) : = \frac { u ^ { T } v } { \| u \| \| v \| } } \end{array}$ where $\lVert \boldsymbol { u } \rVert = ( u ^ { T } u ) ^ { 1 / 2 }$ . Lastly, we set $[ k ] : = \{ 1 , \ldots , k \}$ for $k \in \mathbb N$ .
|
| 63 |
+
|
| 64 |
+
Each network layer increases cosine similarity. We study how the cosine similarity between two intermediate layer outputs changes when weight and bias terms in an MLP are fixed. The following theorem shows that with a high probability cosine similarity increases after one feedforward computation when the number of nodes in the output layer is large.
|
| 65 |
+
|
| 66 |
+
Theorem 1 (Monotonicity of cosine similarity). Suppose $u , v \in \mathbb { R } ^ { d _ { \mathrm { i n } } }$ are any two fixed vectors such that $\| u \| = r \| v \|$ for some $r > 0$ , $\mathbf { W } \in \mathbb { R } ^ { d _ { \mathrm { o u t } } \bar { \times } d _ { \mathrm { i n } } }$ is a random weight matrix where each element $\mathbf { W } _ { k , l } \sim \mathcal { N } ( 0 , d _ { \mathrm { o u t } } ^ { - 1 } ) f o r \ k \in [ d _ { \mathrm { o u t } } ]$ , $l \in [ d _ { \mathrm { i n } } ]$ , and $\mathbf { b } \in \mathbb { R } ^ { d _ { \mathrm { o u t } } }$ is a random bias vector such that ${ \bf b } _ { k } \sim \mathcal N ( 0 , d _ { \mathrm { o u t } } ^ { - 1 } )$ for $k \in [ d _ { \mathrm { o u t } } ]$ . $\begin{array} { r } { I f \cos ( u , v ) < \left( \frac { 1 } { 2 } \left( r + \frac { 1 } { r } \right) \right) ^ { - 1 } } \end{array}$ , then the following holds with probability at least $1 - O ( 1 / d _ { \mathrm { o u t } } )$ .
|
| 67 |
+
|
| 68 |
+
$$
|
| 69 |
+
\mathrm { c o s } ( \phi ( \mathbf { W } u + \mathbf { b } ) , \phi ( \mathbf { W } v + \mathbf { b } ) ) > \mathrm { c o s } ( u , v ) .
|
| 70 |
+
$$
|
| 71 |
+
|
| 72 |
+
Theorem 1 shows that the cosine similarity between two vectors increases with a high probability after one feedforward computation consisting of a linear transformation and ReLU computation. This matches well with the result in Figure 2 (b) where the cosine similarity between samples increases as the intermediate layer gets farther from the input.
|
| 73 |
+
|
| 74 |
+
The bound condition on $\cos ( u , v )$ in Theorem 1 asks that the two vectors before the layer computation are not too close to each other in terms of the direction. This is because the random bias addition can slightly change the angle between the two vectors, leading to a small decrease in cosine similarity when the previous layer’s cosine similarity is too high. This condition is plausible in practice because the $\ell ^ { 2 }$ -norm of intermediate layer outputs is close to one with a high probability when the $\ell ^ { 2 }$ -norm of input data is one [1, Lemma 7.1]. Given that the norm ratio $r$ is close to one, the upper bound condition for $\cos ( u , v )$ is likely to hold because $\begin{array} { r } { ( \frac { 1 } { 2 } ( r + \frac { 1 } { r } ) ) ^ { - 1 } } \end{array}$ is close to 1.
|
| 75 |
+
|
| 76 |
+
Effect of random initialization We now examine the variance of an intermediate output and explain that the variance is mainly due to random initializations as in Figure 2 (c). To be more specific, we denote an intermediate layer output by $h _ { \Theta } ( U ) \in \mathbb { R }$ for some input datum $U$ . Here, $\Theta$ denotes all the random weights and biases that are used in $h _ { \Theta } ( U )$ . The variance of $h _ { \Theta } ( U )$ can be decomposed as
|
| 77 |
+
|
| 78 |
+
$$
|
| 79 |
+
\mathrm { { V a r } } [ h _ { \Theta } ( U ) ] = \underbrace { { \mathbb { E } } [ \mathrm { { V a r } } [ h _ { \Theta } ( U ) \mid \Theta ] ] } _ { \mathrm { { D u e ~ t o ~ t h e ~ r a n d o m n e s s ~ o f ~ d a t a } } } + \underbrace { \mathrm { { V a r } } [ { \mathbb { E } } [ h _ { \Theta } ( U ) \mid \Theta ] ] . } _ { \mathrm { { D u e ~ t o ~ r a n d o m ~ i n i t i a l i z a t i o n s } } }
|
| 80 |
+
$$
|
| 81 |
+
|
| 82 |
+
Here, the inner and outer expectations are over the data $U$ and the random weights $\Theta$ , respectively. The first term on the right hand side explains the within variance after fixing one random initialization, quantifying the randomness of data. In contrast, the second term explains the variance due to different random initializations. The following theorem considers the ratio of the second term to the total variance and shows that the ratio can be very close to one when a deep neural network model is used.
|
| 83 |
+
|
| 84 |
+
Theorem 2 (Informal; Variance due to different random initializations). Let $h _ { \Theta } ( U )$ be an intermediate layer output with an input data $U$ with $\| U \| = 1$ . Under mild assumptions on $\Theta$ , the set of all the random weights and biases, the following inequality holds.
|
| 85 |
+
|
| 86 |
+
$$
|
| 87 |
+
\frac { \mathrm { V a r } [ \mathbb { E } [ h _ { \Theta } ( U ) \mid \Theta ] ] } { \mathrm { V a r } [ h _ { \Theta } ( U ) ] } \ge \beta ,
|
| 88 |
+
$$
|
| 89 |
+
|
| 90 |
+
where $\beta$ is a constant that captures the average cosine similarity of previous layer outputs.
|
| 91 |
+
|
| 92 |
+
Theorem 2 shows that the ratio of the variance due to different random initializations to the total variance is bounded below by the average cosine similarity of previous layer outputs. As Figure 2 (b) illustrated, the average cosine similarity of an intermediate layer output often approaches to one as the layer gets deeper. Accordingly, the lower bound $\beta$ , which captures the average cosine similarity, is close to one when a neural network is deep enough. In Appendix D, we elaborate on the relationship between $\beta$ and the cosine similarity, and provide a detailed statement of the Theorem.
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
Figure 3: Contrastive learning preserves modality gap. (a) Embedding shift experiment. To probe the loss landscape of CLIP, we manually shift the image embeddings and text embeddings towards closing the gap. (b-d) The loss landscapes under different temperatures. Y axis indicates the contrastive loss. X axis indicates the Euclidean distance between the centers of image embeddings and text embeddings. The vertical dash line $x = 0 . 8 2$ indicates CLIP’s original distance between image and text embeddings (i.e., without any shifting). Note that in CLIP, the image embeddings and text embeddings are L2-normalized (Supplementary Figure 12). In other words, the image and text embeddings of CLIP are always on the unit sphere. (e-g) Simulation analysis for the loss landscape. Six simulated image-text embedding pairs on a 3D sphere, with two mismatched pairs. Text embeddings are shifted towards closing the modality gap (i.e., modifying $\theta$ ).
|
| 96 |
+
|
| 97 |
+
# 4 Contrastive learning preserves modality gap
|
| 98 |
+
|
| 99 |
+
# 4.1 Background: Contrastive Loss
|
| 100 |
+
|
| 101 |
+
Given that the modality gap is present at initialization, we investigate why our optimization procedure fails to close the gap. We begin by reviewing contrastive learning, which is a commonly used training strategy for multi-modal models [53, 50, 34]. We illustrate with CLIP due to its wide usage.
|
| 102 |
+
|
| 103 |
+
Given a batch of $N$ (image, text) pairs, CLIP learns to predict which of the $N \times N$ possible (image, text) pairs are aligned. In other words, CLIP learns to maximize the cosine similarity of the image and text embeddings of the $N$ real pairs in the batch while minimizing the cosine similarity of the embeddings of the $N ^ { 2 } - N$ incorrect pairs. Formally, the optimization objective is the average of two losses: one for image-to-text classification:
|
| 104 |
+
|
| 105 |
+
$$
|
| 106 |
+
\mathcal { L } _ { \mathbb { Z } \mathcal { T } } = - \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \log \frac { \exp ( \mathbf { x } _ { i } \cdot \mathbf { y } _ { i } / \tau ) } { \sum _ { j = 1 } ^ { N } \exp ( \mathbf { x } _ { i } \cdot \mathbf { y } _ { j } / \tau ) }
|
| 107 |
+
$$
|
| 108 |
+
|
| 109 |
+
and the other for text-to-image classification:
|
| 110 |
+
|
| 111 |
+
$$
|
| 112 |
+
\mathcal { L } _ { \mathcal { T } \mathcal { T } } = - \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \log \frac { \exp ( \mathbf { x } _ { i } \cdot \mathbf { y } _ { i } / \tau ) } { \sum _ { j = 1 } ^ { N } \exp ( \mathbf { x } _ { j } \cdot \mathbf { y } _ { i } / \tau ) }
|
| 113 |
+
$$
|
| 114 |
+
|
| 115 |
+
Here, $\mathbf { x } _ { i }$ and $\mathbf { y } _ { j }$ are the L2-normalized embedding of image in the $i$ -th pair and that of text in the $j$ -th pair, respectively. $\tau$ is a learned temperature parameter to scale the logits. The final learned temperature is $\begin{array} { r } { \tau = \frac { \textbf { \check { 1 } } } { 1 0 0 } } \end{array}$ in CLIP. See additional illustration in Figure 1(a) and Supp. Figure 12.
|
| 116 |
+
|
| 117 |
+
# 4.2 Embedding Shift Experiment
|
| 118 |
+
|
| 119 |
+
Design We hypothesize that the contrastive learning objective encourages the existence of the modality gap. To testify this hypothesis, we design a loss landscape probing experiment on $n = 5 , 0 0 0$ image-caption pairs5 from the validation set of MSCOCO Caption dataset. We first define the modality gap as the difference between the center of image embeddings and text embeddings:
|
| 120 |
+
|
| 121 |
+
$$
|
| 122 |
+
{ \vec { \Delta } } _ { \mathrm { g a p } } = { \frac { 1 } { n } } \sum _ { i = 1 } ^ { \breve { n } } \mathbf { x } _ { i } - { \frac { 1 } { n } } \sum _ { i = 1 } ^ { n ^ { \breve { } } } \mathbf { y } _ { i }
|
| 123 |
+
$$
|
| 124 |
+
|
| 125 |
+
where $\mathbf { x } _ { i }$ and ${ \bf y } _ { i }$ are the L2-normalized image embedding and text embedding. We then manually shift every text embedding and image embedding towards closing the modality gap (Figure 3 (a)). After shifting, we re-normalize each embedding to the unit hypersphere:
|
| 126 |
+
|
| 127 |
+
$$
|
| 128 |
+
\mathbf { x } _ { i } ^ { \mathrm { s h i f t } } = \mathrm { N o r m a l i z e } ( \mathbf { x } _ { i } - \lambda \vec { \Delta } _ { \mathrm { g a p } } ) , \quad \mathbf { y } _ { i } ^ { \mathrm { s h i f t } } = \mathrm { N o r m a l i z e } ( \mathbf { y } _ { i } + \lambda \vec { \Delta } _ { \mathrm { g a p } } ) .
|
| 129 |
+
$$
|
| 130 |
+
|
| 131 |
+
We vary the scalar $\lambda$ to produce different amounts of shifts. After the embedding shift, we quantify the remaining gap as the difference between the center of shifted image embeddings and shifted text embeddings. The gap distance before shifting is $\| \vec { \Delta } _ { \mathrm { g a p } } \| = 0 . 8 2$ . Here Euclidean distance is a intuitive metric because in CLIP, the image embeddings and text embeddings are L2-normalized (Supplementary Figure 12). In other words, the image and text embeddings of CLIP are always on the unit sphere. Specifically, for any $n$ -dimensional vectors $x$ and $y$ , the cosine similarity is given as $\cos ( x , y ) { \dot { = } } x ^ { T } y$ , and the Euclidean distance is given as $( x - y ) ^ { T } ( x - y ) = 2 ( 1 - x ^ { T } y )$ . Therefore, they have a functional relationship as Euclideandistance $\langle x , y \rangle = 2 ( 1 - \cos ( x , y ) )$ . When the angle between $x$ and $y$ is less than $\pi / 2$ , which is the case as embeddings are in a narrow cone, the small Euclidean distance directly means a high cosine similarity.
|
| 132 |
+
|
| 133 |
+
Results Figure 3(b) shows the contrastive loss landscape on different amount of modality gap under temperature ⌧ = 1100 (i.e., CLIP’s learned final temperature). We found that the default gap distance $\| \vec { \Delta } _ { \mathrm { g a p } } \| = 0 . 8 \bar { 2 }$ actually achieves the global minimum, and shifting toward closing the gap increases the contrastive loss. Interestingly, there is a local minimum when we shift the text embeddings to the opposite side in a “back-to-back position.” Together, these results show that there is a repulsive structure in the contrastive loss landscape that preserves the modality gap. However, when the temperature increases (Figure 3(c,d)), the repulsive structure and the local minimum gradually disappear, and closing the gap becomes more optimal. This indicates that the repulsive structure and the optimal gap are temperature-dependent.
|
| 134 |
+
|
| 135 |
+
Additional Evidence from Fine-tuning To further investigate the impact of temperature on modality gap, we fine-tune CLIP under 6 different temperatures $\begin{array} { r } { \overline { { \tau } } \in \{ \frac { 1 } { 1 0 0 } , \frac { \overline { { 1 } } } { 5 0 } , \frac { 1 } { 3 0 } , \frac { 1 } { 2 0 } , \frac { \overline { { 1 } } } { 1 0 } , 1 \} } \end{array}$ respectively, on MSCOCO Caption training set with batch size 64. We found that a high temperature $( \tau \in \{ \frac { 1 } { 1 0 } , 1 \} )$ ) in fine-tuning significantly reduces or closes the gap, while a low temperature does not. The gap distance $\| \vec { \Delta } _ { \mathrm { g a p } } \|$ decreases monotonically with increasing temperature (Supp. Figure 8).
|
| 136 |
+
|
| 137 |
+
# 4.3 Simulating mismatched data
|
| 138 |
+
|
| 139 |
+
Design We designed a simple simulation to distill the empirical phenomena in the embedding shift experiment. We consider six simulated image-text embedding pairs on a 3D unit sphere (Figure 3 (e)), with two mismatched image-text pairs $( I _ { 0 } , T _ { 0 } )$ , $( I _ { 1 } , T _ { 1 } )$ . Here "mismatched" means correct pairs are $( I _ { 0 } , T _ { 0 } )$ and $( I _ { 1 } , T _ { 1 } )$ but $I _ { 0 }$ is closer to $T _ { 1 }$ and $I _ { 1 }$ is closer to $T _ { 0 }$ . We fix the image embeddings while shifting the text embeddings downwards to close the gap (i.e., modifying $\theta$ , see more details in Appendix A).
|
| 140 |
+
|
| 141 |
+
Results With mismatched data, our simulation model successfully reproduces the temperaturedependent repulsive structure in the optimization landscape. When we remove the mismatch, the repulsive structure disappears (Supp. Figure 9). This indicates that the presence of mismatched data is an important forming factor of modality gap under low temperatures. Although the mismatch here is simulated, in practice mismatched data are common (e.g., hard-to-differentiate images/captions or annotation errors). Investigating how and to what extent the multimodal data misalignment could affect the contrastive loss landscape and thereby the modality gap is an interesting direction for future research.
|
| 142 |
+
|
| 143 |
+
Table 1: Modifying the modality gap can improve zero-shot performances for downstream tasks. Number indicates top-1 accuracy. Direction indicates that whether increasing (") or decreasing (#) the gap leads to optimal performance.
|
| 144 |
+
|
| 145 |
+
<table><tr><td>Dataset</td><td>Original gap</td><td>Modified gap</td><td>Direction</td></tr><tr><td colspan="4">Coarse-grained Classification</td></tr><tr><td>CIFAR10</td><td>0.9013</td><td>0.9081</td><td>→</td></tr><tr><td>CIFAR100</td><td>0.6658</td><td>0.6737</td><td>↓</td></tr><tr><td colspan="4">Fine-grained Classification</td></tr><tr><td>EuroSAT</td><td>0.5410</td><td>0.5645</td><td>←</td></tr><tr><td colspan="4">Optical Character Recognition</td></tr><tr><td>SVHN</td><td>0.5389</td><td>0.5396</td><td>→</td></tr><tr><td>HatefulMemes</td><td>0.5800</td><td>0.5811</td><td>个</td></tr></table>
|
| 146 |
+
|
| 147 |
+
Table 2: Modifying the modality gap reduces biases for all races. Number indicates the fraction FairFace images whose top-1 prediction is offensive. Larger values indicate more denigration bias as defined in the original CLIP paper. Increasing the gap from 0.82 to 0.97 reduces denigration harms consistently for all races.
|
| 148 |
+
|
| 149 |
+
<table><tr><td rowspan="2">Denigration Biases</td><td colspan="3">Original gap</td><td colspan="2">Modified gap</td></tr><tr><td>Crime related human</td><td>Non</td><td>Sum</td><td>Crime Non related human</td><td>Sum</td></tr><tr><td rowspan="5"></td><td>Black 1.0% 0.1%</td><td></td><td> 1.1%</td><td>0.8% 0.1%</td><td>1.0%</td></tr><tr><td>White 15.5%</td><td>0.2%</td><td>15.7%</td><td>13.2% 0.4%</td><td>13.7%</td></tr><tr><td>Indian 1.2%</td><td>0.0%</td><td>1.2%</td><td>1.1% 0.0%</td><td>1.1%</td></tr><tr><td>Latino 2.8%</td><td>0.1%</td><td>2.8%</td><td>1.9% 0.1%</td><td>2.0%</td></tr><tr><td>Middle Eastern 6.3%</td><td>0.0%</td><td>6.3%</td><td>5.2% 0.0%</td><td>5.2%</td></tr><tr><td>Southeast Asian 0.5%</td><td>0.0%</td><td>0.5%</td><td>0.3%</td><td>0.0%</td><td>0.3%</td></tr><tr><td>East Asian 0.7%</td><td>0.0%</td><td>0.7%</td><td>0.6%</td><td>0.0%</td><td>0.6%</td></tr></table>
|
| 150 |
+
|
| 151 |
+
# 4.4 Initialization vs Optimization
|
| 152 |
+
|
| 153 |
+
Design So far, we have shown that (1) modality gap is born at random initialization, and (2) contrastive learning objective encourages the gap. To explore how the final modality gap is affected by a combination of both factors, we train two CLIP models from scratch: one model uses random initialization, where the gap is large $\| \vec { \Delta } _ { \mathrm { g a p } } \| = 1 . 1 8 9 1 \pm 0 . 0 0 1 7$ because of the cone effect discuss in Sec. 2; another model amends the gap at the initialization by transforming text embeddings to be close to the image embeddings, where the gap is almost zero $\| \vec { \Delta } _ { \mathrm { g a p } } \| = 0 . 0 3 8 8 \pm 0 . 0 3 5 1$ Numbers are mean and $9 5 \%$ confidence interval over three runs with different random seeds. The transformation we applied is a common method to align multilingual word embeddings [31]. More specifically, given image embedding $\mathbf { X }$ and text embedding y, we apply an orthogonal matrix to text embedding $\mathbf { y } ^ { \prime } = W \mathbf { y }$ and compute the multi-modal contrastive loss on $\mathbf { X }$ and $\mathbf { y } ^ { \prime }$ . The orthogonal matrix minimizes the distance between image embeddings and transformed text embeddings: $W =$ arg $\mathrm { m i n } _ { W \in O _ { D } } \| X - Y W \|$ where $X , Y \ \in \ \mathbb { R } ^ { N \times D }$ are image embeddings and text embeddings generated from $N$ image-caption pairs, and $O _ { D }$ is the set of $D$ -dimensional orthogonal matrix.
|
| 154 |
+
|
| 155 |
+
Results We train both models on the MSCOCO Caption training set with batch size 64 and temperature ⌧ = 1100 (i.e., CLIP’s learned temperature). After training, the original model gap changes from $1 . 1 8 \mathrm { \dot { 9 } 1 } \pm 0 . 0 0 1 7$ to $1 . 2 9 9 1 \pm 0 . 0 3 8 9$ , while the amended model gap changes from $0 . 0 3 8 8 \pm 0 . 0 3 5 1$ to $0 . 7 4 5 7 \pm 0 . 0 6 3 3$ . Numbers are $9 5 \%$ confidence interval over three runs with different random seeds. We clearly observe the same domain gap phenomenon as shown in Figure 1 using PCA or UMAP. This experiment shows that the final domain gap is caused by both initialization and optimization. When we ablate the domain gap at the initialization, the loss will still encourage the gap, but the gap distance is only $57 \%$ compared to the model without amending the gap.
|
| 156 |
+
|
| 157 |
+
# 5 Modality Gap Implications
|
| 158 |
+
|
| 159 |
+
# 5.1 Zero-shot performance
|
| 160 |
+
|
| 161 |
+
Design One of the most interesting capabilities for CLIP is its strong zero-shot transferability to a variety of downstream tasks without any supervision. We study whether changing the gap will affect CLIP (ViT-B/16)’s performances on various downstream tasks, including coarse-grained classification (CIFAR10 and CIFAR100), fine-grained classification (EuroSAT [22]), and optical character recognition (SVHN, HatefulMemes [28]). Metric and prompt for each task are shown in Supp. Table 3. Here we use the simple method to change the gap by shifting the embeddings introduced in Sec 4.2. The main objective of our paper is to understand the modality gap phenomenon, a general inductive bias that holds across various data modalities and NN architectures. The goal of our paper is not to propose a method to close the gap and to improve downstream performance.
|
| 162 |
+
|
| 163 |
+
Results Modifying the gap by shifting the embeddings can improve different downstream tasks compared to the original gap without shifting embeddings (Table 1). Details of performance vs gap distance curves are shown in Supp. Figure 10. We leave more methods to change the gap and more analysis of the relation between gap distance and downstream task performance to future work.
|
| 164 |
+
|
| 165 |
+
# 5.2 Fairness
|
| 166 |
+
|
| 167 |
+
Design We follow the bias evaluation setup in the CLIP paper to evaluate denigration harms [39, Sec. 7.1]. We performed zero-shot evaluations on CLIP (ViT-B/32) on the evaluation set of the FairFace dataset [26], which has 10,954 images. In addition to the 14 FairFace classes (e.g., ‘white male’, ‘black female’), we added 4 non-human classes (‘animal’, ‘gorilla’, ‘chimpanzee’ and ‘orangutan’) and 3 crime-related classes (‘thief’, ‘criminal’ and ‘suspicious person’). The text prompts are attached in Appendix (Supp. Figure 11). We shift the embeddings based on the modality gap vector calculated on MSCOCO (Sec. 4.2). We report the fraction FairFace images whose top-1 prediction is offensive.
|
| 168 |
+
|
| 169 |
+
Results We found that increasing the gap from 0.82 to 0.97 reduces denigration harms consistently for all races (Table 5). Meanwhile, we only observe a minor 0.0008 top-1 accuracy drop (Appendix B.2). It is encouraging that a simple gap offsetting approach can lead to a consistent bias reduction across all races on such a complex model (i.e., CLIP)6. Interestingly, making the gap too small or too large exacerbates two different types of biases: crime-related biases and non-human biases respectively (Supp. Table 4).
|
| 170 |
+
|
| 171 |
+
# 6 Related Work
|
| 172 |
+
|
| 173 |
+
Contrastive Representation Learning Contrastive representation learning learns an embedding space where similar objects are closer than dissimilar ones, and has achieved great success in vision [7, 20, 6, 9], language [40, 16], and graph [51, 38]. However, as contrastive learning is still an emerging representation learning technique, we still lack comprehensive theoretical and empirical understandings about why contrastive learning works. [48] proposed two ideal objectives for contrastive representation space: alignment (similar samples have similar features) and uniformity (features are uniformly distributed on the hypersphere), and demonstrated these two objectives are highly correlated with downstream task performances. [46] show that low temperatures increase the model’s penalty on hard negative examples, and thus increase uniformity and decrease tolerance (the closeness of semantically similar samples). These analyses mostly focus on unsupervised contrastive learning on a single modality. Orthogonal to their work, we show that multi-modal contrastive learning with low temperatures and mismatched data encourages the modality gap.
|
| 174 |
+
|
| 175 |
+
Multi-modal Contrastive Representation Learning Multi-modal models map inputs from different data modalities (e.g. image and text) into a shared representation space [53, 50, 34, 24, 11]. It has garnered tremendous interest and excitement as a framework for data integration. These models are often pre-trained with contrastive loss [45], as [39] showed that the contrastive learning is $1 2 \times$ more efficient than the generative approaches. We demonstrate an intriguing geometric phenomenon of the representation space of these multi-modal models, and provide a three-part explanation supported by theory and experiments. The idea of mapping images and text into a shared embedding space has been explored in earlier works [42, 49]. There have been recent efforts in formulating images and text embeddings as metric learning [14], multilabel classification [25], n-gram language learning [32], and captioning [10]. Recently there has there has also been work in using a unified encoder to fuse different data modalities [19]. Research into how the modality gap phenomenon generalizes to the multi-modal representations obtained by these alternative methods, or even uni-modal settings with teacher and student model [44, 5] would be a promising direction for future work.
|
| 176 |
+
|
| 177 |
+
Cone Effect Our analyses also provide new insights on the cone effect, which we show is a general phenomenon for deep neural networks. Existing work focuses on the language representations of trained language models such as BERT and GPT-2 [12, 15, 33]. Given that isotropy has both theoretical and empirical benefits for static embeddings [35], the extent of anisotropy in contextualized representations is surprising [12]. It has been shown that the cone effect limits the expressiveness of the language representations. Post-processing methods [33, 43, 2, 35] or modified training objective [15, 47, 16] alleviate the cone effect and improve downstream performance. Existing work attributes the cone effect to the optimization under unbalanced word frequencies distribution [15, 33]. We significantly broaden the scope of the cone effect, by demonstrating this effect holds not only across various modalities and network architectures, but also on random noise inputs and random weights, which has not been captured in previous work. We mathematically characterize the contraction mapping induced by linear layers with ReLU non-linearities to explain the cone effect. Our theory matches well with experiments and provides insights for understanding the general inductive biases of deep neural networks.
|
| 178 |
+
|
| 179 |
+
# 7 Discussion
|
| 180 |
+
|
| 181 |
+
In this work, we investigated an interesting phenomenon in multi-modal contrastive learning — modality gap. We analyzed why the gap exists, i.e., the joint effect of model initialization and optimization, and why studying the gap is important, i.e., it can affect the downstream task performance and fairness. Our work raises several basic questions about representation learning, contrastive learning, and multi-modal contrastive representation learning. For representation learning, prior research in NLP has shown that alleviating the cone effect improves downstream performance. As our work significantly broadens the scope of the cone effect, methods for alleviating the cone effect in other modalities to improve ML performance is an interesting direction of future research.
|
| 182 |
+
|
| 183 |
+
For contrastive learning, our embedding shifting, simulation, and fine-tuning experiments all show that the contrast loss landscape is heavily influenced by temperature. Recent work has found that temperature directly controls the uniformity and affinity of the uni-modal representation space [46]. Our study provides a complementary understanding of the multi-modal representation space. Development of geometric methods for evaluation of representations [37, 30] to further capture the geometric landscape of the modality gap is an interesting direction of future work.
|
| 184 |
+
|
| 185 |
+
For multi-modal contrastive representational learning, we find that changing the modal gap can affect performance and fairness on downstream tasks. Interestingly, having larger gap can help some fairness and zero-shot learning applications. The main objective of our paper is to demonstrate the modality gap phenomenon and explain contraction mapping contribute to this. Systematic analysis of the impact of the gap on applications is an important direction of future work.
|
| 186 |
+
|
| 187 |
+
# Reproducibility Statement
|
| 188 |
+
|
| 189 |
+
We provide open-source implementation of our work at https://github.com/Weixin-Liang/ Modality-Gap. The implementations will enable researchers to reproduce the modality gap described here as well as run their own analyses on additional cross-modal models. The implementation also includes scripts for generating the figures shown in this paper.
|
| 190 |
+
|
| 191 |
+
# References
|
| 192 |
+
|
| 193 |
+
[1] Z. Allen-Zhu, Y. Li, and Z. Song. A convergence theory for deep learning via overparameterization. In ICML, 2019.
|
| 194 |
+
[2] S. Arora, Y. Liang, and T. Ma. A simple but tough-to-beat baseline for sentence embeddings. In ICLR, 2017.
|
| 195 |
+
[3] D. Arpit, S. Jastrzebski, N. Ballas, D. Krueger, E. Bengio, M. S. Kanwal, T. Maharaj, A. Fischer, A. C. Courville, Y. Bengio, and S. Lacoste-Julien. A closer look at memorization in deep networks. In ICML, volume 70 of Proceedings of Machine Learning Research, pages 233–242. PMLR, 2017.
|
| 196 |
+
[4] J. L. Ba, J. R. Kiros, and G. E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016.
|
| 197 |
+
[5] L. Beyer, X. Zhai, A. Royer, L. Markeeva, R. Anil, and A. Kolesnikov. Knowledge distillation: A good teacher is patient and consistent. In CVPR, 2022.
|
| 198 |
+
[6] M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.
|
| 199 |
+
[7] T. Chen, S. Kornblith, M. Norouzi, and G. E. Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.
|
| 200 |
+
[8] X. Chen, H. Fang, T. Lin, R. Vedantam, S. Gupta, P. Dollár, and C. L. Zitnick. Microsoft COCO captions: Data collection and evaluation server. CoRR, abs/1504.00325, 2015.
|
| 201 |
+
[9] X. Chen and K. He. Exploring simple siamese representation learning. In CVPR, 2021.
|
| 202 |
+
[10] K. Desai and J. Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, 2021.
|
| 203 |
+
[11] CLASP: Contrastive Language Aminoacid Sequence Pretraining, 2021.
|
| 204 |
+
[12] K. Ethayarajh. How contextual are contextualized word representations? comparing the geometry of bert, elmo, and GPT-2 embeddings. In EMNLP, 2019.
|
| 205 |
+
[13] J. Frankle and M. Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In ICLR. OpenReview.net, 2019.
|
| 206 |
+
[14] A. Frome, G. S. Corrado, J. Shlens, S. Bengio, J. Dean, M. Ranzato, and T. Mikolov. Devise: A deep visual-semantic embedding model. In NIPS, 2013.
|
| 207 |
+
[15] J. Gao, D. He, X. Tan, T. Qin, L. Wang, and T. Liu. Representation degeneration problem in training natural language generation models. In ICLR, 2019.
|
| 208 |
+
[16] T. Gao, X. Yao, and D. Chen. Simcse: Simple contrastive learning of sentence embeddings. In EMNLP, 2021.
|
| 209 |
+
[17] R. Geirhos, J. Jacobsen, C. Michaelis, R. S. Zemel, W. Brendel, M. Bethge, and F. A. Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2(11):665–673, 2020.
|
| 210 |
+
[18] R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel. Imagenettrained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR. OpenReview.net, 2019.
|
| 211 |
+
[19] R. Girdhar, M. Singh, N. Ravi, L. van der Maaten, A. Joulin, and I. Misra. Omnivore: A single model for many visual modalities. CoRR, abs/2201.08377, 2022.
|
| 212 |
+
[20] J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020.
|
| 213 |
+
[21] W. Guo, J. Wang, and S. Wang. Deep multimodal representation learning: A survey. IEEE Access, 7:63373–63394, 2019.
|
| 214 |
+
[22] P. Helber, B. Bischke, A. Dengel, and D. Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019.
|
| 215 |
+
[23] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015.
|
| 216 |
+
[24] C. Jia, Y. Yang, Y. Xia, Y. Chen, Z. Parekh, H. Pham, Q. V. Le, Y. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.
|
| 217 |
+
[25] A. Joulin, L. van der Maaten, A. Jabri, and N. Vasilache. Learning visual features from large weakly supervised data. In ECCV, 2016.
|
| 218 |
+
[26] K. Kärkkäinen and J. Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In WACV, 2021.
|
| 219 |
+
[27] N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang. On large-batch training for deep learning: Generalization gap and sharp minima. In ICLR. OpenReview.net, 2017.
|
| 220 |
+
[28] D. Kiela, H. Firooz, A. Mohan, V. Goswami, A. Singh, P. Ringshia, and D. Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. In NeurIPS, 2020.
|
| 221 |
+
[29] B. Kim, E. Reif, M. Wattenberg, S. Bengio, and M. C. Mozer. Neural networks trained on natural scenes exhibit gestalt closure. Computational Brain & Behavior, 4(3):251–263, 2021.
|
| 222 |
+
[30] T. Kynkäänniemi, T. Karras, S. Laine, J. Lehtinen, and T. Aila. Improved precision and recall metric for assessing generative models. In NeurIPS, 2019.
|
| 223 |
+
[31] G. Lample, A. Conneau, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel data. In ICLR, 2018.
|
| 224 |
+
[32] A. Li, A. Jabri, A. Joulin, and L. van der Maaten. Learning visual n-grams from web data. In ICCV, 2017.
|
| 225 |
+
[33] B. Li, H. Zhou, J. He, M. Wang, Y. Yang, and L. Li. On the sentence embeddings from pre-trained language models. In EMNLP, 2020.
|
| 226 |
+
[34] J. Li, R. R. Selvaraju, A. D. Gotmare, S. R. Joty, C. Xiong, and S. C. H. Hoi. Align before fuse: Vision and language representation learning with momentum distillation. CoRR, abs/2107.07651, 2021.
|
| 227 |
+
[35] J. Mu and P. Viswanath. All-but-the-top: Simple and effective postprocessing for word representations. In ICLR, 2018.
|
| 228 |
+
[36] B. Neyshabur, Z. Li, S. Bhojanapalli, Y. LeCun, and N. Srebro. The role of over-parametrization in generalization of neural networks. In ICLR, 2019.
|
| 229 |
+
[37] P. Poklukar, V. Polianskii, A. Varava, F. T. Pokorny, and D. K. Jensfelt. Delaunay component analysis for evaluation of data representations. In ICLR, 2022.
|
| 230 |
+
[38] J. Qiu, Q. Chen, Y. Dong, J. Zhang, H. Yang, M. Ding, K. Wang, and J. Tang. Gcc: Graph contrastive coding for graph neural network pre-training. In KDD, 2020.
|
| 231 |
+
[39] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021.
|
| 232 |
+
[40] N. Reimers, I. Gurevych, N. Reimers, I. Gurevych, N. Thakur, N. Reimers, J. Daxenberger, I. Gurevych, N. Reimers, I. Gurevych, et al. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP, 2019.
|
| 233 |
+
[41] T. Sainburg, L. McInnes, and T. Q. Gentner. Parametric umap embeddings for representation and semisupervised learning. Neural Computation, 2021.
|
| 234 |
+
[42] R. Socher and L. Fei-Fei. Connecting modalities: Semi-supervised segmentation and annotation of images using unaligned text corpora. In CVPR, 2010.
|
| 235 |
+
[43] J. Su, J. Cao, W. Liu, and Y. Ou. Whitening sentence representations for better semantics and faster retrieval. CoRR, abs/2103.15316, 2021.
|
| 236 |
+
[44] A. Tarvainen and H. Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In NIPS, 2017.
|
| 237 |
+
[45] A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018.
|
| 238 |
+
[46] F. Wang and H. Liu. Understanding the behaviour of contrastive loss. In CVPR, 2021.
|
| 239 |
+
[47] L. Wang, J. Huang, K. Huang, Z. Hu, G. Wang, and Q. Gu. Improving neural language generation with spectrum control. In ICLR, 2020.
|
| 240 |
+
[48] T. Wang and P. Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In ICML, 2020.
|
| 241 |
+
[49] J. Weston, S. Bengio, and N. Usunier. Large scale image annotation: learning to rank with joint word-image embeddings. Machine learning, 2010.
|
| 242 |
+
[50] H. Xu, G. Ghosh, P. Huang, D. Okhonko, A. Aghajanyan, F. Metze, L. Zettlemoyer, and C. Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In EMNLP, 2021.
|
| 243 |
+
[51] Y. You, T. Chen, Y. Sui, T. Chen, Z. Wang, and Y. Shen. Graph contrastive learning with augmentations. In NeurIPS, 2020.
|
| 244 |
+
[52] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals. Understanding deep learning (still) requires rethinking generalization. Commun. ACM, 64(3):107–115, 2021.
|
| 245 |
+
[53] Y. Zhang, H. Jiang, Y. Miura, C. D. Manning, and C. P. Langlotz. Contrastive learning of medical visual representations from paired images and text. CoRR, abs/2010.00747, 2020.
|
| 246 |
+
|
| 247 |
+
# Checklist
|
| 248 |
+
|
| 249 |
+
1. For all authors...
|
| 250 |
+
|
| 251 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 252 |
+
(b) Did you describe the limitations of your work? [Yes]
|
| 253 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes]
|
| 254 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 255 |
+
|
| 256 |
+
2. If you are including theoretical results...
|
| 257 |
+
|
| 258 |
+
(a) Did you state the full set of assumptions of all theoretical results? [Yes] (b) Did you include complete proofs of all theoretical results? [Yes]
|
| 259 |
+
|
| 260 |
+
3. If you ran experiments...
|
| 261 |
+
|
| 262 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes]
|
| 263 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes]
|
| 264 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [Yes]
|
| 265 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes]
|
| 266 |
+
|
| 267 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 268 |
+
|
| 269 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 270 |
+
(b) Did you mention the license of the assets? [Yes]
|
| 271 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]
|
| 272 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [Yes]
|
| 273 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes]
|
| 274 |
+
|
| 275 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 276 |
+
|
| 277 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 278 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 279 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/S7Evzt9uit3/S7Evzt9uit3_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu.md
ADDED
|
@@ -0,0 +1,361 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# HUMAN MOTION DIFFUSION MODEL
|
| 2 |
+
|
| 3 |
+
Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or and Amit H. Bermano
|
| 4 |
+
|
| 5 |
+
Tel Aviv University, Israel guytevet@mail.tau.ac.il
|
| 6 |
+
|
| 7 |
+
# ABSTRACT
|
| 8 |
+
|
| 9 |
+
Natural and expressive human motion generation is the holy grail of computer animation. It is a challenging task, due to the diversity of possible motion, human perceptual sensitivity to it, and the difficulty of accurately describing it. Therefore, current generative solutions are either low-quality or limited in expressiveness. Diffusion models, which have already shown remarkable generative capabilities in other domains, are promising candidates for human motion due to their many-to-many nature, but they tend to be resource hungry and hard to control. In this paper, we introduce Motion Diffusion Model (MDM), a carefully adapted classifier-free diffusion-based generative model for the human motion domain. MDM is transformer-based, combining insights from motion generation literature. A notable design-choice is the prediction of the sample, rather than the noise, in each diffusion step. This facilitates the use of established geometric losses on the locations and velocities of the motion, such as the foot contact loss. As we demonstrate, MDM is a generic approach, enabling different modes of conditioning, and different generation tasks. We show that our model is trained with lightweight resources and yet achieves state-ofthe-art results on leading benchmarks for text-to-motion and action-to-motion 1. https://guytevet.github.io/mdm-page/.
|
| 10 |
+
|
| 11 |
+
# 1 INTRODUCTION
|
| 12 |
+
|
| 13 |
+
Human motion generation is a fundamental task in computer animation, with applications spanning from gaming to robotics. It is a challenging field, due to several reasons, including the vast span of possible motions, and the difficulty and cost of acquiring high quality data. For the recently emerging text-to-motion setting, where motion is generated from natural language, another inherent problem is data labeling. For example, the label ”kick” could refer to a soccer kick, as well as a Karate one. At the same time, given a specific kick there are many ways to describe it, from how it is performed to the emotions it conveys, constituting a many-to-many problem. Current approaches have shown success in the field, demonstrating plausible mapping from text to motion (Petrovich et al., 2022; Tevet et al., 2022; Ahuja & Morency, 2019). All these approaches, however, still limit the learned distribution since they mainly employ auto-encoders or VAEs (Kingma & Welling, 2013) (implying a one-to-one mapping or a normal latent distribution respectively). In this aspect, diffusion models are a better candidate for human motion generation, as they are free from assumptions on the target distribution, and are known for expressing well the many-to-many distribution matching problem we have described.
|
| 14 |
+
|
| 15 |
+
Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020; Ho et al., 2020) are a generative approach that is gaining significant attention in the computer vision and graphics community. When trained for conditioned generation, recent diffusion models (Ramesh et al., 2022; Saharia et al., 2022b) have shown breakthroughs in terms of image quality and semantics. The competence of these models have also been shown for other domains, including videos (Ho et al., 2022), and 3D point clouds (Luo & Hu, 2021). The problem with such models, however, is that they are notoriously resource demanding and challenging to control.
|
| 16 |
+
|
| 17 |
+
In this paper, we introduce Motion Diffusion Model (MDM) — a carefully adapted diffusion based generative model for the human motion domain. Being diffusion-based, MDM gains from the na“A man runs to the right then runs to the left then back to the middle.”
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
Figure 1: Our Motion Diffusion Model (MDM) reflects the many-to-many nature of text-to-motion mapping by generating diverse motions given a text prompt. Our custom architecture and geometric losses help yielding high-quality motion. Darker color indicates later frames in the sequence.
|
| 23 |
+
|
| 24 |
+
tive aforementioned many-to-many expression of the domain, as evidenced by the resulting motion quality and diversity (Figure 1). In addition, MDM combines insights already well established in the motion generation domain, helping it be significantly more lightweight and controllable.
|
| 25 |
+
|
| 26 |
+
First, instead of the ubiquitous U-net (Ronneberger et al., 2015) backbone, MDM is transformerbased. As we demonstrate, our architecture (Figure 2) is lightweight and better fits the temporal and spatially irregular nature of motion data (represented as a collection of joints). A large volume of motion generation research is devoted to learning using geometric losses (Kocabas et al., 2020; Harvey et al., 2020; Aberman et al., 2020). Some, for example, regulate the velocity of the motion (Petrovich et al., 2021) to prevent jitter, or specifically consider foot sliding using dedicated terms (Shi et al., 2020). Consistently with these works, we show that applying geometric losses in the diffusion setting improves generation.
|
| 27 |
+
|
| 28 |
+
The MDM framework has a generic design enabling different forms of conditioning. We showcase three tasks: text-to-motion, action-to-motion, and unconditioned generation. We train the model in a classifier-free manner (Ho & Salimans, 2022), which enables trading-off diversity to fidelity, and sampling both conditionally and unconditionally from the same model. In the text-to-motion task, our model generates coherent motions (Figure 1) that achieve state-of-the-art results on the HumanML3D (Guo et al., 2022a) and KIT (Plappert et al., 2016) benchmarks. Moreover, our user study shows that human evaluators prefer our generated motions over real motions $4 2 \%$ of the time (Figure 4(a)). In action-to-motion, MDM outperforms the state-of-the-art (Guo et al., 2020; Petrovich et al., 2021), even though they were specifically designed for this task, on the common HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018) benchmarks.
|
| 29 |
+
|
| 30 |
+
Lastly, we also demonstrate completion and editing. By adapting diffusion image-inpainting (Song et al., 2020b; Saharia et al., 2022a), we set a motion prefix and suffix, and use our model to fill in the gap. Doing so under a textual condition guides MDM to fill the gap with a specific motion that still maintains the semantics of the original input. By performing inpainting in the joints space rather than temporally, we also demonstrate the semantic editing of specific body parts, without changing the others (Figure 3).
|
| 31 |
+
|
| 32 |
+
Overall, we introduce Motion Diffusion Model, a motion framework that achieves state-of-the-art quality in several motion generation tasks, while requiring only about three days of training on a single mid-range GPU. It supports geometric losses, which are non trivial to the diffusion setting, but are crucial to the motion domain, and offers the combination of state-of-the-art generative power with well thought-out domain knowledge.
|
| 33 |
+
|
| 34 |
+
# 2 RELATED WORK
|
| 35 |
+
|
| 36 |
+
# 2.1 HUMAN MOTION GENERATION
|
| 37 |
+
|
| 38 |
+
Neural motion generation, learned from motion capture data, can be conditioned by any signal that describes the motion. Many works use parts of the motion itself for guidance. Some predict motion from its prefix poses (Fragkiadaki et al., 2015; Martinez et al., 2017; Hernandez et al., 2019; Guo et al., 2022b). Others (Harvey & Pal, 2018; Kaufmann et al., 2020; Harvey et al., 2020; Duan et al., 2021) solve in-betweening and super-resolution tasks using bi-directional GRU (Cho et al., 2014) and Transformer (Vaswani et al., 2017) architectures. Holden et al. (2016) use auto-encoder to learn motion latent representation, then utilize it to edit and control motion with spatial constraints such as root trajectory and bone lengths. Motion can be controlled with a high-level guidance given from action class (Guo et al., 2020; Petrovich et al., 2021; Cervantes et al., 2022), audio (Li et al., 2021; Aristidou et al., 2022) and natural language (Ahuja & Morency, 2019; Petrovich et al., 2022). In most cases authors suggests a dedicated approach to map each conditioning domain into motion.
|
| 39 |
+
|
| 40 |
+
In recent years, the leading approach for the Text-to-Motion task is to learn a shared latent space for language and motion. JL2P (Ahuja & Morency, 2019) learns the KIT motion-language dataset (Plappert et al., 2016) with an auto-encoder, limiting one-to-one mapping from text to motion. TEMOS (Petrovich et al., 2022) and T2M (Guo et al., 2022a) suggest using a VAE (Kingma & Welling, 2013) to map a text prompt into a normal distribution in latent space. Recently, MotionCLIP (Tevet et al., 2022) leverages the shared text-image latent space learned by CLIP (Radford et al., 2021) to expand text-to-motion out of the data limitations and enabled latent space editing.
|
| 41 |
+
|
| 42 |
+
The human motion manifold can also be learned without labels, as shown by Holden et al. (2016), V-Poser (Pavlakos et al., 2019), and more recently the dedicated MoDi architecture (Raab et al., 2022). We show that our model is capable for such an unsupervised setting as well.
|
| 43 |
+
|
| 44 |
+
# 2.2 DIFFUSION GENERATIVE MODELS
|
| 45 |
+
|
| 46 |
+
Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020) are a class of neural generative models, based on the stochastic diffusion process as it is modeled in Thermodynamics. In this setting, a sample from the data distribution is gradually noised by the diffusion process. Then, a neural model learns the reverse process of gradually denoising the sample. Sampling the learned data distribution is done by denoising a pure initial noise. Ho et al. (2020) and Song et al. (2020a) further developed the practices for image generation applications. For conditioned generation, Dhariwal & Nichol (2021), introduced classifier-guided diffusion, which was later on adapted by GLIDE (Nichol et al., 2021) to enable conditioning over CLIP textual representations. The Classifier-Free Guidance approach Ho & Salimans (2022) enables conditioning while trading-off fidelity and diversity, and achieves better results (Nichol et al., 2021). In this paper, we implement text-to-motion by conditioning on CLIP in a classifier-free manner, similarly to text-to-image (Ramesh et al., 2022; Saharia et al., 2022b). Local editing of images is typically defined as an inpainting problem, where a part of the image is constant, and the inpainted part is denoised by the model, possibly under some condition (Song et al., 2020b; Saharia et al., 2022a). We adapt this technique to edit motion’s specific body parts or temporal intervals (in-betweening) according to an optional condition.
|
| 47 |
+
|
| 48 |
+
Closer to our context, Gu et al. (2022) used the diffusion formulation to model the stochasticity of human trajectory prediction. More recently, concurrent to this work, Zhang et al. (2022) and Kim et al. (2022) have suggested diffusion models for motion generation. Our work requires significantly fewer GPU resources and makes design choices that enable geometric losses, which improve results.
|
| 49 |
+
|
| 50 |
+
# 3 MOTION DIFFUSION MODEL
|
| 51 |
+
|
| 52 |
+
An overview of our method is described in Figure 2. Our goal is to synthesize a human motion $x ^ { 1 : N }$ of length $N$ given an arbitrary condition $c$ . This condition can be any real-world signal that will dictate the synthesis, such as audio (Li et al., 2021; Aristidou et al., 2022), natural language (text-to-motion) (Tevet et al., 2022; Guo et al., 2022a) or a discrete class (action-to-motion) (Guo et al., 2020; Petrovich et al., 2021). Iwhich we denote as the null condition $c = \emptyset$ tion, unconditioned mo. The generated motion $\boldsymbol { x } ^ { 1 : N } = \{ x ^ { i } \} _ { i = 1 } ^ { N }$ s also possible,is a sequences of human poses represented by either joint rotations or positions $\boldsymbol { x } ^ { i } \in \mathbb { R } ^ { J \times D }$ , where $J$ is the number of joints and $D$ is the dimension of the joint representation. MDM can accept motion represented by either locations, rotations, or both (see Section 4).
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
Figure 2: (Left) Motion Diffusion Model (MDM) overview. The model is fed a motion sequence $x _ { t } ^ { 1 : N }$ of length $N$ in a noising step $t$ , as well as $t$ itself and a conditioning code c. c, a CLIP (Radford et al., 2021) based textual embedding in this case, is first randomly masked for classifier-free learning and then projected together with $t$ into the input token $z _ { t k }$ . In each sampling step, the transformerencoder predicts the final clean motion $\hat { x } _ { 0 } ^ { 1 : N }$ . (Right) Sampling MDM. Given a condition $c$ , we sample random noise $x _ { T }$ at the dimensions of the desired motion, then iterate from $T$ to 1. At each step $t$ , MDM predicts the clean sample ${ \hat { x } } _ { 0 }$ , and diffuses it back to $x _ { t - 1 }$ .
|
| 56 |
+
|
| 57 |
+
Framework. Diffusion is modeled as a Markov noising process, $\{ x _ { t } ^ { 1 : N } \} _ { t = 0 } ^ { T }$ , where $x _ { 0 } ^ { 1 : N }$ is drawn from the data distribution and
|
| 58 |
+
|
| 59 |
+
$$
|
| 60 |
+
q ( x _ { t } ^ { 1 : N } | x _ { t - 1 } ^ { 1 : N } ) = \mathcal { N } ( \sqrt { \alpha _ { t } } x _ { t - 1 } ^ { 1 : N } , ( 1 - \alpha _ { t } ) I ) ,
|
| 61 |
+
$$
|
| 62 |
+
|
| 63 |
+
where $\alpha _ { t } \in ( 0 , 1 )$ are constant hyper-parameters. When $\alpha _ { t }$ is small enough, we can approximate $x _ { T } ^ { 1 : N } \sim \mathcal { N } ( 0 , I )$ . From here on we use $x _ { t }$ to denote the full sequence at noising step $t$ .
|
| 64 |
+
|
| 65 |
+
In our context, conditioned motion synthesis models the distribution $p ( x _ { 0 } | c )$ as the reversed diffusion process of gradually cleaning $x _ { T }$ . Instead of predicting $\epsilon _ { t }$ as formulated by Ho et al. (2020), we follow Ramesh et al. (2022) and use an equivalent formulation to predict the signal itself, i.e., $\hat { x } _ { 0 } = G ( x _ { t } , t , c )$ with the simple objective (Ho et al., 2020),
|
| 66 |
+
|
| 67 |
+
$$
|
| 68 |
+
{ \mathcal { L } } _ { \mathrm { s i m p l e } } = E _ { x _ { 0 } \sim q ( x _ { 0 } | c ) , t \sim [ 1 , T ] } [ \| x _ { 0 } - G ( x _ { t } , t , c ) \| _ { 2 } ^ { 2 } ]
|
| 69 |
+
$$
|
| 70 |
+
|
| 71 |
+
Geometric losses. In the motion domain, generative networks are standardly regularized using geometric losses Petrovich et al. (2021); Shi et al. (2020). These losses enforce physical properties and prevent artifacts, encouraging natural and coherent motion. In this work we experiment with three common geometric losses that regulate (1) positions (in case we predict rotations), (2) foot contact, and (3) velocities.
|
| 72 |
+
|
| 73 |
+
$$
|
| 74 |
+
\mathcal { L } _ { \mathrm { p o s } } = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \| F K ( x _ { 0 } ^ { i } ) - F K ( \hat { x } _ { 0 } ^ { i } ) \| _ { 2 } ^ { 2 } ,
|
| 75 |
+
$$
|
| 76 |
+
|
| 77 |
+
$$
|
| 78 |
+
\mathcal { L } _ { \mathrm { f o o t } } = \frac { 1 } { N - 1 } \sum _ { i = 1 } ^ { N - 1 } \| ( F K ( \hat { x } _ { 0 } ^ { i + 1 } ) - F K ( \hat { x } _ { 0 } ^ { i } ) ) \cdot f _ { i } \| _ { 2 } ^ { 2 } ,
|
| 79 |
+
$$
|
| 80 |
+
|
| 81 |
+
$$
|
| 82 |
+
\mathcal { L } _ { \mathrm { v e l } } = \frac { 1 } { N - 1 } \sum _ { i = 1 } ^ { N - 1 } \| ( x _ { 0 } ^ { i + 1 } - x _ { 0 } ^ { i } ) - ( \hat { x } _ { 0 } ^ { i + 1 } - \hat { x } _ { 0 } ^ { i } ) \| _ { 2 } ^ { 2 }
|
| 83 |
+
$$
|
| 84 |
+
|
| 85 |
+
In case we predict joint rotations, $F K ( \cdot )$ denotes the forward kinematic function converting joint rotations into joint positions (otherwise, it denotes the identity function). $f _ { i } \in \{ 0 , 1 \} ^ { J }$ is the binary foot contact mask for each frame $i$ . Relevant only to feet, it indicates whether they touch the ground, and are set according to binary ground truth data (Shi et al., 2020). In essence, it mitigates the foot-sliding effect by nullifying velocities when touching the ground. Overall, our training loss is ${ \mathcal { L } } = { \mathcal { L } } _ { \mathrm { s i m p l e } } + \lambda _ { \mathrm { p o s } } { \bar { \mathcal { L } } } _ { \mathrm { p o s } } + { \bar { \lambda _ { \mathrm { v e l } } } } { \bar { \mathcal { L } } } _ { \mathrm { v e l } } + \lambda _ { \mathrm { f o o t } } { \mathcal { L } } _ { \mathrm { f o o t } }$ .
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
Figure 3: Editing applications. Light blue frames represent motion input and bronze frames are the generated motion. Motion in-betweening (left+center) can be performed conditioned on text or without condition by the same model. Specific body part editing using text is demonstrated on the right: the lower body joints are fixed to the input motion while the upper body is altered to fit the input text prompt.
|
| 89 |
+
|
| 90 |
+
Model. Our model is illustrated in Figure 2. We implement $G$ with a straightforward transformer (Vaswani et al., 2017) encoder-only architecture. The transformer architecture is temporally aware, enabling learning and generating variable-length motions, and is well-proven for the motion domain (Petrovich et al., 2021; Duan et al., 2021; Aksan et al., 2021). The noise time-step $t$ and the condition code $c$ are each projected to the transformer dimension by separate feed-forward networks, then summed to yield the token $z _ { t k }$ . Each frame of the noised input $x _ { t }$ is linearly projected into the transformer dimension and summed with a standard positional embedding. $z _ { t k }$ and the projected frames are then fed to the encoder. Excluding the first output token (corresponding to $z _ { t k }$ ), the encoder result is projected back to the original motion dimensions, and serves as the prediction $\scriptstyle { \hat { x } } _ { 0 }$ . We implement text-to-motion by encoding the text prompt to $c$ with CLIP (Radford et al., 2021) text encoder, and action-to-motion with learned embeddings per class.
|
| 91 |
+
|
| 92 |
+
Sampling from $p ( x _ { 0 } | c )$ is done in an iterative manner, according to Ho et al. (2020). In every time step $t$ we predict the clean sample $\hat { \boldsymbol { x } } _ { 0 } = \boldsymbol { G } ( \boldsymbol { x } _ { t } , t , \boldsymbol { c } )$ and noise it back to $x _ { t - 1 }$ . This is repeated from $t = T$ until $x _ { 0 }$ is achieved (Figure 2 right). We train our model $G$ using classifier-free guidance (Ho & Salimans, 2022). In practice, $G$ learns both the conditioned and the unconditioned distributions by randomly setting $c = \emptyset$ for $1 0 \%$ of the samples, such that $G ( x _ { t } , t , \emptyset )$ approximates $p ( x _ { 0 } )$ . Then, when sampling $G$ we can trade-off diversity and fidelity by interpolating or even extrapolating the two variants using $\boldsymbol { s } \colon G _ { s } ( x _ { t } , t , c ) = G ( x _ { t } , \dot { t , } \emptyset ) + s \cdot ( \dot { G } ( x _ { t } , t , c ) \bar { - } G ( x _ { t } , t , \emptyset ) )$ .
|
| 93 |
+
|
| 94 |
+
Editing. We enable motion in-betweening in the temporal domain, and body part editing in the spatial domain, by adapting diffusion inpainting to motion data. Editing is done only during sampling, without any training involved. Given a subset of the motion sequence inputs, when sampling the model (Figure 2 right), at each iteration we overwrite $\scriptstyle { \hat { x } } _ { 0 }$ with the input part of the motion. This encourages the generation to remain coherent to original input, while completing the missing parts. In the temporal setting, the prefix and suffix frames of the motion sequence are the input, and we solve a motion in-betweening problem (Harvey et al., 2020). Editing can be done either conditionally or unconditionally (by setting $c = \emptyset$ ). In the spatial setting, we show that body parts can be re-synthesized according to a condition $c$ while keeping the rest intact, through the use of the same completion technique.
|
| 95 |
+
|
| 96 |
+
# 4 EXPERIMENTS
|
| 97 |
+
|
| 98 |
+
We implement MDM for three motion generation tasks: Text-to-Motion(4.1), Action-to-Motion(4.2) and unconditioned generation(5.2. Each sub-section reviews the data and metrics of the used benchmarks, provides implementation details, and presents qualitative and quantitative results. Then, we show implementations of motion in-betweening (both conditioned and unconditioned) and bodypart editing by adapting diffusion inpainting to motion (5.1). Our models have been trained with $T = 1 0 0 0$ noising steps and a cosine noise schedule. In Appendix A.1, we experiment with different values of $T$ . All of them have been trained on a single NVIDIA GeForce RTX 2080 Ti GPU for a period of about 3 days.
|
| 99 |
+
|
| 100 |
+
# 4.1 TEXT-TO-MOTION
|
| 101 |
+
|
| 102 |
+
Text-to-motion is the task of generating motion given an input text prompt. The output motion is expected to be both implementing the textual description, and a valid sample from the data distribution (i.e. adhering to general human abilities and the rules of physics). In addition, for each text prompt, we also expect a distribution of motions matching it, rather than just a single result. We evaluate our model using two leading benchmarks - KIT (Plappert et al., 2016) and HumanML3D (Guo et al., 2022a), over the set of metrics suggested by Guo et al. (2022a): R-precision and Multimodal-Dist measure the relevancy of the generated motions to the input prompts, $F I D$ measures the dissimilarity between the generated and ground truth distributions (in latent space), Diversity measures the variability in the resulting motion distribution, and MultiModality is the average variance given a single text prompt. For the full implementation of the metrics, please refer to Guo et al. (2022a). We use HumanML3D as a platform to compare different backbones of our model, discovering that the diffusion framework is relatively agnostic to this attribute. In addition, we conduct a user study comparing our model to current art and ground truth motions.
|
| 103 |
+
|
| 104 |
+
Data. HumanML3D is a recent dataset, textually re-annotating motion capture from the AMASS (Mahmood et al., 2019) and HumanAct12 (Guo et al., 2020) collections. It contains 14, 616 motions annotated by 44, 970 textual descriptions. In addition, it suggests a redundant data representation including a concatenation of root velocity, joint positions, joint velocities, joint rotations and the foot contact binary labels. We also use in this section the same representation for the KIT dataset, brought by the same publishers. Although limited in the number (3, 911) and the diversity of samples, most of the text-to-motion research is based on KIT, hence we view it as important to evaluate using it as well.
|
| 105 |
+
|
| 106 |
+
Implementation. In addition to our Transformer encoder-only backbone (Section 3), we experiment MDM with three more backbones: (1) Transformer decoder injects $z _ { t k }$ through the cross-attention layer, instead of as an input token. (2) Transformer decoder $^ +$ input token, where $z _ { t k }$ is injected both ways, (3) GRU (Cho et al., 2014) concatenate $z _ { t k }$ to each input frame , and (4) a U-net adaptation for motion data (Table 1). U-net is adapted to motion by replacing the 2D convolution filters with 1D convolutions in the temporal axis, such that joints are considered as channels. This is due to the irregular behavior of the joint axis. Our models were trained with batch size 64, 8 layers (except GRU that was optimal at 2), and latent dimension 512. To encode the text we use a frozen CLIPViT-B/32 model. In addition, we experiment with replacing CLIP with sentence-BERT (Reimers & Gurevych, 2019) - a BERT-based (Devlin et al., 2019) text encoder. The full details can be found in our published code1 and Appendix C. Each model was trained for $5 0 0 K$ steps, after which a checkpoint was chosen that minimizes the FID metric to be reported. Since foot contact and joint locations are explicitly represented in HumanML3D, we don’t apply geometric losses in this section. We evaluate our models with guidance-scale $s = 2 . 5$ which provides a diversity-fidelity sweet spot (Figure 4).
|
| 107 |
+
|
| 108 |
+
Quantitative evaluation. We evaluate and compare our models to current art (JL2P Ahuja & Morency (2019), Text2Gesture (Bhattacharya et al., 2021), and T2M (Guo et al., 2022a)) with the metrics suggested by Guo et al. (2022a). As can be seen, MDM achieves state-of-the-art results in FID, Diversity, and MultiModality, indicating high diversity per input text prompt, and high-quality samples, as can also be seen qualitatively in Figure 1.
|
| 109 |
+
|
| 110 |
+
User study. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side view, with both samples generated from the same text prompt randomly sampled from the KIT test set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even preferred over ground truth samples in $4 2 . 3 \%$ of the cases.
|
| 111 |
+
|
| 112 |
+
# 4.2 ACTION-TO-MOTION
|
| 113 |
+
|
| 114 |
+
Action-to-motion is the task of generating motion given an input action class, represented by a scalar. The output motion should faithfully animate the input action, and at the same time be natural and reflect the distribution of the dataset on which the model is trained. Two dataset are commonly used to evaluate action-to-motion models: HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018).
|
| 115 |
+
|
| 116 |
+
Table 1: Quantitative results on the HumanML3D test set. All methods use the real motion length from the ground truth. $\cdot _ { } ,$ means results are better if the metric is closer to the real distribution. We run all the evaluation 20 times (except MultiModality runs 5 times) and $\pm$ indicates the $9 5 \%$ confidence interval. Bold indicates best result.
|
| 117 |
+
|
| 118 |
+
<table><tr><td>Method</td><td>RPrecision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td><td>Multimodality↑</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td></td></tr><tr><td>JL2P</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>1</td></tr><tr><td>Text2Gesture</td><td>0.345±.002</td><td>7.664±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>1</td></tr><tr><td>T2M</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MDM (ours)</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>+ sent-BERT</td><td>0.609±.006</td><td>0.586±.036</td><td>5.504±.03</td><td>9.666±.095</td><td>2.707±.188</td></tr><tr><td>MDM (decoder)</td><td>0.608±.005</td><td>0.767±.085</td><td>5.507±.020</td><td>9.176±.070</td><td>2.927±.125</td></tr><tr><td>+ input token</td><td>0.621±.005</td><td>0.567±.051</td><td>5.424±.022</td><td>9.425±.060</td><td>2.834±.095</td></tr><tr><td>MDM (U-net)</td><td>0.603±.006</td><td>1.137±.008</td><td>5.629±.032</td><td>8.958±.098</td><td>2.636±.214</td></tr><tr><td>MDM (GRU)</td><td>0.645±.005</td><td>4.569±.150</td><td>5.325±.026</td><td>7.688±.082</td><td>1.2646±.024</td></tr></table>
|
| 119 |
+
|
| 120 |
+
Table 2: Quantitative results on the KIT test set.
|
| 121 |
+
|
| 122 |
+
<table><tr><td>Method</td><td>R Precision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td><td>Multimodality↑</td></tr><tr><td>Real</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>JL2P</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Text2Gesture</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td></td></tr><tr><td>T2M</td><td>0.693±.007</td><td>2.770±.109</td><td>3.401±.008</td><td>10.91±.119</td><td>1.482±.065</td></tr><tr><td>MDM (ours)</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.847±.109</td><td>1.907±.214</td></tr></table>
|
| 123 |
+
|
| 124 |
+
We evaluate our model using the set of metrics suggested by Guo et al. (2020), namely Frechet In- ´ ception Distance (FID), action recognition accuracy, diversity and multimodality. The combination of these metrics makes a good measure of the realism and diversity of generated motions.
|
| 125 |
+
|
| 126 |
+
Data. HumanAct12 (Guo et al., 2020) offers approximately 1200 motion clips, organized into 12 action categories, with 47 to 218 samples per label. UESTC (Ji et al., 2018) consists of 40 action classes, 40 subjects and 25K samples, and is split to train and test. We adhere to the cross-subject testing protocol used by current works, with 225-345 samples per action class. For both datasets we use the sequences provided by Petrovich et al. (2021).
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
Figure 4: (a) Text-to-motion user study for the KIT dataset. Each bar represents the preference rate of MDM over the compared model. MDM was preferred over the other models in most of the time, and $4 2 . 3 \%$ of the cases even over ground truth samples. The dashed line marks $5 0 \%$ . (b) Guidance-scale sweep for HumanML3D dataset. $F I D$ (lower is better) and $R$ -precision (higher is better) metrics as a function of the scale $s$ , draws an accuracy-fidelity sweet spot around $s = 2 . 5$ .
|
| 130 |
+
|
| 131 |
+
Table 3: Evaluation of action-to-motion on the HumanAct12 dataset. Our model leads the board in three out of four metrics. Ground-truth evaluation results are slightly different for each of the works, due to implementation differences, such as python package versions. It is important to assess the diversity and multimodality of each model using its own ground-truth results, as they are measured by their distance from GT. We show the GT metrics measured by our model and by the leading compared work, INR (Cervantes et al., 2022). Bold indicates best result, underline indicates second best, $\pm$ indicates $9 5 \%$ confidence interval, indicates that closer to real is better.
|
| 132 |
+
|
| 133 |
+
<table><tr><td>Method</td><td>FID↓</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real (INR)</td><td>0.020±.010</td><td>0.997±.001</td><td>6.850±.050</td><td>2.450±.040</td></tr><tr><td>Real (ours)</td><td>0.050±.000</td><td>0.990±.000</td><td>6.880±.020</td><td>2.590±.010</td></tr><tr><td>Action2Motion (2020)</td><td>0.338±.015</td><td>0.917±.003</td><td>6.879±.066</td><td>2.511±.023</td></tr><tr><td>ACTOR (2021)</td><td>0.120±.000</td><td>0.955±.008</td><td>6.840±.030</td><td>2.530±.020</td></tr><tr><td>INR (2022)</td><td>0.088±.004</td><td>0.973±.001</td><td>6.881±.048</td><td>2.569±.040</td></tr><tr><td>MDM (ours)</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>w/o foot contact</td><td>0.080±.000</td><td>0.990±.000</td><td>6.810±.010</td><td>2.580±.010</td></tr></table>
|
| 134 |
+
|
| 135 |
+
Table 4: Evaluation of action-to-motion on the UESTC dataset. The performance improvement with our model shows a clear gap from state-of-the-art. Bold indicates best result, underline indicates second best, $\pm$ indicates $9 5 \%$ confidence interval, indicates that closer to real is better.
|
| 136 |
+
|
| 137 |
+
<table><tr><td>Method</td><td>FIDtrain </td><td>FIDtest</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real</td><td>2.92±.26</td><td>2.79±.29</td><td>0.988±.001</td><td>33.34±.320</td><td>14.16±.06</td></tr><tr><td>ACTOR (2021)</td><td>20.49±2.31</td><td>23.43±2.20</td><td>0.911±.003</td><td>31.96±.33</td><td>14.52±.09</td></tr><tr><td>INR (2022) (best variation)</td><td>9.55±.06</td><td>15.00±.09</td><td>0.941±.001</td><td>31.59±.19</td><td>14.68±.07</td></tr><tr><td>MDM (ours)</td><td>9.98±1.33</td><td>12.81±1.46</td><td>0.950±.000</td><td>33.02±.28</td><td>14.26±.12</td></tr><tr><td>w/o foot contact</td><td>9.69±.81</td><td>13.08±2.32</td><td>0.960±.000</td><td>33.10±.29</td><td>14.06±.05</td></tr></table>
|
| 138 |
+
|
| 139 |
+
Implementation. The implementation presented in Figure 2 holds for all the variations of our work. In the case of action-to-motion, the only change would be the substitution of the text embedding by an action embedding. Since action is represented by a scalar, its embedding is fairly simple; each input action class scalar is converted into a learned embedding of the transformer dimension.
|
| 140 |
+
|
| 141 |
+
The experiments have been run with batch size 64, a latent dimension of 512, and an encodertransformer architecture. Training on HumanAct12 and UESTC has been carried out for $7 5 0 K$ and $2 M$ steps respectively. In our tables we display the evaluation of the checkpoint that minimizes the FID metric.
|
| 142 |
+
|
| 143 |
+
Quantitative evaluation. Tables 3 and 4 reflect MDM’s performance on the HumanAct12 and UESTC datasets respectively. We conduct 20 evaluations, with 1000 samples in each, and report their average and a $9 5 \%$ confidence interval. We test two variations, with and without foot contact loss. Full ablation study for geometric losses is presented in Appendix A.2. Our model leads the board for both datasets. The variation with no foot contact loss attains slightly better results; nevertheless, as shown in our supplementary video, the contribution of foot contact loss to the quality of results is important, and without it we witness artifacts such as shakiness and unnatural gestures.
|
| 144 |
+
|
| 145 |
+
# 5 ADDITIONAL APPLICATIONS
|
| 146 |
+
|
| 147 |
+
# 5.1 MOTION EDITING
|
| 148 |
+
|
| 149 |
+
In this section we implement two motion editing applications - in-betweening and body part editing, both using the same approach in the temporal and spatial domains correspondingly. For inbetweening, we fix the first and last $2 5 \%$ of the motion, leaving the model to generate the remaining $5 0 \%$ in the middle. For body part editing, we fix the joints we don’t want to edit and leave the
|
| 150 |
+
|
| 151 |
+
model to generate the rest. In particular, we experiment with editing the upper body joints only. In figure 3 we show that in both cases, using the method described in Section 3 generates smooth motions that adhere both to the fixed part of the motion and the condition (if one was given).
|
| 152 |
+
Table 5: Evaluation of unconstrained synthesis on the HumanAct12 dataset. We test MDM in the challenging unconstrained setting, and compare with MoDi (Raab et al., 2022), a work that was specially designed for such setting. We demonstrate that in addition to being able to support any condition, we can achieve plausible results in the unconstrained setting. Bold indicates best result.
|
| 153 |
+
|
| 154 |
+
<table><tr><td>Method</td><td>FID↓</td><td>KID↓</td><td>Precision↑ Recall↑</td><td>Diversity↑</td></tr><tr><td>ACTOR (2021)</td><td>48.80</td><td>0.53</td><td>0.72, 0.74</td><td>14.10</td></tr><tr><td>MoDi (2022)</td><td>13.03</td><td>0.12</td><td>0.71, 0.81</td><td>17.57</td></tr><tr><td>MDM (ours)</td><td>31.92</td><td>0.36</td><td>0.66,0.62</td><td>17.00</td></tr></table>
|
| 155 |
+
|
| 156 |
+
# 5.2 UNCONSTRAINED SYNTHESIS
|
| 157 |
+
|
| 158 |
+
The challenging task of unconstrained synthesis has been studied by only a few (Holden et al., 2016; Raab et al., 2022). In the presence of data labeling, e.g., action classes or text description, the labels work as a supervising factor, and facilitate a structured latent space for the training network. The lack of labeling make training more difficult. The human motion field possesses rich unlabeled datasets (Adobe Systems Inc., 2021), and the ability to train on top of them is an advantage. Daring to test MDM in the challenging unconstrained setting, we follow MoDi(Raab et al., 2022) for evaluation. We use the metrics they suggest (FID, KID, precision/recall and multimodality), and run on an unconstrained version of the HumanAct12 (Guo et al., 2020) dataset.
|
| 159 |
+
|
| 160 |
+
Data. Although annotated, we use HumanAct12 (see Section 4.2) in an unconstrained fashion, ignoring its labels. The choice of HumanAct12 rather than a dataset with no labels (e.g., Mixamo (Adobe Systems Inc., 2021)), is for compatibility with previous publications.
|
| 161 |
+
|
| 162 |
+
Implementation. Our model uses the same architecture for all forms of conditioning, as well as for the unconstrained setting. The only change to the structure shown in Figure 2, is the removal of the conditional input, such that $z _ { t k }$ is composed of the projection of $t$ only. To simulate an unconstrained behavior, ACTOR Petrovich et al. (2021) has been trained by (Raab et al., 2022) with a labeling of one class to all motions.
|
| 163 |
+
|
| 164 |
+
Quantitative evaluation. The results of our evaluation are shown in table 5. We demonstrate superiority over works that were not designed for an unconstrained setting, and get closer to MoDi (Raab et al., 2022). MoDi is carefully molded for unconstrained settings, while our work can be applied to any (or no) constrain, and also provides editing capabilities.
|
| 165 |
+
|
| 166 |
+
# 6 DISCUSSION
|
| 167 |
+
|
| 168 |
+
We have presented MDM, a method that lends itself to various human motion generation tasks. MDM is an untypical classifier-free diffusion model, featuring a transformer-encoder backbone, and predicting the signal, rather than the noise. This yields both a lightweight model, that is unburdening to train, and an accurate one, gaining much from the applicable geometric losses. Our experiments show superiority in conditioned generation, but also that this approach is not very sensitive to the choice of architecture. A notable limitation of the diffusion approach is the long inference time, requiring about 1000 forward passes for a single result. Since our motion model is small anyway, using dimensions order of magnitude smaller than images, our inference time shifts from less than a second to only about a minute, which is an acceptable compromise. As diffusion models continue to evolve, besides better compute, in the future we would be interested in seeing how to incorporate better control into the generation process and widen the options for applications even further.
|
| 169 |
+
|
| 170 |
+
# ACKNOWLEDGEMENTS
|
| 171 |
+
|
| 172 |
+
We thank Rinon Gal for his useful suggestions and references. This research was supported in part by the Israel Science Foundation (grants no. 2492/20 and 3441/21), Len Blavatnik and the Blavatnik family foundation, and The Tel Aviv University Innovation Laboratories (TILabs).
|
| 173 |
+
|
| 174 |
+
# REFERENCES
|
| 175 |
+
|
| 176 |
+
Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62–1, 2020.
|
| 177 |
+
|
| 178 |
+
Adobe Systems Inc. Mixamo, 2021. URL https://www.mixamo.com. Accessed: 2021-12-25.
|
| 179 |
+
|
| 180 |
+
Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pp. 719–728. IEEE, 2019.
|
| 181 |
+
|
| 182 |
+
Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pp. 565–574. IEEE, 2021.
|
| 183 |
+
|
| 184 |
+
A Aristidou, A Yiannakidis, K Aberman, D Cohen-Or, A Shamir, and Y Chrysanthou. Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics, 2022.
|
| 185 |
+
|
| 186 |
+
Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pp. 1–10. IEEE, 2021.
|
| 187 |
+
|
| 188 |
+
Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. arXiv preprint arXiv:2203.13694, 2022.
|
| 189 |
+
|
| 190 |
+
Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨ ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.
|
| 191 |
+
|
| 192 |
+
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https: //aclanthology.org/N19-1423.
|
| 193 |
+
|
| 194 |
+
Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780–8794, 2021.
|
| 195 |
+
|
| 196 |
+
Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021.
|
| 197 |
+
|
| 198 |
+
Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE international conference on computer vision, pp. 4346–4354, 2015.
|
| 199 |
+
|
| 200 |
+
Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17113–17122, 2022.
|
| 201 |
+
|
| 202 |
+
Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pp. 2021–2029, 2020.
|
| 203 |
+
|
| 204 |
+
Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5152–5161, 2022a.
|
| 205 |
+
|
| 206 |
+
Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc MorenoNoguer. Back to mlp: A simple baseline for human motion prediction. arXiv preprint arXiv:2207.01567, 2022b.
|
| 207 |
+
|
| 208 |
+
Felix G Harvey and Christopher Pal. Recurrent transition networks for character locomotion. In ´ SIGGRAPH Asia 2018 Technical Briefs, pp. 1–4. 2018.
|
| 209 |
+
|
| 210 |
+
Felix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in- ´ betweening. ACM Transactions on Graphics (TOG), 39(4):60–1, 2020.
|
| 211 |
+
|
| 212 |
+
Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7134–7143, 2019.
|
| 213 |
+
|
| 214 |
+
Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022.
|
| 215 |
+
|
| 216 |
+
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020.
|
| 217 |
+
|
| 218 |
+
Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022.
|
| 219 |
+
|
| 220 |
+
Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016.
|
| 221 |
+
|
| 222 |
+
Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng. A large-scale rgb-d database for arbitrary-view human action recognition. In Proceedings of the 26th ACM international Conference on Multimedia, pp. 1510–1518, 2018.
|
| 223 |
+
|
| 224 |
+
Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pp. 918–927. IEEE, 2020.
|
| 225 |
+
|
| 226 |
+
Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022.
|
| 227 |
+
|
| 228 |
+
Diederik $\mathrm { \bf P }$ Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
|
| 229 |
+
|
| 230 |
+
Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5253–5263, 2020.
|
| 231 |
+
|
| 232 |
+
Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist $^ { + + }$ . In The IEEE International Conference on Computer Vision (ICCV), 2021.
|
| 233 |
+
|
| 234 |
+
Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2837–2845, 2021.
|
| 235 |
+
|
| 236 |
+
Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pp. 5442–5451, October 2019.
|
| 237 |
+
|
| 238 |
+
Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2891–2900, 2017.
|
| 239 |
+
|
| 240 |
+
Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.
|
| 241 |
+
|
| 242 |
+
Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 10975–10985, 2019.
|
| 243 |
+
|
| 244 |
+
Mathis Petrovich, Michael J. Black, and Gul Varol. Action-conditioned 3D human motion synthesis ¨ with transformer VAE. In International Conference on Computer Vision (ICCV), pp. 10985– 10995, October 2021.
|
| 245 |
+
|
| 246 |
+
Mathis Petrovich, Michael J. Black, and Gul Varol. TEMOS: Generating diverse human motions ¨ from textual descriptions. In European Conference on Computer Vision (ECCV), 2022.
|
| 247 |
+
|
| 248 |
+
Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236–252, 2016.
|
| 249 |
+
|
| 250 |
+
Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel CohenOr. Modi: Unconditional motion synthesis from diverse data. arXiv preprint arXiv:2206.08010, 2022.
|
| 251 |
+
|
| 252 |
+
Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pp. 8748–8763. PMLR, 2021.
|
| 253 |
+
|
| 254 |
+
Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical textconditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
|
| 255 |
+
|
| 256 |
+
Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bertnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pp. 3982–3992, 2019.
|
| 257 |
+
|
| 258 |
+
Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pp. 234–241. Springer, 2015.
|
| 259 |
+
|
| 260 |
+
Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pp. 1–10, 2022a.
|
| 261 |
+
|
| 262 |
+
Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
|
| 263 |
+
|
| 264 |
+
Mingyi Shi, Kfir Aberman, Andreas Aristidou, Taku Komura, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Motionet: 3d human motion reconstruction from monocular video with skeleton consistency. ACM Transactions on Graphics (TOG), 40(1):1–15, 2020.
|
| 265 |
+
|
| 266 |
+
Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp. 2256–2265. PMLR, 2015.
|
| 267 |
+
|
| 268 |
+
Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020a.
|
| 269 |
+
|
| 270 |
+
Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438–12448, 2020.
|
| 271 |
+
|
| 272 |
+
Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020b.
|
| 273 |
+
|
| 274 |
+
Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022.
|
| 275 |
+
|
| 276 |
+
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.
|
| 277 |
+
|
| 278 |
+
Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.
|
| 279 |
+
|
| 280 |
+
# A ADDITIONAL EXPERIMENTS
|
| 281 |
+
|
| 282 |
+
# A.1 DIFFUSION PARAMETERS
|
| 283 |
+
|
| 284 |
+
Learning MDM with different numbers of diffusion steps significantly affects performance and holds the potential to accelerate inference time. Table 6 shows optimal performance for $T = 1 0 0$ , in addition, it enables accelerating inference by a factor of 10 compared to $T = 1 0 0 0$ , which is widely used for images.
|
| 285 |
+
|
| 286 |
+
Table 6: Diffusion steps (HumanML3D test set). We run all the evaluation 20 times. Bold indicates best result, underline indicates second best, $\pm$ indicates $9 5 \%$ confidence interval, indicates that closer to real is better.
|
| 287 |
+
|
| 288 |
+
<table><tr><td>Diffusion steps (T)</td><td>R Precision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td></tr><tr><td>10</td><td>0.574±.006</td><td>1.461±.088</td><td>5.816±.033</td><td>9.369±.058</td></tr><tr><td>100</td><td>0.640±.007</td><td>0.454±.039</td><td>5.336±.029</td><td>9.906±.053</td></tr><tr><td>500</td><td>0.662±.007</td><td>0.553±.055</td><td>5.177±.028</td><td>9.890±.074</td></tr><tr><td>1000</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td></tr></table>
|
| 289 |
+
|
| 290 |
+
# A.2 GEOMETRIC LOSSES
|
| 291 |
+
|
| 292 |
+
We conduct a thorough experiment to evaluate the contribution of geometric losses with the HumanAct12 dataset. The results are presented in Table 7. For alignment with prior work, all metrics are calculated using the deep features of the action recognition network suggested by Guo et al. (2020). In general, MDM scores are too close to the real test distribution (i.e. the evaluation network fails to discriminate between the two). This means that quantitative results comparing the different variants MDM are too similar to evaluate. As a result, we are not able to decide what combination of geometric losses is preferred. We leave for future work experimenting with a different, more expressive, evaluation network.
|
| 293 |
+
|
| 294 |
+
<table><tr><td>Method</td><td>FID↓</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real</td><td>0.050±.000</td><td>0.990±.000</td><td>6.880±.020</td><td>2.590±.010</td></tr><tr><td>MDM (ours)</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>w/o foot contact</td><td>0.080±.000</td><td>0.990±.000</td><td>6.810±.010</td><td>2.580±.010</td></tr><tr><td>w/o geometric losses</td><td>0.090±.000</td><td>0.990±.000</td><td>6.820±.020</td><td>2.550±.010</td></tr><tr><td>foot contact only</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>velocity only</td><td>0.100±.000</td><td>0.990±.000</td><td>6.820±.020</td><td>2.590±.000</td></tr><tr><td>pose only</td><td>0.090±.000</td><td>0.990±.000</td><td>6.830±.020</td><td>2.570±.020</td></tr></table>
|
| 295 |
+
|
| 296 |
+
Table 7: Geometric losses ablation study. (HumanAct12 dataset) the relative $\lambda$ equals 1 when the loss term is included, and 0 when it is excluded.
|
| 297 |
+
|
| 298 |
+
# B EVALUATION METRICS.
|
| 299 |
+
|
| 300 |
+
For the completeness of our work, we describe here the quantitative metrics used throughout the paper, as they originally described and implemented by Guo et al. (2020) for action-to-motion and by Guo et al. (2022a) for text-to-motion.
|
| 301 |
+
|
| 302 |
+
# B.1 ACTION-TO-MOTION
|
| 303 |
+
|
| 304 |
+
The following metrics are based on an RNN action recognition network as it was originally trained by Guo et al. (2020). We refer to it as the evaluator network.
|
| 305 |
+
|
| 306 |
+
Frechet Inception Distance (FID). A widely used metric to evaluate the overall quality for generation tasks. FID is calculated upon features extracted from 1,000 generated motion vs ground truth (real) taken from the test set. To adjust this metric to the motion domain, we extract a deep representation of the motion with the evaluator network instead of the inception neural network, originally used for images. A lower value implies better FID results.
|
| 307 |
+
|
| 308 |
+
Accuracy. We classify 1,000 generated motions using the evaluator network, than we calculate the overall recognition accuracy that indicates the correlation of the motion and its action type.
|
| 309 |
+
|
| 310 |
+
Diversity measures the variance of the generated motions across all action categories. We first randomly sample two subsets of the same size $S _ { d }$ out of a set of all generated motions across all action categories denoted $\{ \mathbf { v } _ { 1 } , . . . , \mathbf { v } _ { S _ { d } } \}$ and $\big \{ \mathbf { v } _ { 1 } ^ { \prime \prime } , . . . , \mathbf { v } _ { S _ { d } } ^ { \prime } \big \}$ . The diversity of those sets of motions is defied as
|
| 311 |
+
|
| 312 |
+
$$
|
| 313 |
+
\mathrm { D i v e r s i t y } = \frac { 1 } { S _ { d } } \sum _ { i = 1 } ^ { S _ { d } } \parallel \mathbf { v } _ { i } - \mathbf { v } _ { i } ^ { \prime } \parallel _ { 2 } .
|
| 314 |
+
$$
|
| 315 |
+
|
| 316 |
+
We use $S _ { d } = 2 0 0$ for our experiments. The diversity value is considered better when closer to the diversity value of the ground truth.
|
| 317 |
+
|
| 318 |
+
Multimodality measures the generated motions diversify within each action class. We randomly sample two subsets with size $S _ { l }$ of the same motion class $c$ $\{ \mathbf { v } _ { c , 1 } , \ldots , \mathbf { v } _ { c , S _ { l } } \}$ and $\{ \mathbf { v } _ { c , 1 } ^ { \prime } , . . . , \mathbf { v } _ { c , S _ { l } } ^ { \prime } \}$ . The multimodality of all action classes $C$ is defined as,
|
| 319 |
+
|
| 320 |
+
$$
|
| 321 |
+
\mathrm { M u l t i m o d a l i t y } = \frac { 1 } { C \times S _ { l } } \sum _ { c = 1 } ^ { C } \sum _ { i = 1 } ^ { S _ { l } } \left\| \mathbf { v } _ { c , i } - \mathbf { v } _ { c , i } ^ { \prime } \right\| _ { 2 } .
|
| 322 |
+
$$
|
| 323 |
+
|
| 324 |
+
We use $S _ { l } = 2 0$ for our experiments.
|
| 325 |
+
|
| 326 |
+
# B.2 TEXT-TO-MOTION
|
| 327 |
+
|
| 328 |
+
Originally suggested by Guo et al. (2022a), the following metrics are based on a text feature extractor and motion feature extractor jointly trained under contrastive loss to produce geometrically close feature vectors for matched text-motion pairs, and vise versa.
|
| 329 |
+
|
| 330 |
+
R Precision. (top-3) For each generated motion, its ground-truth text and a randomly selected missmatched descriptions from the test set. We calculate the euclidean distance between the motion feature and text feature of each description in the pool. We count the average accuracy at top3 places. If the ground truth entry falling into the top-3 candidates, we treat it as True Positive retrieval. We use a batch size 32 (i.e. 31 negative examples).
|
| 331 |
+
|
| 332 |
+
FID. Same as for action-to-motion, using the motion extractor as the evaluator network.
|
| 333 |
+
|
| 334 |
+
Multimodal Distance. We calculate the multimodal distance as the average Euclidean distance between the motion feature of each generated motion and the text feature of its corresponding description in test set. A lower value implies better multimodal distance.
|
| 335 |
+
|
| 336 |
+
Diversity. Same as for action-to-motion but with $S _ { d } = 3 0 0$ .
|
| 337 |
+
|
| 338 |
+
Multimodality. Same as for action-to-motion but with $S _ { m } = 1 0$ .
|
| 339 |
+
|
| 340 |
+
# C IMPLEMENTATION DETAILS
|
| 341 |
+
|
| 342 |
+
The full implementation of MDM can be found in our published code2. In addition, the followings are the hyperparameters and model details for all of our experiments.
|
| 343 |
+
|
| 344 |
+
Diffusion framework. In all of our experiments, we used an implementation of DDPM (Ho et al., 2020) by Dhariwal & Nichol $( 2 0 2 1 ) ^ { 3 }$ . We use $T = 1 , 0 0 0$ diffusion steps, cosine noise scheduling (predefined sigmas). All other hyperparameters are according to the implementation defaults.
|
| 345 |
+
|
| 346 |
+
Transformer architecture. For our transformer architectures, we used the PyTorch implementation4. We used 8 transformer layers, 4 attention heads, latent dimension $d = 5 1 2$ , dropout 0.1, feed-forward size 1024 and gelu activations. The number of learned parameters for each model is stated in Table 8.
|
| 347 |
+
|
| 348 |
+
GRU architecture. We use the PyTorch implementation of GRU (Cho et al., 2014) 5 with two layers and latent dimension 512. The number of learned parameters for each model is stated in Table 8.
|
| 349 |
+
|
| 350 |
+
Learning hyperparameters. For all of our experiments, we use batch size 64, learning rate $1 0 ^ { - 4 }$ .
|
| 351 |
+
|
| 352 |
+
<table><tr><td>Architecture</td><td># Parameters (-106)</td></tr><tr><td>Transformer Encoder</td><td>17.88</td></tr><tr><td>TransformerDecoder</td><td>26.29</td></tr><tr><td>+ input token</td><td>26.29</td></tr><tr><td>U-net</td><td>23.47</td></tr><tr><td>GRU</td><td>4.47</td></tr></table>
|
| 353 |
+
|
| 354 |
+
Table 8: The number of learned parameters per architecture for the text-to-motion task. For the action-to-motion task, there are additional 512 parameters per-class for the class embeddings module.
|
| 355 |
+
|
| 356 |
+
# D USER STUDY
|
| 357 |
+
|
| 358 |
+
In Section 4.1 we conduct a user study for the text-to-motion task. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side view, with both samples generated from the same text prompt randomly sampled from the KIT test set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even preferred over ground truth samples in $4 2 . 3 \%$ of the cases. This user study was designed to measure the precision of the models, i.e. which one better fits the input text. The exact phrasing of the question was “Which animation better fits the following description?”. A sample question from this study is presented in Fig. 5.
|
| 359 |
+
|
| 360 |
+
|
| 361 |
+
Figure 5: An example question for our text-to-motion user study, using the Google Forms platform.
|
parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_content_list.json
ADDED
|
@@ -0,0 +1,1940 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "HUMAN MOTION DIFFUSION MODEL ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
174,
|
| 8 |
+
98,
|
| 9 |
+
619,
|
| 10 |
+
121
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or and Amit H. Bermano ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
183,
|
| 19 |
+
143,
|
| 20 |
+
562,
|
| 21 |
+
172
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Tel Aviv University, Israel guytevet@mail.tau.ac.il ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
184,
|
| 30 |
+
172,
|
| 31 |
+
408,
|
| 32 |
+
200
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "ABSTRACT ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
454,
|
| 42 |
+
237,
|
| 43 |
+
544,
|
| 44 |
+
252
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "Natural and expressive human motion generation is the holy grail of computer animation. It is a challenging task, due to the diversity of possible motion, human perceptual sensitivity to it, and the difficulty of accurately describing it. Therefore, current generative solutions are either low-quality or limited in expressiveness. Diffusion models, which have already shown remarkable generative capabilities in other domains, are promising candidates for human motion due to their many-to-many nature, but they tend to be resource hungry and hard to control. In this paper, we introduce Motion Diffusion Model (MDM), a carefully adapted classifier-free diffusion-based generative model for the human motion domain. MDM is transformer-based, combining insights from motion generation literature. A notable design-choice is the prediction of the sample, rather than the noise, in each diffusion step. This facilitates the use of established geometric losses on the locations and velocities of the motion, such as the foot contact loss. As we demonstrate, MDM is a generic approach, enabling different modes of conditioning, and different generation tasks. We show that our model is trained with lightweight resources and yet achieves state-ofthe-art results on leading benchmarks for text-to-motion and action-to-motion 1. https://guytevet.github.io/mdm-page/. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
267,
|
| 54 |
+
764,
|
| 55 |
+
517
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 INTRODUCTION ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
176,
|
| 65 |
+
542,
|
| 66 |
+
336,
|
| 67 |
+
558
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Human motion generation is a fundamental task in computer animation, with applications spanning from gaming to robotics. It is a challenging field, due to several reasons, including the vast span of possible motions, and the difficulty and cost of acquiring high quality data. For the recently emerging text-to-motion setting, where motion is generated from natural language, another inherent problem is data labeling. For example, the label ”kick” could refer to a soccer kick, as well as a Karate one. At the same time, given a specific kick there are many ways to describe it, from how it is performed to the emotions it conveys, constituting a many-to-many problem. Current approaches have shown success in the field, demonstrating plausible mapping from text to motion (Petrovich et al., 2022; Tevet et al., 2022; Ahuja & Morency, 2019). All these approaches, however, still limit the learned distribution since they mainly employ auto-encoders or VAEs (Kingma & Welling, 2013) (implying a one-to-one mapping or a normal latent distribution respectively). In this aspect, diffusion models are a better candidate for human motion generation, as they are free from assumptions on the target distribution, and are known for expressing well the many-to-many distribution matching problem we have described. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
566,
|
| 77 |
+
825,
|
| 78 |
+
760
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020; Ho et al., 2020) are a generative approach that is gaining significant attention in the computer vision and graphics community. When trained for conditioned generation, recent diffusion models (Ramesh et al., 2022; Saharia et al., 2022b) have shown breakthroughs in terms of image quality and semantics. The competence of these models have also been shown for other domains, including videos (Ho et al., 2022), and 3D point clouds (Luo & Hu, 2021). The problem with such models, however, is that they are notoriously resource demanding and challenging to control. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
768,
|
| 88 |
+
825,
|
| 89 |
+
866
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "In this paper, we introduce Motion Diffusion Model (MDM) — a carefully adapted diffusion based generative model for the human motion domain. Being diffusion-based, MDM gains from the na“A man runs to the right then runs to the left then back to the middle.” ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
178,
|
| 98 |
+
872,
|
| 99 |
+
823,
|
| 100 |
+
901
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "image",
|
| 106 |
+
"img_path": "images/679e88c7d9a9e84834db9606fa91131fa59019df65a3ddf03e3c9875429bf51b.jpg",
|
| 107 |
+
"image_caption": [],
|
| 108 |
+
"image_footnote": [],
|
| 109 |
+
"bbox": [
|
| 110 |
+
178,
|
| 111 |
+
101,
|
| 112 |
+
800,
|
| 113 |
+
253
|
| 114 |
+
],
|
| 115 |
+
"page_idx": 1
|
| 116 |
+
},
|
| 117 |
+
{
|
| 118 |
+
"type": "text",
|
| 119 |
+
"text": "",
|
| 120 |
+
"bbox": [
|
| 121 |
+
173,
|
| 122 |
+
263,
|
| 123 |
+
558,
|
| 124 |
+
276
|
| 125 |
+
],
|
| 126 |
+
"page_idx": 1
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"type": "image",
|
| 130 |
+
"img_path": "images/8f0257ea475bbecd824964d2530d3df4475129bcbbba57fb1a6abf92a0e2e30c.jpg",
|
| 131 |
+
"image_caption": [
|
| 132 |
+
"Figure 1: Our Motion Diffusion Model (MDM) reflects the many-to-many nature of text-to-motion mapping by generating diverse motions given a text prompt. Our custom architecture and geometric losses help yielding high-quality motion. Darker color indicates later frames in the sequence. "
|
| 133 |
+
],
|
| 134 |
+
"image_footnote": [],
|
| 135 |
+
"bbox": [
|
| 136 |
+
174,
|
| 137 |
+
281,
|
| 138 |
+
825,
|
| 139 |
+
405
|
| 140 |
+
],
|
| 141 |
+
"page_idx": 1
|
| 142 |
+
},
|
| 143 |
+
{
|
| 144 |
+
"type": "text",
|
| 145 |
+
"text": "tive aforementioned many-to-many expression of the domain, as evidenced by the resulting motion quality and diversity (Figure 1). In addition, MDM combines insights already well established in the motion generation domain, helping it be significantly more lightweight and controllable. ",
|
| 146 |
+
"bbox": [
|
| 147 |
+
176,
|
| 148 |
+
478,
|
| 149 |
+
820,
|
| 150 |
+
520
|
| 151 |
+
],
|
| 152 |
+
"page_idx": 1
|
| 153 |
+
},
|
| 154 |
+
{
|
| 155 |
+
"type": "text",
|
| 156 |
+
"text": "First, instead of the ubiquitous U-net (Ronneberger et al., 2015) backbone, MDM is transformerbased. As we demonstrate, our architecture (Figure 2) is lightweight and better fits the temporal and spatially irregular nature of motion data (represented as a collection of joints). A large volume of motion generation research is devoted to learning using geometric losses (Kocabas et al., 2020; Harvey et al., 2020; Aberman et al., 2020). Some, for example, regulate the velocity of the motion (Petrovich et al., 2021) to prevent jitter, or specifically consider foot sliding using dedicated terms (Shi et al., 2020). Consistently with these works, we show that applying geometric losses in the diffusion setting improves generation. ",
|
| 157 |
+
"bbox": [
|
| 158 |
+
174,
|
| 159 |
+
526,
|
| 160 |
+
825,
|
| 161 |
+
638
|
| 162 |
+
],
|
| 163 |
+
"page_idx": 1
|
| 164 |
+
},
|
| 165 |
+
{
|
| 166 |
+
"type": "text",
|
| 167 |
+
"text": "The MDM framework has a generic design enabling different forms of conditioning. We showcase three tasks: text-to-motion, action-to-motion, and unconditioned generation. We train the model in a classifier-free manner (Ho & Salimans, 2022), which enables trading-off diversity to fidelity, and sampling both conditionally and unconditionally from the same model. In the text-to-motion task, our model generates coherent motions (Figure 1) that achieve state-of-the-art results on the HumanML3D (Guo et al., 2022a) and KIT (Plappert et al., 2016) benchmarks. Moreover, our user study shows that human evaluators prefer our generated motions over real motions $4 2 \\%$ of the time (Figure 4(a)). In action-to-motion, MDM outperforms the state-of-the-art (Guo et al., 2020; Petrovich et al., 2021), even though they were specifically designed for this task, on the common HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018) benchmarks. ",
|
| 168 |
+
"bbox": [
|
| 169 |
+
174,
|
| 170 |
+
645,
|
| 171 |
+
825,
|
| 172 |
+
784
|
| 173 |
+
],
|
| 174 |
+
"page_idx": 1
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"type": "text",
|
| 178 |
+
"text": "Lastly, we also demonstrate completion and editing. By adapting diffusion image-inpainting (Song et al., 2020b; Saharia et al., 2022a), we set a motion prefix and suffix, and use our model to fill in the gap. Doing so under a textual condition guides MDM to fill the gap with a specific motion that still maintains the semantics of the original input. By performing inpainting in the joints space rather than temporally, we also demonstrate the semantic editing of specific body parts, without changing the others (Figure 3). ",
|
| 179 |
+
"bbox": [
|
| 180 |
+
174,
|
| 181 |
+
791,
|
| 182 |
+
825,
|
| 183 |
+
875
|
| 184 |
+
],
|
| 185 |
+
"page_idx": 1
|
| 186 |
+
},
|
| 187 |
+
{
|
| 188 |
+
"type": "text",
|
| 189 |
+
"text": "Overall, we introduce Motion Diffusion Model, a motion framework that achieves state-of-the-art quality in several motion generation tasks, while requiring only about three days of training on a single mid-range GPU. It supports geometric losses, which are non trivial to the diffusion setting, but are crucial to the motion domain, and offers the combination of state-of-the-art generative power with well thought-out domain knowledge. ",
|
| 190 |
+
"bbox": [
|
| 191 |
+
176,
|
| 192 |
+
882,
|
| 193 |
+
823,
|
| 194 |
+
924
|
| 195 |
+
],
|
| 196 |
+
"page_idx": 1
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"type": "text",
|
| 200 |
+
"text": "",
|
| 201 |
+
"bbox": [
|
| 202 |
+
174,
|
| 203 |
+
103,
|
| 204 |
+
823,
|
| 205 |
+
132
|
| 206 |
+
],
|
| 207 |
+
"page_idx": 2
|
| 208 |
+
},
|
| 209 |
+
{
|
| 210 |
+
"type": "text",
|
| 211 |
+
"text": "2 RELATED WORK ",
|
| 212 |
+
"text_level": 1,
|
| 213 |
+
"bbox": [
|
| 214 |
+
176,
|
| 215 |
+
146,
|
| 216 |
+
344,
|
| 217 |
+
162
|
| 218 |
+
],
|
| 219 |
+
"page_idx": 2
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"type": "text",
|
| 223 |
+
"text": "2.1 HUMAN MOTION GENERATION ",
|
| 224 |
+
"text_level": 1,
|
| 225 |
+
"bbox": [
|
| 226 |
+
176,
|
| 227 |
+
171,
|
| 228 |
+
429,
|
| 229 |
+
185
|
| 230 |
+
],
|
| 231 |
+
"page_idx": 2
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"type": "text",
|
| 235 |
+
"text": "Neural motion generation, learned from motion capture data, can be conditioned by any signal that describes the motion. Many works use parts of the motion itself for guidance. Some predict motion from its prefix poses (Fragkiadaki et al., 2015; Martinez et al., 2017; Hernandez et al., 2019; Guo et al., 2022b). Others (Harvey & Pal, 2018; Kaufmann et al., 2020; Harvey et al., 2020; Duan et al., 2021) solve in-betweening and super-resolution tasks using bi-directional GRU (Cho et al., 2014) and Transformer (Vaswani et al., 2017) architectures. Holden et al. (2016) use auto-encoder to learn motion latent representation, then utilize it to edit and control motion with spatial constraints such as root trajectory and bone lengths. Motion can be controlled with a high-level guidance given from action class (Guo et al., 2020; Petrovich et al., 2021; Cervantes et al., 2022), audio (Li et al., 2021; Aristidou et al., 2022) and natural language (Ahuja & Morency, 2019; Petrovich et al., 2022). In most cases authors suggests a dedicated approach to map each conditioning domain into motion. ",
|
| 236 |
+
"bbox": [
|
| 237 |
+
174,
|
| 238 |
+
191,
|
| 239 |
+
825,
|
| 240 |
+
344
|
| 241 |
+
],
|
| 242 |
+
"page_idx": 2
|
| 243 |
+
},
|
| 244 |
+
{
|
| 245 |
+
"type": "text",
|
| 246 |
+
"text": "In recent years, the leading approach for the Text-to-Motion task is to learn a shared latent space for language and motion. JL2P (Ahuja & Morency, 2019) learns the KIT motion-language dataset (Plappert et al., 2016) with an auto-encoder, limiting one-to-one mapping from text to motion. TEMOS (Petrovich et al., 2022) and T2M (Guo et al., 2022a) suggest using a VAE (Kingma & Welling, 2013) to map a text prompt into a normal distribution in latent space. Recently, MotionCLIP (Tevet et al., 2022) leverages the shared text-image latent space learned by CLIP (Radford et al., 2021) to expand text-to-motion out of the data limitations and enabled latent space editing. ",
|
| 247 |
+
"bbox": [
|
| 248 |
+
174,
|
| 249 |
+
351,
|
| 250 |
+
825,
|
| 251 |
+
449
|
| 252 |
+
],
|
| 253 |
+
"page_idx": 2
|
| 254 |
+
},
|
| 255 |
+
{
|
| 256 |
+
"type": "text",
|
| 257 |
+
"text": "The human motion manifold can also be learned without labels, as shown by Holden et al. (2016), V-Poser (Pavlakos et al., 2019), and more recently the dedicated MoDi architecture (Raab et al., 2022). We show that our model is capable for such an unsupervised setting as well. ",
|
| 258 |
+
"bbox": [
|
| 259 |
+
176,
|
| 260 |
+
455,
|
| 261 |
+
825,
|
| 262 |
+
498
|
| 263 |
+
],
|
| 264 |
+
"page_idx": 2
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"type": "text",
|
| 268 |
+
"text": "2.2 DIFFUSION GENERATIVE MODELS ",
|
| 269 |
+
"text_level": 1,
|
| 270 |
+
"bbox": [
|
| 271 |
+
178,
|
| 272 |
+
508,
|
| 273 |
+
452,
|
| 274 |
+
523
|
| 275 |
+
],
|
| 276 |
+
"page_idx": 2
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"type": "text",
|
| 280 |
+
"text": "Diffusion models (Sohl-Dickstein et al., 2015; Song & Ermon, 2020) are a class of neural generative models, based on the stochastic diffusion process as it is modeled in Thermodynamics. In this setting, a sample from the data distribution is gradually noised by the diffusion process. Then, a neural model learns the reverse process of gradually denoising the sample. Sampling the learned data distribution is done by denoising a pure initial noise. Ho et al. (2020) and Song et al. (2020a) further developed the practices for image generation applications. For conditioned generation, Dhariwal & Nichol (2021), introduced classifier-guided diffusion, which was later on adapted by GLIDE (Nichol et al., 2021) to enable conditioning over CLIP textual representations. The Classifier-Free Guidance approach Ho & Salimans (2022) enables conditioning while trading-off fidelity and diversity, and achieves better results (Nichol et al., 2021). In this paper, we implement text-to-motion by conditioning on CLIP in a classifier-free manner, similarly to text-to-image (Ramesh et al., 2022; Saharia et al., 2022b). Local editing of images is typically defined as an inpainting problem, where a part of the image is constant, and the inpainted part is denoised by the model, possibly under some condition (Song et al., 2020b; Saharia et al., 2022a). We adapt this technique to edit motion’s specific body parts or temporal intervals (in-betweening) according to an optional condition. ",
|
| 281 |
+
"bbox": [
|
| 282 |
+
174,
|
| 283 |
+
530,
|
| 284 |
+
825,
|
| 285 |
+
737
|
| 286 |
+
],
|
| 287 |
+
"page_idx": 2
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"type": "text",
|
| 291 |
+
"text": "Closer to our context, Gu et al. (2022) used the diffusion formulation to model the stochasticity of human trajectory prediction. More recently, concurrent to this work, Zhang et al. (2022) and Kim et al. (2022) have suggested diffusion models for motion generation. Our work requires significantly fewer GPU resources and makes design choices that enable geometric losses, which improve results. ",
|
| 292 |
+
"bbox": [
|
| 293 |
+
174,
|
| 294 |
+
744,
|
| 295 |
+
823,
|
| 296 |
+
800
|
| 297 |
+
],
|
| 298 |
+
"page_idx": 2
|
| 299 |
+
},
|
| 300 |
+
{
|
| 301 |
+
"type": "text",
|
| 302 |
+
"text": "3 MOTION DIFFUSION MODEL ",
|
| 303 |
+
"text_level": 1,
|
| 304 |
+
"bbox": [
|
| 305 |
+
176,
|
| 306 |
+
814,
|
| 307 |
+
442,
|
| 308 |
+
830
|
| 309 |
+
],
|
| 310 |
+
"page_idx": 2
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"type": "text",
|
| 314 |
+
"text": "An overview of our method is described in Figure 2. Our goal is to synthesize a human motion $x ^ { 1 : N }$ of length $N$ given an arbitrary condition $c$ . This condition can be any real-world signal that will dictate the synthesis, such as audio (Li et al., 2021; Aristidou et al., 2022), natural language (text-to-motion) (Tevet et al., 2022; Guo et al., 2022a) or a discrete class (action-to-motion) (Guo et al., 2020; Petrovich et al., 2021). Iwhich we denote as the null condition $c = \\emptyset$ tion, unconditioned mo. The generated motion $\\boldsymbol { x } ^ { 1 : N } = \\{ x ^ { i } \\} _ { i = 1 } ^ { N }$ s also possible,is a sequences of human poses represented by either joint rotations or positions $\\boldsymbol { x } ^ { i } \\in \\mathbb { R } ^ { J \\times D }$ , where $J$ is the number of joints and $D$ is the dimension of the joint representation. MDM can accept motion represented by either locations, rotations, or both (see Section 4). ",
|
| 315 |
+
"bbox": [
|
| 316 |
+
174,
|
| 317 |
+
840,
|
| 318 |
+
825,
|
| 319 |
+
924
|
| 320 |
+
],
|
| 321 |
+
"page_idx": 2
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"type": "image",
|
| 325 |
+
"img_path": "images/aceed64cf2b4fbe7a9a4e3115da6b56c1525a7c001a69455abf0263ac36383e4.jpg",
|
| 326 |
+
"image_caption": [
|
| 327 |
+
"Figure 2: (Left) Motion Diffusion Model (MDM) overview. The model is fed a motion sequence $x _ { t } ^ { 1 : N }$ of length $N$ in a noising step $t$ , as well as $t$ itself and a conditioning code c. c, a CLIP (Radford et al., 2021) based textual embedding in this case, is first randomly masked for classifier-free learning and then projected together with $t$ into the input token $z _ { t k }$ . In each sampling step, the transformerencoder predicts the final clean motion $\\hat { x } _ { 0 } ^ { 1 : N }$ . (Right) Sampling MDM. Given a condition $c$ , we sample random noise $x _ { T }$ at the dimensions of the desired motion, then iterate from $T$ to 1. At each step $t$ , MDM predicts the clean sample ${ \\hat { x } } _ { 0 }$ , and diffuses it back to $x _ { t - 1 }$ . "
|
| 328 |
+
],
|
| 329 |
+
"image_footnote": [],
|
| 330 |
+
"bbox": [
|
| 331 |
+
173,
|
| 332 |
+
94,
|
| 333 |
+
823,
|
| 334 |
+
268
|
| 335 |
+
],
|
| 336 |
+
"page_idx": 3
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"type": "text",
|
| 340 |
+
"text": "",
|
| 341 |
+
"bbox": [
|
| 342 |
+
173,
|
| 343 |
+
388,
|
| 344 |
+
823,
|
| 345 |
+
433
|
| 346 |
+
],
|
| 347 |
+
"page_idx": 3
|
| 348 |
+
},
|
| 349 |
+
{
|
| 350 |
+
"type": "text",
|
| 351 |
+
"text": "Framework. Diffusion is modeled as a Markov noising process, $\\{ x _ { t } ^ { 1 : N } \\} _ { t = 0 } ^ { T }$ , where $x _ { 0 } ^ { 1 : N }$ is drawn from the data distribution and ",
|
| 352 |
+
"bbox": [
|
| 353 |
+
171,
|
| 354 |
+
438,
|
| 355 |
+
823,
|
| 356 |
+
468
|
| 357 |
+
],
|
| 358 |
+
"page_idx": 3
|
| 359 |
+
},
|
| 360 |
+
{
|
| 361 |
+
"type": "equation",
|
| 362 |
+
"img_path": "images/11fd050c7836e50f9c8d71696e606b1b1ef5c767db8996739a597ba46ce164a5.jpg",
|
| 363 |
+
"text": "$$\nq ( x _ { t } ^ { 1 : N } | x _ { t - 1 } ^ { 1 : N } ) = \\mathcal { N } ( \\sqrt { \\alpha _ { t } } x _ { t - 1 } ^ { 1 : N } , ( 1 - \\alpha _ { t } ) I ) ,\n$$",
|
| 364 |
+
"text_format": "latex",
|
| 365 |
+
"bbox": [
|
| 366 |
+
357,
|
| 367 |
+
472,
|
| 368 |
+
638,
|
| 369 |
+
492
|
| 370 |
+
],
|
| 371 |
+
"page_idx": 3
|
| 372 |
+
},
|
| 373 |
+
{
|
| 374 |
+
"type": "text",
|
| 375 |
+
"text": "where $\\alpha _ { t } \\in ( 0 , 1 )$ are constant hyper-parameters. When $\\alpha _ { t }$ is small enough, we can approximate $x _ { T } ^ { 1 : N } \\sim \\mathcal { N } ( 0 , I )$ . From here on we use $x _ { t }$ to denote the full sequence at noising step $t$ . ",
|
| 376 |
+
"bbox": [
|
| 377 |
+
173,
|
| 378 |
+
498,
|
| 379 |
+
826,
|
| 380 |
+
527
|
| 381 |
+
],
|
| 382 |
+
"page_idx": 3
|
| 383 |
+
},
|
| 384 |
+
{
|
| 385 |
+
"type": "text",
|
| 386 |
+
"text": "In our context, conditioned motion synthesis models the distribution $p ( x _ { 0 } | c )$ as the reversed diffusion process of gradually cleaning $x _ { T }$ . Instead of predicting $\\epsilon _ { t }$ as formulated by Ho et al. (2020), we follow Ramesh et al. (2022) and use an equivalent formulation to predict the signal itself, i.e., $\\hat { x } _ { 0 } = G ( x _ { t } , t , c )$ with the simple objective (Ho et al., 2020), ",
|
| 387 |
+
"bbox": [
|
| 388 |
+
173,
|
| 389 |
+
534,
|
| 390 |
+
825,
|
| 391 |
+
589
|
| 392 |
+
],
|
| 393 |
+
"page_idx": 3
|
| 394 |
+
},
|
| 395 |
+
{
|
| 396 |
+
"type": "equation",
|
| 397 |
+
"img_path": "images/3c8560f67f4f45f2c98fdecc22a60f73367236e6abdf30d227a76e25ffda0db5.jpg",
|
| 398 |
+
"text": "$$\n{ \\mathcal { L } } _ { \\mathrm { s i m p l e } } = E _ { x _ { 0 } \\sim q ( x _ { 0 } | c ) , t \\sim [ 1 , T ] } [ \\| x _ { 0 } - G ( x _ { t } , t , c ) \\| _ { 2 } ^ { 2 } ]\n$$",
|
| 399 |
+
"text_format": "latex",
|
| 400 |
+
"bbox": [
|
| 401 |
+
334,
|
| 402 |
+
597,
|
| 403 |
+
663,
|
| 404 |
+
616
|
| 405 |
+
],
|
| 406 |
+
"page_idx": 3
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"type": "text",
|
| 410 |
+
"text": "Geometric losses. In the motion domain, generative networks are standardly regularized using geometric losses Petrovich et al. (2021); Shi et al. (2020). These losses enforce physical properties and prevent artifacts, encouraging natural and coherent motion. In this work we experiment with three common geometric losses that regulate (1) positions (in case we predict rotations), (2) foot contact, and (3) velocities. ",
|
| 411 |
+
"bbox": [
|
| 412 |
+
173,
|
| 413 |
+
623,
|
| 414 |
+
825,
|
| 415 |
+
693
|
| 416 |
+
],
|
| 417 |
+
"page_idx": 3
|
| 418 |
+
},
|
| 419 |
+
{
|
| 420 |
+
"type": "equation",
|
| 421 |
+
"img_path": "images/83dc1a4ccd51b4d9ef24b1014f16302b9d7f9ed28ca5ce6413072eeed5daf8bd.jpg",
|
| 422 |
+
"text": "$$\n\\mathcal { L } _ { \\mathrm { p o s } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\| F K ( x _ { 0 } ^ { i } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 } ,\n$$",
|
| 423 |
+
"text_format": "latex",
|
| 424 |
+
"bbox": [
|
| 425 |
+
366,
|
| 426 |
+
685,
|
| 427 |
+
630,
|
| 428 |
+
728
|
| 429 |
+
],
|
| 430 |
+
"page_idx": 3
|
| 431 |
+
},
|
| 432 |
+
{
|
| 433 |
+
"type": "equation",
|
| 434 |
+
"img_path": "images/82159991daf78998072aa721251cf1cc292514d5cc6571d9271180114bcd0601.jpg",
|
| 435 |
+
"text": "$$\n\\mathcal { L } _ { \\mathrm { f o o t } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( F K ( \\hat { x } _ { 0 } ^ { i + 1 } ) - F K ( \\hat { x } _ { 0 } ^ { i } ) ) \\cdot f _ { i } \\| _ { 2 } ^ { 2 } ,\n$$",
|
| 436 |
+
"text_format": "latex",
|
| 437 |
+
"bbox": [
|
| 438 |
+
323,
|
| 439 |
+
739,
|
| 440 |
+
673,
|
| 441 |
+
782
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 3
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "equation",
|
| 447 |
+
"img_path": "images/8b86f289003d23a8b1f4d9c86a75e25912e7dcd00b9e615165e4137137ac3796.jpg",
|
| 448 |
+
"text": "$$\n\\mathcal { L } _ { \\mathrm { v e l } } = \\frac { 1 } { N - 1 } \\sum _ { i = 1 } ^ { N - 1 } \\| ( x _ { 0 } ^ { i + 1 } - x _ { 0 } ^ { i } ) - ( \\hat { x } _ { 0 } ^ { i + 1 } - \\hat { x } _ { 0 } ^ { i } ) \\| _ { 2 } ^ { 2 }\n$$",
|
| 449 |
+
"text_format": "latex",
|
| 450 |
+
"bbox": [
|
| 451 |
+
330,
|
| 452 |
+
794,
|
| 453 |
+
668,
|
| 454 |
+
837
|
| 455 |
+
],
|
| 456 |
+
"page_idx": 3
|
| 457 |
+
},
|
| 458 |
+
{
|
| 459 |
+
"type": "text",
|
| 460 |
+
"text": "In case we predict joint rotations, $F K ( \\cdot )$ denotes the forward kinematic function converting joint rotations into joint positions (otherwise, it denotes the identity function). $f _ { i } \\in \\{ 0 , 1 \\} ^ { J }$ is the binary foot contact mask for each frame $i$ . Relevant only to feet, it indicates whether they touch the ground, and are set according to binary ground truth data (Shi et al., 2020). In essence, it mitigates the foot-sliding effect by nullifying velocities when touching the ground. Overall, our training loss is ${ \\mathcal { L } } = { \\mathcal { L } } _ { \\mathrm { s i m p l e } } + \\lambda _ { \\mathrm { p o s } } { \\bar { \\mathcal { L } } } _ { \\mathrm { p o s } } + { \\bar { \\lambda _ { \\mathrm { v e l } } } } { \\bar { \\mathcal { L } } } _ { \\mathrm { v e l } } + \\lambda _ { \\mathrm { f o o t } } { \\mathcal { L } } _ { \\mathrm { f o o t } }$ . ",
|
| 461 |
+
"bbox": [
|
| 462 |
+
173,
|
| 463 |
+
839,
|
| 464 |
+
825,
|
| 465 |
+
926
|
| 466 |
+
],
|
| 467 |
+
"page_idx": 3
|
| 468 |
+
},
|
| 469 |
+
{
|
| 470 |
+
"type": "image",
|
| 471 |
+
"img_path": "images/3c301f5cc7f24dee220ca678e53e2e875ab4f7d26b316945108c82aaf3924b98.jpg",
|
| 472 |
+
"image_caption": [
|
| 473 |
+
"Figure 3: Editing applications. Light blue frames represent motion input and bronze frames are the generated motion. Motion in-betweening (left+center) can be performed conditioned on text or without condition by the same model. Specific body part editing using text is demonstrated on the right: the lower body joints are fixed to the input motion while the upper body is altered to fit the input text prompt. "
|
| 474 |
+
],
|
| 475 |
+
"image_footnote": [],
|
| 476 |
+
"bbox": [
|
| 477 |
+
176,
|
| 478 |
+
97,
|
| 479 |
+
800,
|
| 480 |
+
309
|
| 481 |
+
],
|
| 482 |
+
"page_idx": 4
|
| 483 |
+
},
|
| 484 |
+
{
|
| 485 |
+
"type": "text",
|
| 486 |
+
"text": "Model. Our model is illustrated in Figure 2. We implement $G$ with a straightforward transformer (Vaswani et al., 2017) encoder-only architecture. The transformer architecture is temporally aware, enabling learning and generating variable-length motions, and is well-proven for the motion domain (Petrovich et al., 2021; Duan et al., 2021; Aksan et al., 2021). The noise time-step $t$ and the condition code $c$ are each projected to the transformer dimension by separate feed-forward networks, then summed to yield the token $z _ { t k }$ . Each frame of the noised input $x _ { t }$ is linearly projected into the transformer dimension and summed with a standard positional embedding. $z _ { t k }$ and the projected frames are then fed to the encoder. Excluding the first output token (corresponding to $z _ { t k }$ ), the encoder result is projected back to the original motion dimensions, and serves as the prediction $\\scriptstyle { \\hat { x } } _ { 0 }$ . We implement text-to-motion by encoding the text prompt to $c$ with CLIP (Radford et al., 2021) text encoder, and action-to-motion with learned embeddings per class. ",
|
| 487 |
+
"bbox": [
|
| 488 |
+
173,
|
| 489 |
+
397,
|
| 490 |
+
825,
|
| 491 |
+
550
|
| 492 |
+
],
|
| 493 |
+
"page_idx": 4
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"type": "text",
|
| 497 |
+
"text": "Sampling from $p ( x _ { 0 } | c )$ is done in an iterative manner, according to Ho et al. (2020). In every time step $t$ we predict the clean sample $\\hat { \\boldsymbol { x } } _ { 0 } = \\boldsymbol { G } ( \\boldsymbol { x } _ { t } , t , \\boldsymbol { c } )$ and noise it back to $x _ { t - 1 }$ . This is repeated from $t = T$ until $x _ { 0 }$ is achieved (Figure 2 right). We train our model $G$ using classifier-free guidance (Ho & Salimans, 2022). In practice, $G$ learns both the conditioned and the unconditioned distributions by randomly setting $c = \\emptyset$ for $1 0 \\%$ of the samples, such that $G ( x _ { t } , t , \\emptyset )$ approximates $p ( x _ { 0 } )$ . Then, when sampling $G$ we can trade-off diversity and fidelity by interpolating or even extrapolating the two variants using $\\boldsymbol { s } \\colon G _ { s } ( x _ { t } , t , c ) = G ( x _ { t } , \\dot { t , } \\emptyset ) + s \\cdot ( \\dot { G } ( x _ { t } , t , c ) \\bar { - } G ( x _ { t } , t , \\emptyset ) )$ . ",
|
| 498 |
+
"bbox": [
|
| 499 |
+
173,
|
| 500 |
+
558,
|
| 501 |
+
825,
|
| 502 |
+
656
|
| 503 |
+
],
|
| 504 |
+
"page_idx": 4
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"type": "text",
|
| 508 |
+
"text": "Editing. We enable motion in-betweening in the temporal domain, and body part editing in the spatial domain, by adapting diffusion inpainting to motion data. Editing is done only during sampling, without any training involved. Given a subset of the motion sequence inputs, when sampling the model (Figure 2 right), at each iteration we overwrite $\\scriptstyle { \\hat { x } } _ { 0 }$ with the input part of the motion. This encourages the generation to remain coherent to original input, while completing the missing parts. In the temporal setting, the prefix and suffix frames of the motion sequence are the input, and we solve a motion in-betweening problem (Harvey et al., 2020). Editing can be done either conditionally or unconditionally (by setting $c = \\emptyset$ ). In the spatial setting, we show that body parts can be re-synthesized according to a condition $c$ while keeping the rest intact, through the use of the same completion technique. ",
|
| 509 |
+
"bbox": [
|
| 510 |
+
173,
|
| 511 |
+
661,
|
| 512 |
+
825,
|
| 513 |
+
801
|
| 514 |
+
],
|
| 515 |
+
"page_idx": 4
|
| 516 |
+
},
|
| 517 |
+
{
|
| 518 |
+
"type": "text",
|
| 519 |
+
"text": "4 EXPERIMENTS ",
|
| 520 |
+
"text_level": 1,
|
| 521 |
+
"bbox": [
|
| 522 |
+
176,
|
| 523 |
+
815,
|
| 524 |
+
326,
|
| 525 |
+
830
|
| 526 |
+
],
|
| 527 |
+
"page_idx": 4
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"type": "text",
|
| 531 |
+
"text": "We implement MDM for three motion generation tasks: Text-to-Motion(4.1), Action-to-Motion(4.2) and unconditioned generation(5.2. Each sub-section reviews the data and metrics of the used benchmarks, provides implementation details, and presents qualitative and quantitative results. Then, we show implementations of motion in-betweening (both conditioned and unconditioned) and bodypart editing by adapting diffusion inpainting to motion (5.1). Our models have been trained with $T = 1 0 0 0$ noising steps and a cosine noise schedule. In Appendix A.1, we experiment with different values of $T$ . All of them have been trained on a single NVIDIA GeForce RTX 2080 Ti GPU for a period of about 3 days. ",
|
| 532 |
+
"bbox": [
|
| 533 |
+
174,
|
| 534 |
+
840,
|
| 535 |
+
825,
|
| 536 |
+
924
|
| 537 |
+
],
|
| 538 |
+
"page_idx": 4
|
| 539 |
+
},
|
| 540 |
+
{
|
| 541 |
+
"type": "text",
|
| 542 |
+
"text": "",
|
| 543 |
+
"bbox": [
|
| 544 |
+
173,
|
| 545 |
+
103,
|
| 546 |
+
823,
|
| 547 |
+
132
|
| 548 |
+
],
|
| 549 |
+
"page_idx": 5
|
| 550 |
+
},
|
| 551 |
+
{
|
| 552 |
+
"type": "text",
|
| 553 |
+
"text": "4.1 TEXT-TO-MOTION ",
|
| 554 |
+
"text_level": 1,
|
| 555 |
+
"bbox": [
|
| 556 |
+
174,
|
| 557 |
+
145,
|
| 558 |
+
343,
|
| 559 |
+
159
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 5
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "text",
|
| 565 |
+
"text": "Text-to-motion is the task of generating motion given an input text prompt. The output motion is expected to be both implementing the textual description, and a valid sample from the data distribution (i.e. adhering to general human abilities and the rules of physics). In addition, for each text prompt, we also expect a distribution of motions matching it, rather than just a single result. We evaluate our model using two leading benchmarks - KIT (Plappert et al., 2016) and HumanML3D (Guo et al., 2022a), over the set of metrics suggested by Guo et al. (2022a): R-precision and Multimodal-Dist measure the relevancy of the generated motions to the input prompts, $F I D$ measures the dissimilarity between the generated and ground truth distributions (in latent space), Diversity measures the variability in the resulting motion distribution, and MultiModality is the average variance given a single text prompt. For the full implementation of the metrics, please refer to Guo et al. (2022a). We use HumanML3D as a platform to compare different backbones of our model, discovering that the diffusion framework is relatively agnostic to this attribute. In addition, we conduct a user study comparing our model to current art and ground truth motions. ",
|
| 566 |
+
"bbox": [
|
| 567 |
+
174,
|
| 568 |
+
165,
|
| 569 |
+
825,
|
| 570 |
+
345
|
| 571 |
+
],
|
| 572 |
+
"page_idx": 5
|
| 573 |
+
},
|
| 574 |
+
{
|
| 575 |
+
"type": "text",
|
| 576 |
+
"text": "Data. HumanML3D is a recent dataset, textually re-annotating motion capture from the AMASS (Mahmood et al., 2019) and HumanAct12 (Guo et al., 2020) collections. It contains 14, 616 motions annotated by 44, 970 textual descriptions. In addition, it suggests a redundant data representation including a concatenation of root velocity, joint positions, joint velocities, joint rotations and the foot contact binary labels. We also use in this section the same representation for the KIT dataset, brought by the same publishers. Although limited in the number (3, 911) and the diversity of samples, most of the text-to-motion research is based on KIT, hence we view it as important to evaluate using it as well. ",
|
| 577 |
+
"bbox": [
|
| 578 |
+
174,
|
| 579 |
+
353,
|
| 580 |
+
825,
|
| 581 |
+
464
|
| 582 |
+
],
|
| 583 |
+
"page_idx": 5
|
| 584 |
+
},
|
| 585 |
+
{
|
| 586 |
+
"type": "text",
|
| 587 |
+
"text": "Implementation. In addition to our Transformer encoder-only backbone (Section 3), we experiment MDM with three more backbones: (1) Transformer decoder injects $z _ { t k }$ through the cross-attention layer, instead of as an input token. (2) Transformer decoder $^ +$ input token, where $z _ { t k }$ is injected both ways, (3) GRU (Cho et al., 2014) concatenate $z _ { t k }$ to each input frame , and (4) a U-net adaptation for motion data (Table 1). U-net is adapted to motion by replacing the 2D convolution filters with 1D convolutions in the temporal axis, such that joints are considered as channels. This is due to the irregular behavior of the joint axis. Our models were trained with batch size 64, 8 layers (except GRU that was optimal at 2), and latent dimension 512. To encode the text we use a frozen CLIPViT-B/32 model. In addition, we experiment with replacing CLIP with sentence-BERT (Reimers & Gurevych, 2019) - a BERT-based (Devlin et al., 2019) text encoder. The full details can be found in our published code1 and Appendix C. Each model was trained for $5 0 0 K$ steps, after which a checkpoint was chosen that minimizes the FID metric to be reported. Since foot contact and joint locations are explicitly represented in HumanML3D, we don’t apply geometric losses in this section. We evaluate our models with guidance-scale $s = 2 . 5$ which provides a diversity-fidelity sweet spot (Figure 4). ",
|
| 588 |
+
"bbox": [
|
| 589 |
+
174,
|
| 590 |
+
472,
|
| 591 |
+
825,
|
| 592 |
+
679
|
| 593 |
+
],
|
| 594 |
+
"page_idx": 5
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"type": "text",
|
| 598 |
+
"text": "Quantitative evaluation. We evaluate and compare our models to current art (JL2P Ahuja & Morency (2019), Text2Gesture (Bhattacharya et al., 2021), and T2M (Guo et al., 2022a)) with the metrics suggested by Guo et al. (2022a). As can be seen, MDM achieves state-of-the-art results in FID, Diversity, and MultiModality, indicating high diversity per input text prompt, and high-quality samples, as can also be seen qualitatively in Figure 1. ",
|
| 599 |
+
"bbox": [
|
| 600 |
+
174,
|
| 601 |
+
686,
|
| 602 |
+
823,
|
| 603 |
+
756
|
| 604 |
+
],
|
| 605 |
+
"page_idx": 5
|
| 606 |
+
},
|
| 607 |
+
{
|
| 608 |
+
"type": "text",
|
| 609 |
+
"text": "User study. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side view, with both samples generated from the same text prompt randomly sampled from the KIT test set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even preferred over ground truth samples in $4 2 . 3 \\%$ of the cases. ",
|
| 610 |
+
"bbox": [
|
| 611 |
+
174,
|
| 612 |
+
763,
|
| 613 |
+
825,
|
| 614 |
+
847
|
| 615 |
+
],
|
| 616 |
+
"page_idx": 5
|
| 617 |
+
},
|
| 618 |
+
{
|
| 619 |
+
"type": "text",
|
| 620 |
+
"text": "4.2 ACTION-TO-MOTION ",
|
| 621 |
+
"text_level": 1,
|
| 622 |
+
"bbox": [
|
| 623 |
+
176,
|
| 624 |
+
848,
|
| 625 |
+
361,
|
| 626 |
+
861
|
| 627 |
+
],
|
| 628 |
+
"page_idx": 5
|
| 629 |
+
},
|
| 630 |
+
{
|
| 631 |
+
"type": "text",
|
| 632 |
+
"text": "Action-to-motion is the task of generating motion given an input action class, represented by a scalar. The output motion should faithfully animate the input action, and at the same time be natural and reflect the distribution of the dataset on which the model is trained. Two dataset are commonly used to evaluate action-to-motion models: HumanAct12 (Guo et al., 2020) and UESTC (Ji et al., 2018). ",
|
| 633 |
+
"bbox": [
|
| 634 |
+
176,
|
| 635 |
+
868,
|
| 636 |
+
823,
|
| 637 |
+
922
|
| 638 |
+
],
|
| 639 |
+
"page_idx": 5
|
| 640 |
+
},
|
| 641 |
+
{
|
| 642 |
+
"type": "table",
|
| 643 |
+
"img_path": "images/a8390e456fc75cdd5ea1504b7990139329e00b053fc9fa17c46ccd72412e5cff.jpg",
|
| 644 |
+
"table_caption": [
|
| 645 |
+
"Table 1: Quantitative results on the HumanML3D test set. All methods use the real motion length from the ground truth. $\\cdot _ { } ,$ means results are better if the metric is closer to the real distribution. We run all the evaluation 20 times (except MultiModality runs 5 times) and $\\pm$ indicates the $9 5 \\%$ confidence interval. Bold indicates best result. "
|
| 646 |
+
],
|
| 647 |
+
"table_footnote": [],
|
| 648 |
+
"table_body": "<table><tr><td>Method</td><td>RPrecision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td><td>Multimodality↑</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td></td></tr><tr><td>JL2P</td><td>0.486±.002</td><td>11.02±.046</td><td>5.296±.008</td><td>7.676±.058</td><td>1</td></tr><tr><td>Text2Gesture</td><td>0.345±.002</td><td>7.664±.030</td><td>6.030±.008</td><td>6.409±.071</td><td>1</td></tr><tr><td>T2M</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MDM (ours)</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>+ sent-BERT</td><td>0.609±.006</td><td>0.586±.036</td><td>5.504±.03</td><td>9.666±.095</td><td>2.707±.188</td></tr><tr><td>MDM (decoder)</td><td>0.608±.005</td><td>0.767±.085</td><td>5.507±.020</td><td>9.176±.070</td><td>2.927±.125</td></tr><tr><td>+ input token</td><td>0.621±.005</td><td>0.567±.051</td><td>5.424±.022</td><td>9.425±.060</td><td>2.834±.095</td></tr><tr><td>MDM (U-net)</td><td>0.603±.006</td><td>1.137±.008</td><td>5.629±.032</td><td>8.958±.098</td><td>2.636±.214</td></tr><tr><td>MDM (GRU)</td><td>0.645±.005</td><td>4.569±.150</td><td>5.325±.026</td><td>7.688±.082</td><td>1.2646±.024</td></tr></table>",
|
| 649 |
+
"bbox": [
|
| 650 |
+
176,
|
| 651 |
+
88,
|
| 652 |
+
818,
|
| 653 |
+
271
|
| 654 |
+
],
|
| 655 |
+
"page_idx": 6
|
| 656 |
+
},
|
| 657 |
+
{
|
| 658 |
+
"type": "table",
|
| 659 |
+
"img_path": "images/b4c373214b840b1b1680f5c0558314e7810c86b9be268ce24730e3e98f5a8f4c.jpg",
|
| 660 |
+
"table_caption": [
|
| 661 |
+
"Table 2: Quantitative results on the KIT test set. "
|
| 662 |
+
],
|
| 663 |
+
"table_footnote": [],
|
| 664 |
+
"table_body": "<table><tr><td>Method</td><td>R Precision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td><td>Multimodality↑</td></tr><tr><td>Real</td><td>0.779±.006</td><td>0.031±.004</td><td>2.788±.012</td><td>11.08±.097</td><td>-</td></tr><tr><td>JL2P</td><td>0.483±.005</td><td>6.545±.072</td><td>5.147±.030</td><td>9.073±.100</td><td>-</td></tr><tr><td>Text2Gesture</td><td>0.338±.005</td><td>12.12±.183</td><td>6.964±.029</td><td>9.334±.079</td><td></td></tr><tr><td>T2M</td><td>0.693±.007</td><td>2.770±.109</td><td>3.401±.008</td><td>10.91±.119</td><td>1.482±.065</td></tr><tr><td>MDM (ours)</td><td>0.396±.004</td><td>0.497±.021</td><td>9.191±.022</td><td>10.847±.109</td><td>1.907±.214</td></tr></table>",
|
| 665 |
+
"bbox": [
|
| 666 |
+
176,
|
| 667 |
+
353,
|
| 668 |
+
818,
|
| 669 |
+
474
|
| 670 |
+
],
|
| 671 |
+
"page_idx": 6
|
| 672 |
+
},
|
| 673 |
+
{
|
| 674 |
+
"type": "text",
|
| 675 |
+
"text": "We evaluate our model using the set of metrics suggested by Guo et al. (2020), namely Frechet In- ´ ception Distance (FID), action recognition accuracy, diversity and multimodality. The combination of these metrics makes a good measure of the realism and diversity of generated motions. ",
|
| 676 |
+
"bbox": [
|
| 677 |
+
174,
|
| 678 |
+
535,
|
| 679 |
+
825,
|
| 680 |
+
577
|
| 681 |
+
],
|
| 682 |
+
"page_idx": 6
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"type": "text",
|
| 686 |
+
"text": "Data. HumanAct12 (Guo et al., 2020) offers approximately 1200 motion clips, organized into 12 action categories, with 47 to 218 samples per label. UESTC (Ji et al., 2018) consists of 40 action classes, 40 subjects and 25K samples, and is split to train and test. We adhere to the cross-subject testing protocol used by current works, with 225-345 samples per action class. For both datasets we use the sequences provided by Petrovich et al. (2021). ",
|
| 687 |
+
"bbox": [
|
| 688 |
+
173,
|
| 689 |
+
583,
|
| 690 |
+
825,
|
| 691 |
+
654
|
| 692 |
+
],
|
| 693 |
+
"page_idx": 6
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "image",
|
| 697 |
+
"img_path": "images/d7a9d635bd4c7d3904d77117eb65b18a0c3237c9cb38308379a700f6d61b9f3c.jpg",
|
| 698 |
+
"image_caption": [
|
| 699 |
+
"Figure 4: (a) Text-to-motion user study for the KIT dataset. Each bar represents the preference rate of MDM over the compared model. MDM was preferred over the other models in most of the time, and $4 2 . 3 \\%$ of the cases even over ground truth samples. The dashed line marks $5 0 \\%$ . (b) Guidance-scale sweep for HumanML3D dataset. $F I D$ (lower is better) and $R$ -precision (higher is better) metrics as a function of the scale $s$ , draws an accuracy-fidelity sweet spot around $s = 2 . 5$ . "
|
| 700 |
+
],
|
| 701 |
+
"image_footnote": [],
|
| 702 |
+
"bbox": [
|
| 703 |
+
253,
|
| 704 |
+
704,
|
| 705 |
+
727,
|
| 706 |
+
852
|
| 707 |
+
],
|
| 708 |
+
"page_idx": 6
|
| 709 |
+
},
|
| 710 |
+
{
|
| 711 |
+
"type": "table",
|
| 712 |
+
"img_path": "images/71e7b9aeb113ee6e6b17585a4c1630b39f58d6ffe56a9d3d7cc375c2ed3f08c2.jpg",
|
| 713 |
+
"table_caption": [
|
| 714 |
+
"Table 3: Evaluation of action-to-motion on the HumanAct12 dataset. Our model leads the board in three out of four metrics. Ground-truth evaluation results are slightly different for each of the works, due to implementation differences, such as python package versions. It is important to assess the diversity and multimodality of each model using its own ground-truth results, as they are measured by their distance from GT. We show the GT metrics measured by our model and by the leading compared work, INR (Cervantes et al., 2022). Bold indicates best result, underline indicates second best, $\\pm$ indicates $9 5 \\%$ confidence interval, indicates that closer to real is better. "
|
| 715 |
+
],
|
| 716 |
+
"table_footnote": [],
|
| 717 |
+
"table_body": "<table><tr><td>Method</td><td>FID↓</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real (INR)</td><td>0.020±.010</td><td>0.997±.001</td><td>6.850±.050</td><td>2.450±.040</td></tr><tr><td>Real (ours)</td><td>0.050±.000</td><td>0.990±.000</td><td>6.880±.020</td><td>2.590±.010</td></tr><tr><td>Action2Motion (2020)</td><td>0.338±.015</td><td>0.917±.003</td><td>6.879±.066</td><td>2.511±.023</td></tr><tr><td>ACTOR (2021)</td><td>0.120±.000</td><td>0.955±.008</td><td>6.840±.030</td><td>2.530±.020</td></tr><tr><td>INR (2022)</td><td>0.088±.004</td><td>0.973±.001</td><td>6.881±.048</td><td>2.569±.040</td></tr><tr><td>MDM (ours)</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>w/o foot contact</td><td>0.080±.000</td><td>0.990±.000</td><td>6.810±.010</td><td>2.580±.010</td></tr></table>",
|
| 718 |
+
"bbox": [
|
| 719 |
+
192,
|
| 720 |
+
101,
|
| 721 |
+
808,
|
| 722 |
+
262
|
| 723 |
+
],
|
| 724 |
+
"page_idx": 7
|
| 725 |
+
},
|
| 726 |
+
{
|
| 727 |
+
"type": "table",
|
| 728 |
+
"img_path": "images/ec3d68c861da7d767ff2deaf102626966cf953b43a64342352012a725813007f.jpg",
|
| 729 |
+
"table_caption": [
|
| 730 |
+
"Table 4: Evaluation of action-to-motion on the UESTC dataset. The performance improvement with our model shows a clear gap from state-of-the-art. Bold indicates best result, underline indicates second best, $\\pm$ indicates $9 5 \\%$ confidence interval, indicates that closer to real is better. "
|
| 731 |
+
],
|
| 732 |
+
"table_footnote": [],
|
| 733 |
+
"table_body": "<table><tr><td>Method</td><td>FIDtrain </td><td>FIDtest</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real</td><td>2.92±.26</td><td>2.79±.29</td><td>0.988±.001</td><td>33.34±.320</td><td>14.16±.06</td></tr><tr><td>ACTOR (2021)</td><td>20.49±2.31</td><td>23.43±2.20</td><td>0.911±.003</td><td>31.96±.33</td><td>14.52±.09</td></tr><tr><td>INR (2022) (best variation)</td><td>9.55±.06</td><td>15.00±.09</td><td>0.941±.001</td><td>31.59±.19</td><td>14.68±.07</td></tr><tr><td>MDM (ours)</td><td>9.98±1.33</td><td>12.81±1.46</td><td>0.950±.000</td><td>33.02±.28</td><td>14.26±.12</td></tr><tr><td>w/o foot contact</td><td>9.69±.81</td><td>13.08±2.32</td><td>0.960±.000</td><td>33.10±.29</td><td>14.06±.05</td></tr></table>",
|
| 734 |
+
"bbox": [
|
| 735 |
+
176,
|
| 736 |
+
386,
|
| 737 |
+
825,
|
| 738 |
+
502
|
| 739 |
+
],
|
| 740 |
+
"page_idx": 7
|
| 741 |
+
},
|
| 742 |
+
{
|
| 743 |
+
"type": "text",
|
| 744 |
+
"text": "Implementation. The implementation presented in Figure 2 holds for all the variations of our work. In the case of action-to-motion, the only change would be the substitution of the text embedding by an action embedding. Since action is represented by a scalar, its embedding is fairly simple; each input action class scalar is converted into a learned embedding of the transformer dimension. ",
|
| 745 |
+
"bbox": [
|
| 746 |
+
174,
|
| 747 |
+
583,
|
| 748 |
+
823,
|
| 749 |
+
638
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 7
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "text",
|
| 755 |
+
"text": "The experiments have been run with batch size 64, a latent dimension of 512, and an encodertransformer architecture. Training on HumanAct12 and UESTC has been carried out for $7 5 0 K$ and $2 M$ steps respectively. In our tables we display the evaluation of the checkpoint that minimizes the FID metric. ",
|
| 756 |
+
"bbox": [
|
| 757 |
+
174,
|
| 758 |
+
645,
|
| 759 |
+
825,
|
| 760 |
+
702
|
| 761 |
+
],
|
| 762 |
+
"page_idx": 7
|
| 763 |
+
},
|
| 764 |
+
{
|
| 765 |
+
"type": "text",
|
| 766 |
+
"text": "Quantitative evaluation. Tables 3 and 4 reflect MDM’s performance on the HumanAct12 and UESTC datasets respectively. We conduct 20 evaluations, with 1000 samples in each, and report their average and a $9 5 \\%$ confidence interval. We test two variations, with and without foot contact loss. Full ablation study for geometric losses is presented in Appendix A.2. Our model leads the board for both datasets. The variation with no foot contact loss attains slightly better results; nevertheless, as shown in our supplementary video, the contribution of foot contact loss to the quality of results is important, and without it we witness artifacts such as shakiness and unnatural gestures. ",
|
| 767 |
+
"bbox": [
|
| 768 |
+
173,
|
| 769 |
+
708,
|
| 770 |
+
825,
|
| 771 |
+
806
|
| 772 |
+
],
|
| 773 |
+
"page_idx": 7
|
| 774 |
+
},
|
| 775 |
+
{
|
| 776 |
+
"type": "text",
|
| 777 |
+
"text": "5 ADDITIONAL APPLICATIONS ",
|
| 778 |
+
"text_level": 1,
|
| 779 |
+
"bbox": [
|
| 780 |
+
176,
|
| 781 |
+
821,
|
| 782 |
+
442,
|
| 783 |
+
837
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 7
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "5.1 MOTION EDITING ",
|
| 790 |
+
"text_level": 1,
|
| 791 |
+
"bbox": [
|
| 792 |
+
174,
|
| 793 |
+
847,
|
| 794 |
+
339,
|
| 795 |
+
862
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 7
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "text",
|
| 801 |
+
"text": "In this section we implement two motion editing applications - in-betweening and body part editing, both using the same approach in the temporal and spatial domains correspondingly. For inbetweening, we fix the first and last $2 5 \\%$ of the motion, leaving the model to generate the remaining $5 0 \\%$ in the middle. For body part editing, we fix the joints we don’t want to edit and leave the ",
|
| 802 |
+
"bbox": [
|
| 803 |
+
174,
|
| 804 |
+
868,
|
| 805 |
+
823,
|
| 806 |
+
924
|
| 807 |
+
],
|
| 808 |
+
"page_idx": 7
|
| 809 |
+
},
|
| 810 |
+
{
|
| 811 |
+
"type": "table",
|
| 812 |
+
"img_path": "images/089bb7a3f1346cccaada3c6555301bd17b789d1dbb439201a15aa8001fc8c5cb.jpg",
|
| 813 |
+
"table_caption": [
|
| 814 |
+
"model to generate the rest. In particular, we experiment with editing the upper body joints only. In figure 3 we show that in both cases, using the method described in Section 3 generates smooth motions that adhere both to the fixed part of the motion and the condition (if one was given). ",
|
| 815 |
+
"Table 5: Evaluation of unconstrained synthesis on the HumanAct12 dataset. We test MDM in the challenging unconstrained setting, and compare with MoDi (Raab et al., 2022), a work that was specially designed for such setting. We demonstrate that in addition to being able to support any condition, we can achieve plausible results in the unconstrained setting. Bold indicates best result. "
|
| 816 |
+
],
|
| 817 |
+
"table_footnote": [],
|
| 818 |
+
"table_body": "<table><tr><td>Method</td><td>FID↓</td><td>KID↓</td><td>Precision↑ Recall↑</td><td>Diversity↑</td></tr><tr><td>ACTOR (2021)</td><td>48.80</td><td>0.53</td><td>0.72, 0.74</td><td>14.10</td></tr><tr><td>MoDi (2022)</td><td>13.03</td><td>0.12</td><td>0.71, 0.81</td><td>17.57</td></tr><tr><td>MDM (ours)</td><td>31.92</td><td>0.36</td><td>0.66,0.62</td><td>17.00</td></tr></table>",
|
| 819 |
+
"bbox": [
|
| 820 |
+
227,
|
| 821 |
+
159,
|
| 822 |
+
771,
|
| 823 |
+
244
|
| 824 |
+
],
|
| 825 |
+
"page_idx": 8
|
| 826 |
+
},
|
| 827 |
+
{
|
| 828 |
+
"type": "text",
|
| 829 |
+
"text": "5.2 UNCONSTRAINED SYNTHESIS ",
|
| 830 |
+
"text_level": 1,
|
| 831 |
+
"bbox": [
|
| 832 |
+
176,
|
| 833 |
+
318,
|
| 834 |
+
421,
|
| 835 |
+
332
|
| 836 |
+
],
|
| 837 |
+
"page_idx": 8
|
| 838 |
+
},
|
| 839 |
+
{
|
| 840 |
+
"type": "text",
|
| 841 |
+
"text": "The challenging task of unconstrained synthesis has been studied by only a few (Holden et al., 2016; Raab et al., 2022). In the presence of data labeling, e.g., action classes or text description, the labels work as a supervising factor, and facilitate a structured latent space for the training network. The lack of labeling make training more difficult. The human motion field possesses rich unlabeled datasets (Adobe Systems Inc., 2021), and the ability to train on top of them is an advantage. Daring to test MDM in the challenging unconstrained setting, we follow MoDi(Raab et al., 2022) for evaluation. We use the metrics they suggest (FID, KID, precision/recall and multimodality), and run on an unconstrained version of the HumanAct12 (Guo et al., 2020) dataset. ",
|
| 842 |
+
"bbox": [
|
| 843 |
+
174,
|
| 844 |
+
338,
|
| 845 |
+
825,
|
| 846 |
+
449
|
| 847 |
+
],
|
| 848 |
+
"page_idx": 8
|
| 849 |
+
},
|
| 850 |
+
{
|
| 851 |
+
"type": "text",
|
| 852 |
+
"text": "Data. Although annotated, we use HumanAct12 (see Section 4.2) in an unconstrained fashion, ignoring its labels. The choice of HumanAct12 rather than a dataset with no labels (e.g., Mixamo (Adobe Systems Inc., 2021)), is for compatibility with previous publications. ",
|
| 853 |
+
"bbox": [
|
| 854 |
+
174,
|
| 855 |
+
455,
|
| 856 |
+
825,
|
| 857 |
+
498
|
| 858 |
+
],
|
| 859 |
+
"page_idx": 8
|
| 860 |
+
},
|
| 861 |
+
{
|
| 862 |
+
"type": "text",
|
| 863 |
+
"text": "Implementation. Our model uses the same architecture for all forms of conditioning, as well as for the unconstrained setting. The only change to the structure shown in Figure 2, is the removal of the conditional input, such that $z _ { t k }$ is composed of the projection of $t$ only. To simulate an unconstrained behavior, ACTOR Petrovich et al. (2021) has been trained by (Raab et al., 2022) with a labeling of one class to all motions. ",
|
| 864 |
+
"bbox": [
|
| 865 |
+
174,
|
| 866 |
+
505,
|
| 867 |
+
825,
|
| 868 |
+
574
|
| 869 |
+
],
|
| 870 |
+
"page_idx": 8
|
| 871 |
+
},
|
| 872 |
+
{
|
| 873 |
+
"type": "text",
|
| 874 |
+
"text": "Quantitative evaluation. The results of our evaluation are shown in table 5. We demonstrate superiority over works that were not designed for an unconstrained setting, and get closer to MoDi (Raab et al., 2022). MoDi is carefully molded for unconstrained settings, while our work can be applied to any (or no) constrain, and also provides editing capabilities. ",
|
| 875 |
+
"bbox": [
|
| 876 |
+
174,
|
| 877 |
+
582,
|
| 878 |
+
825,
|
| 879 |
+
637
|
| 880 |
+
],
|
| 881 |
+
"page_idx": 8
|
| 882 |
+
},
|
| 883 |
+
{
|
| 884 |
+
"type": "text",
|
| 885 |
+
"text": "6 DISCUSSION ",
|
| 886 |
+
"text_level": 1,
|
| 887 |
+
"bbox": [
|
| 888 |
+
174,
|
| 889 |
+
651,
|
| 890 |
+
310,
|
| 891 |
+
667
|
| 892 |
+
],
|
| 893 |
+
"page_idx": 8
|
| 894 |
+
},
|
| 895 |
+
{
|
| 896 |
+
"type": "text",
|
| 897 |
+
"text": "We have presented MDM, a method that lends itself to various human motion generation tasks. MDM is an untypical classifier-free diffusion model, featuring a transformer-encoder backbone, and predicting the signal, rather than the noise. This yields both a lightweight model, that is unburdening to train, and an accurate one, gaining much from the applicable geometric losses. Our experiments show superiority in conditioned generation, but also that this approach is not very sensitive to the choice of architecture. A notable limitation of the diffusion approach is the long inference time, requiring about 1000 forward passes for a single result. Since our motion model is small anyway, using dimensions order of magnitude smaller than images, our inference time shifts from less than a second to only about a minute, which is an acceptable compromise. As diffusion models continue to evolve, besides better compute, in the future we would be interested in seeing how to incorporate better control into the generation process and widen the options for applications even further. ",
|
| 898 |
+
"bbox": [
|
| 899 |
+
174,
|
| 900 |
+
676,
|
| 901 |
+
825,
|
| 902 |
+
829
|
| 903 |
+
],
|
| 904 |
+
"page_idx": 8
|
| 905 |
+
},
|
| 906 |
+
{
|
| 907 |
+
"type": "text",
|
| 908 |
+
"text": "ACKNOWLEDGEMENTS ",
|
| 909 |
+
"text_level": 1,
|
| 910 |
+
"bbox": [
|
| 911 |
+
176,
|
| 912 |
+
852,
|
| 913 |
+
367,
|
| 914 |
+
866
|
| 915 |
+
],
|
| 916 |
+
"page_idx": 8
|
| 917 |
+
},
|
| 918 |
+
{
|
| 919 |
+
"type": "text",
|
| 920 |
+
"text": "We thank Rinon Gal for his useful suggestions and references. This research was supported in part by the Israel Science Foundation (grants no. 2492/20 and 3441/21), Len Blavatnik and the Blavatnik family foundation, and The Tel Aviv University Innovation Laboratories (TILabs). ",
|
| 921 |
+
"bbox": [
|
| 922 |
+
176,
|
| 923 |
+
882,
|
| 924 |
+
823,
|
| 925 |
+
922
|
| 926 |
+
],
|
| 927 |
+
"page_idx": 8
|
| 928 |
+
},
|
| 929 |
+
{
|
| 930 |
+
"type": "text",
|
| 931 |
+
"text": "REFERENCES ",
|
| 932 |
+
"text_level": 1,
|
| 933 |
+
"bbox": [
|
| 934 |
+
176,
|
| 935 |
+
103,
|
| 936 |
+
287,
|
| 937 |
+
117
|
| 938 |
+
],
|
| 939 |
+
"page_idx": 9
|
| 940 |
+
},
|
| 941 |
+
{
|
| 942 |
+
"type": "text",
|
| 943 |
+
"text": "Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62–1, 2020. ",
|
| 944 |
+
"bbox": [
|
| 945 |
+
176,
|
| 946 |
+
126,
|
| 947 |
+
823,
|
| 948 |
+
169
|
| 949 |
+
],
|
| 950 |
+
"page_idx": 9
|
| 951 |
+
},
|
| 952 |
+
{
|
| 953 |
+
"type": "text",
|
| 954 |
+
"text": "Adobe Systems Inc. Mixamo, 2021. URL https://www.mixamo.com. Accessed: 2021-12-25. ",
|
| 955 |
+
"bbox": [
|
| 956 |
+
171,
|
| 957 |
+
178,
|
| 958 |
+
823,
|
| 959 |
+
194
|
| 960 |
+
],
|
| 961 |
+
"page_idx": 9
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"type": "text",
|
| 965 |
+
"text": "Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pp. 719–728. IEEE, 2019. ",
|
| 966 |
+
"bbox": [
|
| 967 |
+
171,
|
| 968 |
+
202,
|
| 969 |
+
823,
|
| 970 |
+
232
|
| 971 |
+
],
|
| 972 |
+
"page_idx": 9
|
| 973 |
+
},
|
| 974 |
+
{
|
| 975 |
+
"type": "text",
|
| 976 |
+
"text": "Emre Aksan, Manuel Kaufmann, Peng Cao, and Otmar Hilliges. A spatio-temporal transformer for 3d human motion prediction. In 2021 International Conference on 3D Vision (3DV), pp. 565–574. IEEE, 2021. ",
|
| 977 |
+
"bbox": [
|
| 978 |
+
173,
|
| 979 |
+
241,
|
| 980 |
+
823,
|
| 981 |
+
284
|
| 982 |
+
],
|
| 983 |
+
"page_idx": 9
|
| 984 |
+
},
|
| 985 |
+
{
|
| 986 |
+
"type": "text",
|
| 987 |
+
"text": "A Aristidou, A Yiannakidis, K Aberman, D Cohen-Or, A Shamir, and Y Chrysanthou. Rhythm is a dancer: Music-driven motion synthesis with global structure. IEEE Transactions on Visualization and Computer Graphics, 2022. ",
|
| 988 |
+
"bbox": [
|
| 989 |
+
174,
|
| 990 |
+
292,
|
| 991 |
+
825,
|
| 992 |
+
335
|
| 993 |
+
],
|
| 994 |
+
"page_idx": 9
|
| 995 |
+
},
|
| 996 |
+
{
|
| 997 |
+
"type": "text",
|
| 998 |
+
"text": "Uttaran Bhattacharya, Nicholas Rewkowski, Abhishek Banerjee, Pooja Guhan, Aniket Bera, and Dinesh Manocha. Text2gestures: A transformer-based network for generating emotive body gestures for virtual agents. In 2021 IEEE Virtual Reality and 3D User Interfaces (VR), pp. 1–10. IEEE, 2021. ",
|
| 999 |
+
"bbox": [
|
| 1000 |
+
173,
|
| 1001 |
+
344,
|
| 1002 |
+
825,
|
| 1003 |
+
401
|
| 1004 |
+
],
|
| 1005 |
+
"page_idx": 9
|
| 1006 |
+
},
|
| 1007 |
+
{
|
| 1008 |
+
"type": "text",
|
| 1009 |
+
"text": "Pablo Cervantes, Yusuke Sekikawa, Ikuro Sato, and Koichi Shinoda. Implicit neural representations for variable length human motion generation. arXiv preprint arXiv:2203.13694, 2022. ",
|
| 1010 |
+
"bbox": [
|
| 1011 |
+
173,
|
| 1012 |
+
411,
|
| 1013 |
+
823,
|
| 1014 |
+
440
|
| 1015 |
+
],
|
| 1016 |
+
"page_idx": 9
|
| 1017 |
+
},
|
| 1018 |
+
{
|
| 1019 |
+
"type": "text",
|
| 1020 |
+
"text": "Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Hol- ¨ ger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. ",
|
| 1021 |
+
"bbox": [
|
| 1022 |
+
173,
|
| 1023 |
+
449,
|
| 1024 |
+
826,
|
| 1025 |
+
492
|
| 1026 |
+
],
|
| 1027 |
+
"page_idx": 9
|
| 1028 |
+
},
|
| 1029 |
+
{
|
| 1030 |
+
"type": "text",
|
| 1031 |
+
"text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 4171–4186, Minneapolis, Minnesota, June 2019. Association for Computational Linguistics. doi: 10.18653/v1/N19-1423. URL https: //aclanthology.org/N19-1423. ",
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
173,
|
| 1034 |
+
501,
|
| 1035 |
+
825,
|
| 1036 |
+
587
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 9
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "text",
|
| 1042 |
+
"text": "Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780–8794, 2021. ",
|
| 1043 |
+
"bbox": [
|
| 1044 |
+
173,
|
| 1045 |
+
595,
|
| 1046 |
+
823,
|
| 1047 |
+
625
|
| 1048 |
+
],
|
| 1049 |
+
"page_idx": 9
|
| 1050 |
+
},
|
| 1051 |
+
{
|
| 1052 |
+
"type": "text",
|
| 1053 |
+
"text": "Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. ",
|
| 1054 |
+
"bbox": [
|
| 1055 |
+
173,
|
| 1056 |
+
633,
|
| 1057 |
+
821,
|
| 1058 |
+
662
|
| 1059 |
+
],
|
| 1060 |
+
"page_idx": 9
|
| 1061 |
+
},
|
| 1062 |
+
{
|
| 1063 |
+
"type": "text",
|
| 1064 |
+
"text": "Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Jitendra Malik. Recurrent network models for human dynamics. In Proceedings of the IEEE international conference on computer vision, pp. 4346–4354, 2015. ",
|
| 1065 |
+
"bbox": [
|
| 1066 |
+
174,
|
| 1067 |
+
671,
|
| 1068 |
+
825,
|
| 1069 |
+
715
|
| 1070 |
+
],
|
| 1071 |
+
"page_idx": 9
|
| 1072 |
+
},
|
| 1073 |
+
{
|
| 1074 |
+
"type": "text",
|
| 1075 |
+
"text": "Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17113–17122, 2022. ",
|
| 1076 |
+
"bbox": [
|
| 1077 |
+
173,
|
| 1078 |
+
724,
|
| 1079 |
+
825,
|
| 1080 |
+
767
|
| 1081 |
+
],
|
| 1082 |
+
"page_idx": 9
|
| 1083 |
+
},
|
| 1084 |
+
{
|
| 1085 |
+
"type": "text",
|
| 1086 |
+
"text": "Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pp. 2021–2029, 2020. ",
|
| 1087 |
+
"bbox": [
|
| 1088 |
+
176,
|
| 1089 |
+
776,
|
| 1090 |
+
823,
|
| 1091 |
+
820
|
| 1092 |
+
],
|
| 1093 |
+
"page_idx": 9
|
| 1094 |
+
},
|
| 1095 |
+
{
|
| 1096 |
+
"type": "text",
|
| 1097 |
+
"text": "Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5152–5161, 2022a. ",
|
| 1098 |
+
"bbox": [
|
| 1099 |
+
176,
|
| 1100 |
+
829,
|
| 1101 |
+
823,
|
| 1102 |
+
872
|
| 1103 |
+
],
|
| 1104 |
+
"page_idx": 9
|
| 1105 |
+
},
|
| 1106 |
+
{
|
| 1107 |
+
"type": "text",
|
| 1108 |
+
"text": "Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc MorenoNoguer. Back to mlp: A simple baseline for human motion prediction. arXiv preprint arXiv:2207.01567, 2022b. ",
|
| 1109 |
+
"bbox": [
|
| 1110 |
+
176,
|
| 1111 |
+
881,
|
| 1112 |
+
823,
|
| 1113 |
+
924
|
| 1114 |
+
],
|
| 1115 |
+
"page_idx": 9
|
| 1116 |
+
},
|
| 1117 |
+
{
|
| 1118 |
+
"type": "text",
|
| 1119 |
+
"text": "Felix G Harvey and Christopher Pal. Recurrent transition networks for character locomotion. In ´ SIGGRAPH Asia 2018 Technical Briefs, pp. 1–4. 2018. ",
|
| 1120 |
+
"bbox": [
|
| 1121 |
+
173,
|
| 1122 |
+
103,
|
| 1123 |
+
823,
|
| 1124 |
+
133
|
| 1125 |
+
],
|
| 1126 |
+
"page_idx": 10
|
| 1127 |
+
},
|
| 1128 |
+
{
|
| 1129 |
+
"type": "text",
|
| 1130 |
+
"text": "Felix G Harvey, Mike Yurick, Derek Nowrouzezahrai, and Christopher Pal. Robust motion in- ´ betweening. ACM Transactions on Graphics (TOG), 39(4):60–1, 2020. ",
|
| 1131 |
+
"bbox": [
|
| 1132 |
+
171,
|
| 1133 |
+
140,
|
| 1134 |
+
821,
|
| 1135 |
+
170
|
| 1136 |
+
],
|
| 1137 |
+
"page_idx": 10
|
| 1138 |
+
},
|
| 1139 |
+
{
|
| 1140 |
+
"type": "text",
|
| 1141 |
+
"text": "Alejandro Hernandez, Jurgen Gall, and Francesc Moreno-Noguer. Human motion prediction via spatio-temporal inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7134–7143, 2019. ",
|
| 1142 |
+
"bbox": [
|
| 1143 |
+
176,
|
| 1144 |
+
178,
|
| 1145 |
+
823,
|
| 1146 |
+
222
|
| 1147 |
+
],
|
| 1148 |
+
"page_idx": 10
|
| 1149 |
+
},
|
| 1150 |
+
{
|
| 1151 |
+
"type": "text",
|
| 1152 |
+
"text": "Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. ",
|
| 1153 |
+
"bbox": [
|
| 1154 |
+
169,
|
| 1155 |
+
229,
|
| 1156 |
+
825,
|
| 1157 |
+
258
|
| 1158 |
+
],
|
| 1159 |
+
"page_idx": 10
|
| 1160 |
+
},
|
| 1161 |
+
{
|
| 1162 |
+
"type": "text",
|
| 1163 |
+
"text": "Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. ",
|
| 1164 |
+
"bbox": [
|
| 1165 |
+
173,
|
| 1166 |
+
267,
|
| 1167 |
+
825,
|
| 1168 |
+
296
|
| 1169 |
+
],
|
| 1170 |
+
"page_idx": 10
|
| 1171 |
+
},
|
| 1172 |
+
{
|
| 1173 |
+
"type": "text",
|
| 1174 |
+
"text": "Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. ",
|
| 1175 |
+
"bbox": [
|
| 1176 |
+
174,
|
| 1177 |
+
304,
|
| 1178 |
+
823,
|
| 1179 |
+
334
|
| 1180 |
+
],
|
| 1181 |
+
"page_idx": 10
|
| 1182 |
+
},
|
| 1183 |
+
{
|
| 1184 |
+
"type": "text",
|
| 1185 |
+
"text": "Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Transactions on Graphics (TOG), 35(4):1–11, 2016. ",
|
| 1186 |
+
"bbox": [
|
| 1187 |
+
173,
|
| 1188 |
+
342,
|
| 1189 |
+
823,
|
| 1190 |
+
372
|
| 1191 |
+
],
|
| 1192 |
+
"page_idx": 10
|
| 1193 |
+
},
|
| 1194 |
+
{
|
| 1195 |
+
"type": "text",
|
| 1196 |
+
"text": "Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng. A large-scale rgb-d database for arbitrary-view human action recognition. In Proceedings of the 26th ACM international Conference on Multimedia, pp. 1510–1518, 2018. ",
|
| 1197 |
+
"bbox": [
|
| 1198 |
+
176,
|
| 1199 |
+
380,
|
| 1200 |
+
821,
|
| 1201 |
+
424
|
| 1202 |
+
],
|
| 1203 |
+
"page_idx": 10
|
| 1204 |
+
},
|
| 1205 |
+
{
|
| 1206 |
+
"type": "text",
|
| 1207 |
+
"text": "Manuel Kaufmann, Emre Aksan, Jie Song, Fabrizio Pece, Remo Ziegler, and Otmar Hilliges. Convolutional autoencoders for human motion infilling. In 2020 International Conference on 3D Vision (3DV), pp. 918–927. IEEE, 2020. ",
|
| 1208 |
+
"bbox": [
|
| 1209 |
+
176,
|
| 1210 |
+
431,
|
| 1211 |
+
823,
|
| 1212 |
+
474
|
| 1213 |
+
],
|
| 1214 |
+
"page_idx": 10
|
| 1215 |
+
},
|
| 1216 |
+
{
|
| 1217 |
+
"type": "text",
|
| 1218 |
+
"text": "Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. ",
|
| 1219 |
+
"bbox": [
|
| 1220 |
+
173,
|
| 1221 |
+
483,
|
| 1222 |
+
821,
|
| 1223 |
+
512
|
| 1224 |
+
],
|
| 1225 |
+
"page_idx": 10
|
| 1226 |
+
},
|
| 1227 |
+
{
|
| 1228 |
+
"type": "text",
|
| 1229 |
+
"text": "Diederik $\\mathrm { \\bf P }$ Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. ",
|
| 1230 |
+
"bbox": [
|
| 1231 |
+
173,
|
| 1232 |
+
520,
|
| 1233 |
+
821,
|
| 1234 |
+
550
|
| 1235 |
+
],
|
| 1236 |
+
"page_idx": 10
|
| 1237 |
+
},
|
| 1238 |
+
{
|
| 1239 |
+
"type": "text",
|
| 1240 |
+
"text": "Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5253–5263, 2020. ",
|
| 1241 |
+
"bbox": [
|
| 1242 |
+
173,
|
| 1243 |
+
558,
|
| 1244 |
+
823,
|
| 1245 |
+
602
|
| 1246 |
+
],
|
| 1247 |
+
"page_idx": 10
|
| 1248 |
+
},
|
| 1249 |
+
{
|
| 1250 |
+
"type": "text",
|
| 1251 |
+
"text": "Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist $^ { + + }$ . In The IEEE International Conference on Computer Vision (ICCV), 2021. ",
|
| 1252 |
+
"bbox": [
|
| 1253 |
+
173,
|
| 1254 |
+
609,
|
| 1255 |
+
823,
|
| 1256 |
+
652
|
| 1257 |
+
],
|
| 1258 |
+
"page_idx": 10
|
| 1259 |
+
},
|
| 1260 |
+
{
|
| 1261 |
+
"type": "text",
|
| 1262 |
+
"text": "Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2837–2845, 2021. ",
|
| 1263 |
+
"bbox": [
|
| 1264 |
+
174,
|
| 1265 |
+
661,
|
| 1266 |
+
825,
|
| 1267 |
+
704
|
| 1268 |
+
],
|
| 1269 |
+
"page_idx": 10
|
| 1270 |
+
},
|
| 1271 |
+
{
|
| 1272 |
+
"type": "text",
|
| 1273 |
+
"text": "Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In International Conference on Computer Vision, pp. 5442–5451, October 2019. ",
|
| 1274 |
+
"bbox": [
|
| 1275 |
+
176,
|
| 1276 |
+
712,
|
| 1277 |
+
823,
|
| 1278 |
+
756
|
| 1279 |
+
],
|
| 1280 |
+
"page_idx": 10
|
| 1281 |
+
},
|
| 1282 |
+
{
|
| 1283 |
+
"type": "text",
|
| 1284 |
+
"text": "Julieta Martinez, Michael J Black, and Javier Romero. On human motion prediction using recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2891–2900, 2017. ",
|
| 1285 |
+
"bbox": [
|
| 1286 |
+
173,
|
| 1287 |
+
763,
|
| 1288 |
+
823,
|
| 1289 |
+
808
|
| 1290 |
+
],
|
| 1291 |
+
"page_idx": 10
|
| 1292 |
+
},
|
| 1293 |
+
{
|
| 1294 |
+
"type": "text",
|
| 1295 |
+
"text": "Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. ",
|
| 1296 |
+
"bbox": [
|
| 1297 |
+
174,
|
| 1298 |
+
815,
|
| 1299 |
+
823,
|
| 1300 |
+
859
|
| 1301 |
+
],
|
| 1302 |
+
"page_idx": 10
|
| 1303 |
+
},
|
| 1304 |
+
{
|
| 1305 |
+
"type": "text",
|
| 1306 |
+
"text": "Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pp. 10975–10985, 2019. ",
|
| 1307 |
+
"bbox": [
|
| 1308 |
+
174,
|
| 1309 |
+
867,
|
| 1310 |
+
825,
|
| 1311 |
+
924
|
| 1312 |
+
],
|
| 1313 |
+
"page_idx": 10
|
| 1314 |
+
},
|
| 1315 |
+
{
|
| 1316 |
+
"type": "text",
|
| 1317 |
+
"text": "Mathis Petrovich, Michael J. Black, and Gul Varol. Action-conditioned 3D human motion synthesis ¨ with transformer VAE. In International Conference on Computer Vision (ICCV), pp. 10985– 10995, October 2021. ",
|
| 1318 |
+
"bbox": [
|
| 1319 |
+
173,
|
| 1320 |
+
103,
|
| 1321 |
+
823,
|
| 1322 |
+
146
|
| 1323 |
+
],
|
| 1324 |
+
"page_idx": 11
|
| 1325 |
+
},
|
| 1326 |
+
{
|
| 1327 |
+
"type": "text",
|
| 1328 |
+
"text": "Mathis Petrovich, Michael J. Black, and Gul Varol. TEMOS: Generating diverse human motions ¨ from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. ",
|
| 1329 |
+
"bbox": [
|
| 1330 |
+
173,
|
| 1331 |
+
154,
|
| 1332 |
+
821,
|
| 1333 |
+
184
|
| 1334 |
+
],
|
| 1335 |
+
"page_idx": 11
|
| 1336 |
+
},
|
| 1337 |
+
{
|
| 1338 |
+
"type": "text",
|
| 1339 |
+
"text": "Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big data, 4(4):236–252, 2016. ",
|
| 1340 |
+
"bbox": [
|
| 1341 |
+
173,
|
| 1342 |
+
191,
|
| 1343 |
+
823,
|
| 1344 |
+
220
|
| 1345 |
+
],
|
| 1346 |
+
"page_idx": 11
|
| 1347 |
+
},
|
| 1348 |
+
{
|
| 1349 |
+
"type": "text",
|
| 1350 |
+
"text": "Sigal Raab, Inbal Leibovitch, Peizhuo Li, Kfir Aberman, Olga Sorkine-Hornung, and Daniel CohenOr. Modi: Unconditional motion synthesis from diverse data. arXiv preprint arXiv:2206.08010, 2022. ",
|
| 1351 |
+
"bbox": [
|
| 1352 |
+
176,
|
| 1353 |
+
228,
|
| 1354 |
+
823,
|
| 1355 |
+
272
|
| 1356 |
+
],
|
| 1357 |
+
"page_idx": 11
|
| 1358 |
+
},
|
| 1359 |
+
{
|
| 1360 |
+
"type": "text",
|
| 1361 |
+
"text": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pp. 8748–8763. PMLR, 2021. ",
|
| 1362 |
+
"bbox": [
|
| 1363 |
+
174,
|
| 1364 |
+
280,
|
| 1365 |
+
825,
|
| 1366 |
+
337
|
| 1367 |
+
],
|
| 1368 |
+
"page_idx": 11
|
| 1369 |
+
},
|
| 1370 |
+
{
|
| 1371 |
+
"type": "text",
|
| 1372 |
+
"text": "Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical textconditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. ",
|
| 1373 |
+
"bbox": [
|
| 1374 |
+
173,
|
| 1375 |
+
344,
|
| 1376 |
+
820,
|
| 1377 |
+
375
|
| 1378 |
+
],
|
| 1379 |
+
"page_idx": 11
|
| 1380 |
+
},
|
| 1381 |
+
{
|
| 1382 |
+
"type": "text",
|
| 1383 |
+
"text": "Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bertnetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pp. 3982–3992, 2019. ",
|
| 1384 |
+
"bbox": [
|
| 1385 |
+
173,
|
| 1386 |
+
382,
|
| 1387 |
+
825,
|
| 1388 |
+
439
|
| 1389 |
+
],
|
| 1390 |
+
"page_idx": 11
|
| 1391 |
+
},
|
| 1392 |
+
{
|
| 1393 |
+
"type": "text",
|
| 1394 |
+
"text": "Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computerassisted intervention, pp. 234–241. Springer, 2015. ",
|
| 1395 |
+
"bbox": [
|
| 1396 |
+
174,
|
| 1397 |
+
446,
|
| 1398 |
+
825,
|
| 1399 |
+
491
|
| 1400 |
+
],
|
| 1401 |
+
"page_idx": 11
|
| 1402 |
+
},
|
| 1403 |
+
{
|
| 1404 |
+
"type": "text",
|
| 1405 |
+
"text": "Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pp. 1–10, 2022a. ",
|
| 1406 |
+
"bbox": [
|
| 1407 |
+
173,
|
| 1408 |
+
498,
|
| 1409 |
+
825,
|
| 1410 |
+
542
|
| 1411 |
+
],
|
| 1412 |
+
"page_idx": 11
|
| 1413 |
+
},
|
| 1414 |
+
{
|
| 1415 |
+
"type": "text",
|
| 1416 |
+
"text": "Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b. ",
|
| 1417 |
+
"bbox": [
|
| 1418 |
+
173,
|
| 1419 |
+
550,
|
| 1420 |
+
825,
|
| 1421 |
+
607
|
| 1422 |
+
],
|
| 1423 |
+
"page_idx": 11
|
| 1424 |
+
},
|
| 1425 |
+
{
|
| 1426 |
+
"type": "text",
|
| 1427 |
+
"text": "Mingyi Shi, Kfir Aberman, Andreas Aristidou, Taku Komura, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. Motionet: 3d human motion reconstruction from monocular video with skeleton consistency. ACM Transactions on Graphics (TOG), 40(1):1–15, 2020. ",
|
| 1428 |
+
"bbox": [
|
| 1429 |
+
173,
|
| 1430 |
+
614,
|
| 1431 |
+
823,
|
| 1432 |
+
659
|
| 1433 |
+
],
|
| 1434 |
+
"page_idx": 11
|
| 1435 |
+
},
|
| 1436 |
+
{
|
| 1437 |
+
"type": "text",
|
| 1438 |
+
"text": "Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp. 2256–2265. PMLR, 2015. ",
|
| 1439 |
+
"bbox": [
|
| 1440 |
+
173,
|
| 1441 |
+
666,
|
| 1442 |
+
825,
|
| 1443 |
+
709
|
| 1444 |
+
],
|
| 1445 |
+
"page_idx": 11
|
| 1446 |
+
},
|
| 1447 |
+
{
|
| 1448 |
+
"type": "text",
|
| 1449 |
+
"text": "Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020a. ",
|
| 1450 |
+
"bbox": [
|
| 1451 |
+
169,
|
| 1452 |
+
718,
|
| 1453 |
+
823,
|
| 1454 |
+
747
|
| 1455 |
+
],
|
| 1456 |
+
"page_idx": 11
|
| 1457 |
+
},
|
| 1458 |
+
{
|
| 1459 |
+
"type": "text",
|
| 1460 |
+
"text": "Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438–12448, 2020. ",
|
| 1461 |
+
"bbox": [
|
| 1462 |
+
169,
|
| 1463 |
+
755,
|
| 1464 |
+
823,
|
| 1465 |
+
785
|
| 1466 |
+
],
|
| 1467 |
+
"page_idx": 11
|
| 1468 |
+
},
|
| 1469 |
+
{
|
| 1470 |
+
"type": "text",
|
| 1471 |
+
"text": "Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020b. ",
|
| 1472 |
+
"bbox": [
|
| 1473 |
+
173,
|
| 1474 |
+
792,
|
| 1475 |
+
825,
|
| 1476 |
+
835
|
| 1477 |
+
],
|
| 1478 |
+
"page_idx": 11
|
| 1479 |
+
},
|
| 1480 |
+
{
|
| 1481 |
+
"type": "text",
|
| 1482 |
+
"text": "Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. arXiv preprint arXiv:2203.08063, 2022. ",
|
| 1483 |
+
"bbox": [
|
| 1484 |
+
171,
|
| 1485 |
+
843,
|
| 1486 |
+
823,
|
| 1487 |
+
873
|
| 1488 |
+
],
|
| 1489 |
+
"page_idx": 11
|
| 1490 |
+
},
|
| 1491 |
+
{
|
| 1492 |
+
"type": "text",
|
| 1493 |
+
"text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. ",
|
| 1494 |
+
"bbox": [
|
| 1495 |
+
174,
|
| 1496 |
+
882,
|
| 1497 |
+
823,
|
| 1498 |
+
924
|
| 1499 |
+
],
|
| 1500 |
+
"page_idx": 11
|
| 1501 |
+
},
|
| 1502 |
+
{
|
| 1503 |
+
"type": "text",
|
| 1504 |
+
"text": "Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. ",
|
| 1505 |
+
"bbox": [
|
| 1506 |
+
174,
|
| 1507 |
+
103,
|
| 1508 |
+
823,
|
| 1509 |
+
146
|
| 1510 |
+
],
|
| 1511 |
+
"page_idx": 12
|
| 1512 |
+
},
|
| 1513 |
+
{
|
| 1514 |
+
"type": "text",
|
| 1515 |
+
"text": "A ADDITIONAL EXPERIMENTS ",
|
| 1516 |
+
"text_level": 1,
|
| 1517 |
+
"bbox": [
|
| 1518 |
+
178,
|
| 1519 |
+
186,
|
| 1520 |
+
444,
|
| 1521 |
+
203
|
| 1522 |
+
],
|
| 1523 |
+
"page_idx": 12
|
| 1524 |
+
},
|
| 1525 |
+
{
|
| 1526 |
+
"type": "text",
|
| 1527 |
+
"text": "A.1 DIFFUSION PARAMETERS ",
|
| 1528 |
+
"text_level": 1,
|
| 1529 |
+
"bbox": [
|
| 1530 |
+
176,
|
| 1531 |
+
223,
|
| 1532 |
+
395,
|
| 1533 |
+
238
|
| 1534 |
+
],
|
| 1535 |
+
"page_idx": 12
|
| 1536 |
+
},
|
| 1537 |
+
{
|
| 1538 |
+
"type": "text",
|
| 1539 |
+
"text": "Learning MDM with different numbers of diffusion steps significantly affects performance and holds the potential to accelerate inference time. Table 6 shows optimal performance for $T = 1 0 0$ , in addition, it enables accelerating inference by a factor of 10 compared to $T = 1 0 0 0$ , which is widely used for images. ",
|
| 1540 |
+
"bbox": [
|
| 1541 |
+
173,
|
| 1542 |
+
252,
|
| 1543 |
+
825,
|
| 1544 |
+
309
|
| 1545 |
+
],
|
| 1546 |
+
"page_idx": 12
|
| 1547 |
+
},
|
| 1548 |
+
{
|
| 1549 |
+
"type": "table",
|
| 1550 |
+
"img_path": "images/e2fec044e57faea9d2df7a47f4b8d49941af16ad176262c5e777eeb8e6e950ab.jpg",
|
| 1551 |
+
"table_caption": [
|
| 1552 |
+
"Table 6: Diffusion steps (HumanML3D test set). We run all the evaluation 20 times. Bold indicates best result, underline indicates second best, $\\pm$ indicates $9 5 \\%$ confidence interval, indicates that closer to real is better. "
|
| 1553 |
+
],
|
| 1554 |
+
"table_footnote": [],
|
| 1555 |
+
"table_body": "<table><tr><td>Diffusion steps (T)</td><td>R Precision (top 3)↑</td><td>FID↓</td><td>Multimodal Dist↓</td><td>Diversity→</td></tr><tr><td>Real</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td></tr><tr><td>10</td><td>0.574±.006</td><td>1.461±.088</td><td>5.816±.033</td><td>9.369±.058</td></tr><tr><td>100</td><td>0.640±.007</td><td>0.454±.039</td><td>5.336±.029</td><td>9.906±.053</td></tr><tr><td>500</td><td>0.662±.007</td><td>0.553±.055</td><td>5.177±.028</td><td>9.890±.074</td></tr><tr><td>1000</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td></tr></table>",
|
| 1556 |
+
"bbox": [
|
| 1557 |
+
207,
|
| 1558 |
+
329,
|
| 1559 |
+
781,
|
| 1560 |
+
465
|
| 1561 |
+
],
|
| 1562 |
+
"page_idx": 12
|
| 1563 |
+
},
|
| 1564 |
+
{
|
| 1565 |
+
"type": "text",
|
| 1566 |
+
"text": "A.2 GEOMETRIC LOSSES ",
|
| 1567 |
+
"text_level": 1,
|
| 1568 |
+
"bbox": [
|
| 1569 |
+
176,
|
| 1570 |
+
559,
|
| 1571 |
+
362,
|
| 1572 |
+
574
|
| 1573 |
+
],
|
| 1574 |
+
"page_idx": 12
|
| 1575 |
+
},
|
| 1576 |
+
{
|
| 1577 |
+
"type": "text",
|
| 1578 |
+
"text": "We conduct a thorough experiment to evaluate the contribution of geometric losses with the HumanAct12 dataset. The results are presented in Table 7. For alignment with prior work, all metrics are calculated using the deep features of the action recognition network suggested by Guo et al. (2020). In general, MDM scores are too close to the real test distribution (i.e. the evaluation network fails to discriminate between the two). This means that quantitative results comparing the different variants MDM are too similar to evaluate. As a result, we are not able to decide what combination of geometric losses is preferred. We leave for future work experimenting with a different, more expressive, evaluation network. ",
|
| 1579 |
+
"bbox": [
|
| 1580 |
+
173,
|
| 1581 |
+
588,
|
| 1582 |
+
825,
|
| 1583 |
+
700
|
| 1584 |
+
],
|
| 1585 |
+
"page_idx": 12
|
| 1586 |
+
},
|
| 1587 |
+
{
|
| 1588 |
+
"type": "table",
|
| 1589 |
+
"img_path": "images/328d4ad3f45e6215f84fbbaa5b0be5c491d036657b33b272e8c9d01cf9d28fc0.jpg",
|
| 1590 |
+
"table_caption": [],
|
| 1591 |
+
"table_footnote": [
|
| 1592 |
+
"Table 7: Geometric losses ablation study. (HumanAct12 dataset) the relative $\\lambda$ equals 1 when the loss term is included, and 0 when it is excluded. "
|
| 1593 |
+
],
|
| 1594 |
+
"table_body": "<table><tr><td>Method</td><td>FID↓</td><td>Accuracy↑</td><td>Diversity→</td><td>Multimodality→</td></tr><tr><td>Real</td><td>0.050±.000</td><td>0.990±.000</td><td>6.880±.020</td><td>2.590±.010</td></tr><tr><td>MDM (ours)</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>w/o foot contact</td><td>0.080±.000</td><td>0.990±.000</td><td>6.810±.010</td><td>2.580±.010</td></tr><tr><td>w/o geometric losses</td><td>0.090±.000</td><td>0.990±.000</td><td>6.820±.020</td><td>2.550±.010</td></tr><tr><td>foot contact only</td><td>0.100±.000</td><td>0.990±.000</td><td>6.860±.050</td><td>2.520±.010</td></tr><tr><td>velocity only</td><td>0.100±.000</td><td>0.990±.000</td><td>6.820±.020</td><td>2.590±.000</td></tr><tr><td>pose only</td><td>0.090±.000</td><td>0.990±.000</td><td>6.830±.020</td><td>2.570±.020</td></tr></table>",
|
| 1595 |
+
"bbox": [
|
| 1596 |
+
196,
|
| 1597 |
+
717,
|
| 1598 |
+
805,
|
| 1599 |
+
866
|
| 1600 |
+
],
|
| 1601 |
+
"page_idx": 12
|
| 1602 |
+
},
|
| 1603 |
+
{
|
| 1604 |
+
"type": "text",
|
| 1605 |
+
"text": "B EVALUATION METRICS. ",
|
| 1606 |
+
"text_level": 1,
|
| 1607 |
+
"bbox": [
|
| 1608 |
+
176,
|
| 1609 |
+
102,
|
| 1610 |
+
406,
|
| 1611 |
+
118
|
| 1612 |
+
],
|
| 1613 |
+
"page_idx": 13
|
| 1614 |
+
},
|
| 1615 |
+
{
|
| 1616 |
+
"type": "text",
|
| 1617 |
+
"text": "For the completeness of our work, we describe here the quantitative metrics used throughout the paper, as they originally described and implemented by Guo et al. (2020) for action-to-motion and by Guo et al. (2022a) for text-to-motion. ",
|
| 1618 |
+
"bbox": [
|
| 1619 |
+
174,
|
| 1620 |
+
136,
|
| 1621 |
+
825,
|
| 1622 |
+
179
|
| 1623 |
+
],
|
| 1624 |
+
"page_idx": 13
|
| 1625 |
+
},
|
| 1626 |
+
{
|
| 1627 |
+
"type": "text",
|
| 1628 |
+
"text": "B.1 ACTION-TO-MOTION ",
|
| 1629 |
+
"text_level": 1,
|
| 1630 |
+
"bbox": [
|
| 1631 |
+
176,
|
| 1632 |
+
200,
|
| 1633 |
+
364,
|
| 1634 |
+
214
|
| 1635 |
+
],
|
| 1636 |
+
"page_idx": 13
|
| 1637 |
+
},
|
| 1638 |
+
{
|
| 1639 |
+
"type": "text",
|
| 1640 |
+
"text": "The following metrics are based on an RNN action recognition network as it was originally trained by Guo et al. (2020). We refer to it as the evaluator network. ",
|
| 1641 |
+
"bbox": [
|
| 1642 |
+
173,
|
| 1643 |
+
228,
|
| 1644 |
+
823,
|
| 1645 |
+
257
|
| 1646 |
+
],
|
| 1647 |
+
"page_idx": 13
|
| 1648 |
+
},
|
| 1649 |
+
{
|
| 1650 |
+
"type": "text",
|
| 1651 |
+
"text": "Frechet Inception Distance (FID). A widely used metric to evaluate the overall quality for generation tasks. FID is calculated upon features extracted from 1,000 generated motion vs ground truth (real) taken from the test set. To adjust this metric to the motion domain, we extract a deep representation of the motion with the evaluator network instead of the inception neural network, originally used for images. A lower value implies better FID results. ",
|
| 1652 |
+
"bbox": [
|
| 1653 |
+
174,
|
| 1654 |
+
263,
|
| 1655 |
+
823,
|
| 1656 |
+
333
|
| 1657 |
+
],
|
| 1658 |
+
"page_idx": 13
|
| 1659 |
+
},
|
| 1660 |
+
{
|
| 1661 |
+
"type": "text",
|
| 1662 |
+
"text": "Accuracy. We classify 1,000 generated motions using the evaluator network, than we calculate the overall recognition accuracy that indicates the correlation of the motion and its action type. ",
|
| 1663 |
+
"bbox": [
|
| 1664 |
+
171,
|
| 1665 |
+
340,
|
| 1666 |
+
823,
|
| 1667 |
+
368
|
| 1668 |
+
],
|
| 1669 |
+
"page_idx": 13
|
| 1670 |
+
},
|
| 1671 |
+
{
|
| 1672 |
+
"type": "text",
|
| 1673 |
+
"text": "Diversity measures the variance of the generated motions across all action categories. We first randomly sample two subsets of the same size $S _ { d }$ out of a set of all generated motions across all action categories denoted $\\{ \\mathbf { v } _ { 1 } , . . . , \\mathbf { v } _ { S _ { d } } \\}$ and $\\big \\{ \\mathbf { v } _ { 1 } ^ { \\prime \\prime } , . . . , \\mathbf { v } _ { S _ { d } } ^ { \\prime } \\big \\}$ . The diversity of those sets of motions is defied as ",
|
| 1674 |
+
"bbox": [
|
| 1675 |
+
173,
|
| 1676 |
+
375,
|
| 1677 |
+
825,
|
| 1678 |
+
431
|
| 1679 |
+
],
|
| 1680 |
+
"page_idx": 13
|
| 1681 |
+
},
|
| 1682 |
+
{
|
| 1683 |
+
"type": "equation",
|
| 1684 |
+
"img_path": "images/3355b57415a6194344bff9111e3022e8e7a9f043e753343692e778c758f0c6ee.jpg",
|
| 1685 |
+
"text": "$$\n\\mathrm { D i v e r s i t y } = \\frac { 1 } { S _ { d } } \\sum _ { i = 1 } ^ { S _ { d } } \\parallel \\mathbf { v } _ { i } - \\mathbf { v } _ { i } ^ { \\prime } \\parallel _ { 2 } .\n$$",
|
| 1686 |
+
"text_format": "latex",
|
| 1687 |
+
"bbox": [
|
| 1688 |
+
380,
|
| 1689 |
+
433,
|
| 1690 |
+
617,
|
| 1691 |
+
477
|
| 1692 |
+
],
|
| 1693 |
+
"page_idx": 13
|
| 1694 |
+
},
|
| 1695 |
+
{
|
| 1696 |
+
"type": "text",
|
| 1697 |
+
"text": "We use $S _ { d } = 2 0 0$ for our experiments. The diversity value is considered better when closer to the diversity value of the ground truth. ",
|
| 1698 |
+
"bbox": [
|
| 1699 |
+
173,
|
| 1700 |
+
484,
|
| 1701 |
+
825,
|
| 1702 |
+
513
|
| 1703 |
+
],
|
| 1704 |
+
"page_idx": 13
|
| 1705 |
+
},
|
| 1706 |
+
{
|
| 1707 |
+
"type": "text",
|
| 1708 |
+
"text": "Multimodality measures the generated motions diversify within each action class. We randomly sample two subsets with size $S _ { l }$ of the same motion class $c$ $\\{ \\mathbf { v } _ { c , 1 } , \\ldots , \\mathbf { v } _ { c , S _ { l } } \\}$ and $\\{ \\mathbf { v } _ { c , 1 } ^ { \\prime } , . . . , \\mathbf { v } _ { c , S _ { l } } ^ { \\prime } \\}$ . The multimodality of all action classes $C$ is defined as, ",
|
| 1709 |
+
"bbox": [
|
| 1710 |
+
174,
|
| 1711 |
+
520,
|
| 1712 |
+
825,
|
| 1713 |
+
563
|
| 1714 |
+
],
|
| 1715 |
+
"page_idx": 13
|
| 1716 |
+
},
|
| 1717 |
+
{
|
| 1718 |
+
"type": "equation",
|
| 1719 |
+
"img_path": "images/572d12d24732b9d2cb80866ed300b6a0e5d20bc1f029326f9d5ffe8d676a6419.jpg",
|
| 1720 |
+
"text": "$$\n\\mathrm { M u l t i m o d a l i t y } = \\frac { 1 } { C \\times S _ { l } } \\sum _ { c = 1 } ^ { C } \\sum _ { i = 1 } ^ { S _ { l } } \\left\\| \\mathbf { v } _ { c , i } - \\mathbf { v } _ { c , i } ^ { \\prime } \\right\\| _ { 2 } .\n$$",
|
| 1721 |
+
"text_format": "latex",
|
| 1722 |
+
"bbox": [
|
| 1723 |
+
328,
|
| 1724 |
+
573,
|
| 1725 |
+
668,
|
| 1726 |
+
617
|
| 1727 |
+
],
|
| 1728 |
+
"page_idx": 13
|
| 1729 |
+
},
|
| 1730 |
+
{
|
| 1731 |
+
"type": "text",
|
| 1732 |
+
"text": "We use $S _ { l } = 2 0$ for our experiments. ",
|
| 1733 |
+
"bbox": [
|
| 1734 |
+
176,
|
| 1735 |
+
627,
|
| 1736 |
+
416,
|
| 1737 |
+
642
|
| 1738 |
+
],
|
| 1739 |
+
"page_idx": 13
|
| 1740 |
+
},
|
| 1741 |
+
{
|
| 1742 |
+
"type": "text",
|
| 1743 |
+
"text": "B.2 TEXT-TO-MOTION ",
|
| 1744 |
+
"text_level": 1,
|
| 1745 |
+
"bbox": [
|
| 1746 |
+
176,
|
| 1747 |
+
662,
|
| 1748 |
+
346,
|
| 1749 |
+
678
|
| 1750 |
+
],
|
| 1751 |
+
"page_idx": 13
|
| 1752 |
+
},
|
| 1753 |
+
{
|
| 1754 |
+
"type": "text",
|
| 1755 |
+
"text": "Originally suggested by Guo et al. (2022a), the following metrics are based on a text feature extractor and motion feature extractor jointly trained under contrastive loss to produce geometrically close feature vectors for matched text-motion pairs, and vise versa. ",
|
| 1756 |
+
"bbox": [
|
| 1757 |
+
174,
|
| 1758 |
+
691,
|
| 1759 |
+
823,
|
| 1760 |
+
734
|
| 1761 |
+
],
|
| 1762 |
+
"page_idx": 13
|
| 1763 |
+
},
|
| 1764 |
+
{
|
| 1765 |
+
"type": "text",
|
| 1766 |
+
"text": "R Precision. (top-3) For each generated motion, its ground-truth text and a randomly selected missmatched descriptions from the test set. We calculate the euclidean distance between the motion feature and text feature of each description in the pool. We count the average accuracy at top3 places. If the ground truth entry falling into the top-3 candidates, we treat it as True Positive retrieval. We use a batch size 32 (i.e. 31 negative examples). ",
|
| 1767 |
+
"bbox": [
|
| 1768 |
+
173,
|
| 1769 |
+
741,
|
| 1770 |
+
825,
|
| 1771 |
+
810
|
| 1772 |
+
],
|
| 1773 |
+
"page_idx": 13
|
| 1774 |
+
},
|
| 1775 |
+
{
|
| 1776 |
+
"type": "text",
|
| 1777 |
+
"text": "FID. Same as for action-to-motion, using the motion extractor as the evaluator network. ",
|
| 1778 |
+
"bbox": [
|
| 1779 |
+
174,
|
| 1780 |
+
818,
|
| 1781 |
+
746,
|
| 1782 |
+
832
|
| 1783 |
+
],
|
| 1784 |
+
"page_idx": 13
|
| 1785 |
+
},
|
| 1786 |
+
{
|
| 1787 |
+
"type": "text",
|
| 1788 |
+
"text": "Multimodal Distance. We calculate the multimodal distance as the average Euclidean distance between the motion feature of each generated motion and the text feature of its corresponding description in test set. A lower value implies better multimodal distance. ",
|
| 1789 |
+
"bbox": [
|
| 1790 |
+
173,
|
| 1791 |
+
838,
|
| 1792 |
+
823,
|
| 1793 |
+
881
|
| 1794 |
+
],
|
| 1795 |
+
"page_idx": 13
|
| 1796 |
+
},
|
| 1797 |
+
{
|
| 1798 |
+
"type": "text",
|
| 1799 |
+
"text": "Diversity. Same as for action-to-motion but with $S _ { d } = 3 0 0$ . ",
|
| 1800 |
+
"bbox": [
|
| 1801 |
+
173,
|
| 1802 |
+
887,
|
| 1803 |
+
566,
|
| 1804 |
+
902
|
| 1805 |
+
],
|
| 1806 |
+
"page_idx": 13
|
| 1807 |
+
},
|
| 1808 |
+
{
|
| 1809 |
+
"type": "text",
|
| 1810 |
+
"text": "Multimodality. Same as for action-to-motion but with $S _ { m } = 1 0$ . ",
|
| 1811 |
+
"bbox": [
|
| 1812 |
+
173,
|
| 1813 |
+
909,
|
| 1814 |
+
601,
|
| 1815 |
+
924
|
| 1816 |
+
],
|
| 1817 |
+
"page_idx": 13
|
| 1818 |
+
},
|
| 1819 |
+
{
|
| 1820 |
+
"type": "text",
|
| 1821 |
+
"text": "C IMPLEMENTATION DETAILS ",
|
| 1822 |
+
"text_level": 1,
|
| 1823 |
+
"bbox": [
|
| 1824 |
+
176,
|
| 1825 |
+
102,
|
| 1826 |
+
439,
|
| 1827 |
+
118
|
| 1828 |
+
],
|
| 1829 |
+
"page_idx": 14
|
| 1830 |
+
},
|
| 1831 |
+
{
|
| 1832 |
+
"type": "text",
|
| 1833 |
+
"text": "The full implementation of MDM can be found in our published code2. In addition, the followings are the hyperparameters and model details for all of our experiments. ",
|
| 1834 |
+
"bbox": [
|
| 1835 |
+
173,
|
| 1836 |
+
133,
|
| 1837 |
+
823,
|
| 1838 |
+
161
|
| 1839 |
+
],
|
| 1840 |
+
"page_idx": 14
|
| 1841 |
+
},
|
| 1842 |
+
{
|
| 1843 |
+
"type": "text",
|
| 1844 |
+
"text": "Diffusion framework. In all of our experiments, we used an implementation of DDPM (Ho et al., 2020) by Dhariwal & Nichol $( 2 0 2 1 ) ^ { 3 }$ . We use $T = 1 , 0 0 0$ diffusion steps, cosine noise scheduling (predefined sigmas). All other hyperparameters are according to the implementation defaults. ",
|
| 1845 |
+
"bbox": [
|
| 1846 |
+
173,
|
| 1847 |
+
167,
|
| 1848 |
+
825,
|
| 1849 |
+
210
|
| 1850 |
+
],
|
| 1851 |
+
"page_idx": 14
|
| 1852 |
+
},
|
| 1853 |
+
{
|
| 1854 |
+
"type": "text",
|
| 1855 |
+
"text": "Transformer architecture. For our transformer architectures, we used the PyTorch implementation4. We used 8 transformer layers, 4 attention heads, latent dimension $d = 5 1 2$ , dropout 0.1, feed-forward size 1024 and gelu activations. The number of learned parameters for each model is stated in Table 8. ",
|
| 1856 |
+
"bbox": [
|
| 1857 |
+
173,
|
| 1858 |
+
217,
|
| 1859 |
+
825,
|
| 1860 |
+
273
|
| 1861 |
+
],
|
| 1862 |
+
"page_idx": 14
|
| 1863 |
+
},
|
| 1864 |
+
{
|
| 1865 |
+
"type": "text",
|
| 1866 |
+
"text": "GRU architecture. We use the PyTorch implementation of GRU (Cho et al., 2014) 5 with two layers and latent dimension 512. The number of learned parameters for each model is stated in Table 8. ",
|
| 1867 |
+
"bbox": [
|
| 1868 |
+
173,
|
| 1869 |
+
280,
|
| 1870 |
+
823,
|
| 1871 |
+
309
|
| 1872 |
+
],
|
| 1873 |
+
"page_idx": 14
|
| 1874 |
+
},
|
| 1875 |
+
{
|
| 1876 |
+
"type": "table",
|
| 1877 |
+
"img_path": "images/91c4e97cfd198b03973344c786adc67ac1c38f2757b6f589d480cba5d087d2cc.jpg",
|
| 1878 |
+
"table_caption": [
|
| 1879 |
+
"Learning hyperparameters. For all of our experiments, we use batch size 64, learning rate $1 0 ^ { - 4 }$ . "
|
| 1880 |
+
],
|
| 1881 |
+
"table_footnote": [],
|
| 1882 |
+
"table_body": "<table><tr><td>Architecture</td><td># Parameters (-106)</td></tr><tr><td>Transformer Encoder</td><td>17.88</td></tr><tr><td>TransformerDecoder</td><td>26.29</td></tr><tr><td>+ input token</td><td>26.29</td></tr><tr><td>U-net</td><td>23.47</td></tr><tr><td>GRU</td><td>4.47</td></tr></table>",
|
| 1883 |
+
"bbox": [
|
| 1884 |
+
338,
|
| 1885 |
+
343,
|
| 1886 |
+
650,
|
| 1887 |
+
444
|
| 1888 |
+
],
|
| 1889 |
+
"page_idx": 14
|
| 1890 |
+
},
|
| 1891 |
+
{
|
| 1892 |
+
"type": "text",
|
| 1893 |
+
"text": "Table 8: The number of learned parameters per architecture for the text-to-motion task. For the action-to-motion task, there are additional 512 parameters per-class for the class embeddings module. ",
|
| 1894 |
+
"bbox": [
|
| 1895 |
+
173,
|
| 1896 |
+
454,
|
| 1897 |
+
825,
|
| 1898 |
+
496
|
| 1899 |
+
],
|
| 1900 |
+
"page_idx": 14
|
| 1901 |
+
},
|
| 1902 |
+
{
|
| 1903 |
+
"type": "text",
|
| 1904 |
+
"text": "D USER STUDY ",
|
| 1905 |
+
"text_level": 1,
|
| 1906 |
+
"bbox": [
|
| 1907 |
+
174,
|
| 1908 |
+
523,
|
| 1909 |
+
320,
|
| 1910 |
+
541
|
| 1911 |
+
],
|
| 1912 |
+
"page_idx": 14
|
| 1913 |
+
},
|
| 1914 |
+
{
|
| 1915 |
+
"type": "text",
|
| 1916 |
+
"text": "In Section 4.1 we conduct a user study for the text-to-motion task. We asked 31 users to choose between MDM and state-of-the-art works in a side-by-side view, with both samples generated from the same text prompt randomly sampled from the KIT test set. We repeated this process with 10 samples per model and 10 repetitions per sample. This user study enabled a comparison with the recent TEMOS model (Petrovich et al., 2022), which was not included in the HumanML3D benchmark. Fig. 4 shows that most of the time, MDM was preferred over the compared models, and even preferred over ground truth samples in $4 2 . 3 \\%$ of the cases. This user study was designed to measure the precision of the models, i.e. which one better fits the input text. The exact phrasing of the question was “Which animation better fits the following description?”. A sample question from this study is presented in Fig. 5. ",
|
| 1917 |
+
"bbox": [
|
| 1918 |
+
173,
|
| 1919 |
+
556,
|
| 1920 |
+
825,
|
| 1921 |
+
696
|
| 1922 |
+
],
|
| 1923 |
+
"page_idx": 14
|
| 1924 |
+
},
|
| 1925 |
+
{
|
| 1926 |
+
"type": "image",
|
| 1927 |
+
"img_path": "images/08840b6b387e4d6c1ee66ffc62da485b77bf28066ea8a312462a7e54006beead.jpg",
|
| 1928 |
+
"image_caption": [
|
| 1929 |
+
"Figure 5: An example question for our text-to-motion user study, using the Google Forms platform. "
|
| 1930 |
+
],
|
| 1931 |
+
"image_footnote": [],
|
| 1932 |
+
"bbox": [
|
| 1933 |
+
269,
|
| 1934 |
+
382,
|
| 1935 |
+
723,
|
| 1936 |
+
617
|
| 1937 |
+
],
|
| 1938 |
+
"page_idx": 15
|
| 1939 |
+
}
|
| 1940 |
+
]
|
parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/SJ1kSyO2jwu/SJ1kSyO2jwu_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Uczck6TlSZ/Uczck6TlSZ.md
ADDED
|
@@ -0,0 +1,383 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Generating Images with Multimodal Language Models
|
| 2 |
+
|
| 3 |
+
Jing Yu Koh Carnegie Mellon University jingyuk@cs.cmu.edu
|
| 4 |
+
|
| 5 |
+
Daniel Fried Carnegie Mellon University dfried@cs.cmu.edu
|
| 6 |
+
|
| 7 |
+
Ruslan Salakhutdinov Carnegie Mellon University rsalakhu@cs.cmu.edu
|
| 8 |
+
|
| 9 |
+
# Abstract
|
| 10 |
+
|
| 11 |
+
We propose a method to fuse frozen text-only large language models (LLMs) with pre-trained image encoder and decoder models, by mapping between their embedding spaces. Our model demonstrates a wide suite of multimodal capabilities: image retrieval, novel image generation, and multimodal dialogue. Ours is the first approach capable of conditioning on arbitrarily interleaved image and text inputs to generate coherent image (and text) outputs. To achieve strong performance on image generation, we propose an efficient mapping network to ground the LLM to an off-the-shelf text-to-image generation model. This mapping network translates hidden representations of text into the embedding space of the visual models, enabling us to leverage the strong text representations of the LLM for visual outputs. Our approach outperforms baseline generation models on tasks with longer and more complex language. In addition to novel image generation, our model is also capable of image retrieval from a prespecified dataset, and decides whether to retrieve or generate at inference time. This is done with a learnt decision module which conditions on the hidden representations of the LLM. Our model exhibits a wider range of capabilities compared to prior multimodal language models. It can process image-and-text inputs, and produce retrieved images, generated images, and generated text — outperforming non-LLM based generation models across several text-to-image tasks that measure context dependence.
|
| 12 |
+
|
| 13 |
+
# 1 Introduction
|
| 14 |
+
|
| 15 |
+
Autoregressive language models (LMs) and large language models (LLMs) trained on text corpora have shown impressive abilities to efficiently adapt to other modalities. Prior work showcased the effectiveness of grounding text-only LMs to images for vision-and-language tasks [56, 4, 29, 33, 31, 35], to embodied settings for robotics [3, 18], offline reinforcement learning [48], and more. These methods typically keep most of the LLM weights frozen. This allows them to leverage the capabilities that the LLM learns during large scale text-only pretraining, such as the ability to learn from in-context examples [9], more effectively process longer context, and condition on inputs more strongly.
|
| 16 |
+
|
| 17 |
+
In this work, we tackle the task of extending multimodal language models to generate novel images. Our approach, Generating Images with Large Language Models (GILL), is capable of processing arbitrarily interleaved image-and-text inputs to generate text, retrieve images, and generate novel images (Fig. 1). Our findings show that it is possible to efficiently map the output embedding space of a frozen text-only LLM to that of a frozen generation model (in this work, Stable Diffusion [49]) despite both models using entirely different text encoders. We achieve this by finetuning a small number of parameters on image-caption pairs [52], in contrast to other methods which require interleaved image-text data [4, 2]. Our approach is computationally efficient and does not require running the image generation model at training time. To achieve strong image generation performance, we propose efficient architectural changes to learn the LLM-to-generation mapping effectively with the GILLMapper module. GILLMapper is a lightweight Transformer [57] conditioned on special learnt text tokens. We train it by minimizing the $l _ { 2 }$ distance between its outputs and the outputs of the text encoder of a text-to-image generation model. This distillation training allows us to use the image decoder of the text-to-image model at inference time. Despite its simplicity, we show that this allows us to outperform the baseline text-to-image generation model on several tasks that measure language context dependence. Finally, to decide whether to produce a retrieved image or a generated one at inference time, we train a decision model that outputs a decision conditioned on the LM hidden representations. This allows us to both generate and retrieve in output sequences, as shown in Fig. 1.
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
Figure 1: Our model is capable of generating text, retrieving images, generating novel images, and interleaving results into coherent multimodal dialogue.
|
| 21 |
+
|
| 22 |
+
Our experimental results demonstrate that GILL is more effective than Stable Diffusion at processing longer-form text, including dialogue and discourse. We show on dialogue-conditioned image generation that GILL can outperform non-LLM based generation models, and benefit from multimodal context: generating images that match text better than the backbone generation models that we distill from. In addition, GILL can process arbitrarily interleaved image-text inputs, unlike typical text-to-image models which only process text. GILL is the first model capable of outputting retrieved images, novel images, and text — interleaving these for coherent multimodal dialogue generation.1
|
| 23 |
+
|
| 24 |
+
# 2 Related Work
|
| 25 |
+
|
| 26 |
+
Multimodal Language Models Several prior works have developed multimodal language models which process image and text inputs to generate text outputs. Frozen [56] showed that it is possible to finetune a visual encoder to map images into the hidden space of a text-only LLM, and that this exhibits compelling few-shot, captioning, and question answering abilities. Other methods improve upon this approach by introducing adapters [19], scaling up model and data sizes [4, 64], improving the visual encoder [4, 33], finetuning on instructions [35], and training unified models on multi-task objectives [36, 63, 42]. CM3 [2, 62] trained multimodal LMs on HTML webpages consisting of interleaved images and text. Many state-of-the-art models also require significant computational resources to train. For example, Flamingo [4] is trained on 1535 TPUs for 15 days, while RA-CM3 [62] use 256 GPUs for 5 days. In contrast, our efficient adaptation method is trained on 2 GPUs for 2 days. The most similar work to our approach is FROMAGe [31], which trains a multimodal language model capable of processing arbitrarily interleaved image and text inputs to generate text interleaved with retrieved images. While FROMAGe can only retrieve images in their outputs, GILL is capable of both image retrieval and image generation, which allows it to outperform retrieval-only models when they are limited by their candidate retrieval set (Fig. 5).
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
Figure 2: GILL model architecture overview. It is trained with a captioning loss to learn to process images (left), and losses for image retrieval and image generation to learn to produce images (right).
|
| 30 |
+
|
| 31 |
+
Large Language Models Our work leverages recent advances in Transformer-based [57] LLMs. When trained at large enough scale, LLMs exhibit compelling properties, such as the ability to learn from few-shot in-context examples [9, 11] and generate and process long text inputs [61, 59, 53, 7]. Our approach also builds upon recent efforts on open sourced LLM weights [69, 55].
|
| 32 |
+
|
| 33 |
+
Text-to-Image Generation Text-to-image generation is the task of synthesizing a realistic image conditioned on natural language descriptions. [47] was one of the first to tackle this with a conditional GAN [23]. Later work improved upon this by introducing multi-stage models [67], attention mechanisms [60], and contrastive methods [73, 66]. Several recent papers also formulate the text-to-image generation task as a sequence modeling problem [45, 17, 13], training large Transformer [57] models on discretized image tokens [46]. [20, 65] improved upon this approach by introducing stronger image quantizers and scaling up model parameters. Several recent methods [38, 44, 49] apply diffusion models [26] to improve generated image quality. [50, 65] scale up text encoder models to achieve significant gains in generating relevant images. In contrast with computationally intensive methods that train end-to-end, GILL does not require running the image generation model during training.
|
| 34 |
+
|
| 35 |
+
# 3 Method
|
| 36 |
+
|
| 37 |
+
We efficiently adapt a pretrained autoregressive language model of text, to process image and text inputs and produce image and text outputs. Most of the model weights (including those of the base LLM and image generator) are kept frozen, and we finetune a small number of parameters on image caption data (Fig. 2) to achieve a wide range of capabilities (Fig. 5). There are several challenges that need to be resolved. The model needs to learn to process image-and-text content (Sec. 3.1). It also needs to learn to produce images (either retrieved or generated), and determine whether to produce text or images at each step (Sec. 3.2). Finally, whenever an image is produced, the model needs to decide whether image retrieval (from a candidate set) or generation is more appropriate (Sec. 3.3).
|
| 38 |
+
|
| 39 |
+
# 3.1 Learning to Process Images
|
| 40 |
+
|
| 41 |
+
Given an image $x$ and its text caption $y$ (tokenized as $( s _ { 1 } , \ldots , s _ { T } ) ,$ ), our goal is to adapt a frozen LLM to enable it to complete any sequence of arbitrarily interleaved image and text inputs. For example, inputs for the Visual Storytelling dataset [28] consist of 5 images and 5 text descriptions, interleaved in a manner such as $( x _ { 1 } , y _ { 1 } , \dotsc , x _ { 5 } , y _ { 5 } )$ . We follow prior work [56, 19, 35, 31] in learning translation parameters that map from image features to text embedding space.
|
| 42 |
+
|
| 43 |
+
We first extract visual embeddings $v _ { \phi } ( x ) \in \mathbb { R } ^ { d }$ with a pretrained visual backbone (its weights $\phi$ and LLM weights $\theta$ are kept frozen). We learn a linear mapping $\mathbf { W } _ { \mathrm { c a p } } \in \mathbb { R } ^ { d \times k e }$ which maps $v _ { \phi } ( x )$ into a sequence of $k$ $e$ -dimensional vectors that we use as inputs to the LLM (Fig. 2, left), where $e$ is the
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
Figure 3: Inference time procedure for GILL. The model takes in image and text inputs, and produces text interleaved with image embeddings. After deciding whether to retrieve or generate for a particular set of tokens, it returns the appropriate image outputs.
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
Figure 4: GILLMapper model architecture. It is conditioned on the hidden [IMG] representations and a sequence of learnt query embedding vectors.
|
| 50 |
+
|
| 51 |
+
LLM input embedding dimension. We train $\mathbf { W } _ { \mathrm { c a p } }$ on image-caption pairs (details in Sec. 3.4), by minimizing the negative log-likelihood loss of the token sequence $( s _ { 1 } , \ldots , s _ { T } )$ :
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
l _ { c } ( x , y ) = - \sum _ { t = 1 } ^ { T } \log p _ { \boldsymbol \theta } ( s _ { t } \mid v _ { \boldsymbol \phi } ( x ) ^ { T } \mathbf { W } _ { \mathrm { c a p } } , s _ { 1 } , \ldots , s _ { t - 1 } )
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
Intuitively, this objective trains a mapping $\mathbf { W } _ { \mathrm { c a p } }$ that allows us to translate images into embedding vectors in the token embedding space of the LLM (illustrated in Fig. 2, left).
|
| 58 |
+
|
| 59 |
+
# 3.2 Learning to Produce Images
|
| 60 |
+
|
| 61 |
+
In order to enable the model to produce image outputs, we add special [IMG] tokens to the vocabulary of the LLM, similar to [71, 31] which introduce special tokens correspond to images that should be output by the model. The hidden states that the LLM produces for these tokens will be used to retrieve or generate images. While [31] use a single token for their image retrieval model, we observed in our experiments that image generation requires much more finegrained textual information (Sec. 5). In order to improve the expressivity of the frozen LLM for novel image generation, we generalize to use $r$ tokens $[ \mathrm { I M G 1 } ] , \dots , [ \mathrm { I M G } \{ \mathbf { r } \} ]$ for representing visual outputs.
|
| 62 |
+
|
| 63 |
+
Concretely, we add a trainable matrix $\mathbf { E } _ { \mathrm { i m g } } \in \mathbb { R } ^ { r \times e }$ to the embedding matrix of the frozen LLM, which represents the $r$ [IMG] token embeddings. We wish to train the model to learn when it should produce [IMG] tokens. This is done by minimizing the negative log-likelihood of producing the first [IMG] token conditioned on previously generated tokens:
|
| 64 |
+
|
| 65 |
+
$$
|
| 66 |
+
l _ { p } ( y ) = - \log p _ { \{ \theta \cup { \bf E } _ { \mathrm { i m g } } \} } ( { \tt \small { [ I M G 1 ] } } \ | \ s _ { 1 } , \dots , s _ { t } )
|
| 67 |
+
$$
|
| 68 |
+
|
| 69 |
+
The LLM weights $\theta$ are kept frozen, and we only update $\mathbf { E } _ { \mathrm { i m g } }$ . During inference, we always generate the [IMG2], . . . , $\left[ \mathrm { I M G } \{ \bf r \} \right]$ tokens whenever the first [IMG1] token is produced. During training, the [IMG] tokens are appended to each caption (Fig. 2). The LLM hidden states of the [IMG] tokens are used for image retrieval and generation, as described in the following sections.
|
| 70 |
+
|
| 71 |
+
Novel Image Generation In order for our LLM to produce image outputs, the [IMG] tokens need to be mapped into a semantically meaningful region of the input space of an image generation model $G _ { \psi }$ (such as that of the Stable Diffusion [49] image decoder). In initial experiments, we found that training a simple linear mapping such as those used in previous work on retrieval [31] was insufficient, and that such a model was unable to handle more complex prompts (see Sec. 5 for analysis). Hence, we propose GILLMapper (Fig. 4), a lightweight 4-layer encoder-decoder transformer model with trainable weights $\omega$ . The GILLMapper module $f _ { \omega }$ conditions on $h _ { \{ \theta \cup { \bf E } _ { \mathrm { i m g } } \} } \left( y , \thinspace [ \mathrm { I M G } ] \right)$ (the [IMG] representations from the last hidden layer of the LLM) and $L$ learnt query embeddings $( q _ { 1 } , \dots , q _ { L } ) \in \mathbb { R } ^ { L \times m }$ (where $L$ is the maximum input sequence length of the text-to-image generation backbone $G _ { \psi }$ ).
|
| 72 |
+
|
| 73 |
+
The purpose of introducing learnable query embeddings is to enable GILLMapper to extract sequences of $L$ features from the LLM [IMG] hidden states. This is similar to the queries introduced in DETR [10] for object detection and BLIP-2 [33] for extracting image features. We optimize the GILL trainable weights $( q _ { 1 } , \dots , q _ { L }$ and $\omega$ ) by minimizing the MSE loss of the GILLMapper model outputs against the embeddings produced by the text encoder $( T _ { \psi } )$ of a frozen text-to-image generation model:
|
| 74 |
+
|
| 75 |
+
$$
|
| 76 |
+
l _ { g } ( y ) = \parallel f _ { \omega } \left( h _ { \{ \theta \cup { \bf E } _ { \mathrm { i n g } } \} } ( y , [ \mathrm { I M G } \{ 1 \} ] ) , \dots , h _ { \{ \theta \cup { \bf E } _ { \mathrm { i n g } } \} } ( y , [ \mathrm { I M G } \{ { \bf r } \} ] ) , q _ { 1 } , \dots , q _ { L } \right) - T _ { \psi } ( y ) \parallel _ { 2 } ^ { 2 } .
|
| 77 |
+
$$
|
| 78 |
+
|
| 79 |
+
This is essentially distilling from $T _ { \psi }$ to learn a valid mapping from the output representations of our frozen LLM to the input space of $G _ { \psi }$ . Note that this does not require $G _ { \psi }$ during training, so we can precompute $T _ { \psi } ( y )$ ahead of time, making training highly efficient. During inference, when [IMG] tokens are generated, we can synthesize an image by applying GILLMapper and the decoder $G _ { \psi }$ :
|
| 80 |
+
|
| 81 |
+
$$
|
| 82 |
+
\mathrm { G e n e r a t e d ~ I m a g e } = G _ { \psi } ( f _ { \omega } ( h _ { \{ \theta \cup \mathbf { E } _ { \mathrm { i n g } } \} } ( y , [ \mathbb { M } \mathbb { G } ( 1 ) ] ) , \dots , h _ { \{ \theta \cup \mathbf { E } _ { \mathrm { i n g } } \} } ( y , [ \mathbb { M } \mathbb { G } ( \mathbf { r } ) ] ) , q _ { 1 } , \dots , q _ { L } ) )
|
| 83 |
+
$$
|
| 84 |
+
|
| 85 |
+
where $h _ { \{ \theta \cup { \bf E } _ { \mathrm { i m g } } \} } \left( y , \thinspace [ \mathrm { I M G } \{ \mathrm { i } \} ] \right)$ represents the hidden states from the last hidden layer of the modified LLM corresponding to the $i ^ { t h }$ [IMG] token. The learnt query embeddings $( q _ { 1 } , \dots , q _ { L } )$ are part of the GILLMapper model weights, and are hence kept fixed during inference.
|
| 86 |
+
|
| 87 |
+
Image Retrieval Similar to [31], we learn a linear mapping $\mathbf { W _ { \mathrm { t } 2 i } } \in \mathbb { R } ^ { e \times p }$ that maps the first token ([IMG1] to a $p$ -dimensional vector. We also learn a linear mapping $\mathbf { W _ { \mathrm { i 2 t } } } \in \mathbb { R } ^ { d \times p }$ that map the pooled visual output of the image encoder $v _ { \phi } ( x )$ to a $p$ -dimensional space. These represent image and text embeddings, and we train the model by minimizing the InfoNCE loss [39]:
|
| 88 |
+
|
| 89 |
+
$$
|
| 90 |
+
l _ { r } ( \mathbf { x } _ { i } , \mathbf { y } _ { i } ) = - \log \frac { \exp ( \sin ( \mathbf { x } _ { i } , \mathbf { y } _ { i } , \mathbf { W } _ { \mathrm { t 2 i } } ) / \tau ) } { \sum _ { j = 1 } ^ { N } \exp ( \sin ( \mathbf { x } _ { j } , \mathbf { y } _ { i } , \mathbf { W } _ { \mathrm { t 2 i } } ) / \tau ) } - \log \frac { \exp ( \sin ( \mathbf { x } _ { i } , \mathbf { y } _ { i } , \mathbf { W } _ { \mathrm { i 2 t } } ) ) / \tau ) } { \sum _ { j = 1 } ^ { N } \exp ( \sin ( \mathbf { x } _ { i } , \mathbf { y } _ { j } , \mathbf { W } _ { \mathrm { i 2 t } } ) ) / \tau ) }
|
| 91 |
+
$$
|
| 92 |
+
|
| 93 |
+
where the similarity is computed as
|
| 94 |
+
|
| 95 |
+
$$
|
| 96 |
+
\sin ( x , y , \mathbf { W } ) = \frac { \left( \mathbf { W } ^ { T } v _ { \phi } ( x ) \right) ^ { T } \left( \mathbf { W } ^ { T } h _ { \{ \theta \cup \mathbf { E } _ { \mathrm { i m g } } \} } ( y , \ \left[ \mathrm { I M G 1 } \right] ) \right) } { \| \mathbf { W } ^ { T } v _ { \phi } ( x ) \| \left\| \mathbf { W } ^ { T } h _ { \{ \theta \cup \mathbf { E } _ { \mathrm { i m g } } \} } ( y , \ \left[ \mathrm { I M G 1 } \right] ) \right\| }
|
| 97 |
+
$$
|
| 98 |
+
|
| 99 |
+
During inference, we follow standard procedure [43] in retrieving the image with the highest cosine similarity (between image embeddings and the [IMG] tokens) from a candidate set of images.
|
| 100 |
+
|
| 101 |
+
# 3.3 Deciding to Generate or Retrieve
|
| 102 |
+
|
| 103 |
+
While learning to produce [IMG] tokens allows us to decide when to interleave images in text , the task of deciding whether to retrieve or generate from [IMG] tokens remains. Intuitively, for a given prompt, we would like to retrieve when there is a strong match from our set of candidate images, and generate otherwise. In order to evaluate this, we collect human annotations on PartiPrompts (P2) [65], a collection of prompts used to benchmark image generation models. P2 contains some prompts that are well-represented by naturally occurring images, but others that are unlikely to occur in natural image sets, making it a test of generative models. For each of the 1,632 examples in P2, we generate an image with the text-to-image generation model $G _ { \psi }$ , and use the CLIP ViT-L [43] model to retrieve the top ranked image from CC3M [52] according to the cosine similarity of image embeddings $v _ { \phi }$ .
|
| 104 |
+
|
| 105 |
+
We have 5 independent human annotators (details in the appendix) select which of the two images for each prompt, retrieved or generated, is better matched to the prompt. We labeled the examples where the generated image was selected as ‘gen’ (indicating prompts which we should generate an image for) and ‘ret’ for prompts that should have an image retrieved. We extract the most confident set of these annotations (retaining roughly 900 examples with an inter-annotator agreement of at least $4 / 5$ ), and split them into a $67 \%$ train (600) and $33 \%$ test (300) split. We use this to train a linear classifier on the LLM [IMG] hidden states as a decision model for deciding when to retrieve or generate (more details and baselines are provided in the appendix). Although these annotations of retrieving versus generating are somewhat model dependent, we believe that this data is still a valuable metric during model development. We will release our annotations to encourage future work in this space.
|
| 106 |
+
|
| 107 |
+
# 3.4 Data and Implementation Details
|
| 108 |
+
|
| 109 |
+
The final training objective for a batch of image-text pairs $\displaystyle ( \mathbf { x } , \mathbf { y } )$ is the sum of the captioning loss $l _ { c }$ (Eq. 1), image token prediction loss $l _ { p }$ (Eq. 2), generation loss $l _ { g }$ (Eq. 3) and retrieval loss $l _ { r }$ (Eq. 4):
|
| 110 |
+
|
| 111 |
+
$$
|
| 112 |
+
\operatorname* { m i n } _ { \mathbf { W _ { i 2 1 } , W _ { r 2 4 } , W _ { c a p } , E _ { i m g } , } \omega , q _ { 1 : L } } \frac { 1 } { N } \sum _ { i = 1 } ^ { N } \left( l _ { c } ( \mathbf { x } _ { i } , \mathbf { y } _ { i } ) + l _ { p } ( \mathbf { y } _ { i } ) + l _ { g } ( \mathbf { y } _ { i } ) + l _ { r } ( \mathbf { x } _ { i } , \mathbf { y } _ { i } ) \right)
|
| 113 |
+
$$
|
| 114 |
+
|
| 115 |
+
The decision model is trained separately after convergence of the other components. The multitask loss (Eq. 5) trains GILL to process images $( l _ { c } )$ , produce [IMG] tokens $( l _ { p } )$ , generate images $( l _ { g } )$ , and retrieve images $( l _ { r } )$ . This enables it to generalize to a wide range of vision-and-language tasks.
|
| 116 |
+
|
| 117 |
+
We train on Conceptual Captions (CC3M) [52], which consists of $3 . 3 { \mathrm { M } }$ image-text pairs. Following [31], we pack two random examples together during training with probability 0.5 (i.e., $50 \%$ of the time, the input is a single image and caption example, while the other $50 \%$ of the time the input consists of a sequence consisting of two interleaved images and captions). We use the OPT-6.7B [69] model as the LLM backbone (which produce hidden states $h _ { \theta }$ with embedding dim $e = 4 0 9 6$ ). For the visual model used to extract features $v _ { \phi }$ for captioning and retrieval, we use the CLIP [43] ViT-L model. For our text-to-image generation backbone $G _ { \psi }$ , we use the Stable Diffusion [49] v1.5 model (with $L = 7 7$ input vectors).2 We use $k = 4$ visual tokens, and $r = 8$ learnt [IMG] tokens. We set the GILLMapper query embedding dimension $m = 5 1 2$ . For retrieval, we use an embedding dimension $p = 2 5 6$ . All pretrained model weights are kept frozen, and we only train the linear layers $\mathbf { W } _ { \mathrm { i 2 t } }$ , $\mathbf { W } _ { \mathrm { t 2 i } }$ , $\mathbf { W } _ { \mathrm { c a p } }$ , the [IMG] embedding matrix $\mathbf { E } _ { \mathrm { i m g } }$ , and the GILLMapper parameters $\omega$ and query vectors $q _ { 1 : L }$ . In total, there are 50M trainable parameters, significantly fewer than in the frozen LLM and visual models (which total approximately 8B parameters). We use bfloat16 precision [1], and optimize using Adam [30] $\beta _ { 1 } = 0 . 9$ , $\beta _ { 2 } = 0 . 9 5 )$ ) with a learning rate of 0.001. We train with a batch size of 200 for 20K iterations, which takes 2 days on 2 A6000 GPUs. We follow [31] and concatenate captions to encourage the model to attend to relevant images within an image-text sequence.
|
| 118 |
+
|
| 119 |
+
# 4 Experiments
|
| 120 |
+
|
| 121 |
+
GILL is the first multimodal language model capable of conditioning on image-and-text inputs to generate meaningful images interleaved with text. Hence, our experiments primarily focus on evaluating its ability to produce novel images (Sec. 4.1). Our results show that GILL improves over Stable Diffusion [49] on tasks that require processing long-form text such as dialogue and discourse. We also benchmark the performance of models in deciding whether to retrieve or generate (see appendix). GILL is capable of generating text, retrieving images, and generating images. Despite being more general than prior work [56, 4, 31], we find that GILL performs comparably to or better than existing multimodal LMs on contextual image retrieval and text generation tasks (see Sec. 5).
|
| 122 |
+
|
| 123 |
+
# 4.1 Contextual Image Generation
|
| 124 |
+
|
| 125 |
+
To test the ability of our model against baseline methods for novel image generation, we run experiments on the VIST [28] and VisDial [16] datasets. These are the same datasets used in prior work [31] for benchmarking image retrieval conditioned on multimodal text-and-image context.
|
| 126 |
+
|
| 127 |
+
Evaluation Metrics The focus of our evaluation is on the ability of generative models to handle complex language descriptions. Hence, we compute metrics which measure the relevance of the generated image content. We evaluate models with two metrics:
|
| 128 |
+
|
| 129 |
+
1. CLIP Similarity: We use the CLIP [43] ViT-L image encoder to produce pooled representations of generated images and the corresponding real images, and report their cosine similarity. A higher score indicates that a generated image is more similar to the real image. 2. Learned Perceptual Image Patch Similarity (LPIPS): LPIPS [68] evaluates the distance between image patches. We measure LPIPS between real and generated images. A lower value indicates that two images are closer in perceptual space (i.e., more similar), while a higher value indicates that two images are more dissimilar.
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
Our model composites multimodal information to produce relevant image and text outputs. It can outperform baseline models that are limited to image retrieval.
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
Figure 5: Qualitative results over various input and output modalities. GILL is able to process contextual multimodal cues to retrieve and generate appropriate image and text outputs.
|
| 136 |
+
|
| 137 |
+
Table 1: Results on contextual image generation on VIST [28] (averaged over 5 random seeds). Our model can process longer (possibly multimodel) inputs to outperform baseline models.
|
| 138 |
+
|
| 139 |
+
<table><tr><td rowspan="2">Model</td><td colspan="3">CLIP Similarity (↑)</td><td colspan="3">LPIPS (↓)</td></tr><tr><td>1 caption</td><td>5 captions</td><td>5 caps,4images</td><td>1 caption</td><td>5 captions</td><td>5 caps,4images</td></tr><tr><td>GLIDE [38]</td><td>0.582</td><td>0.591</td><td>-</td><td>0.753</td><td>0.745</td><td>-</td></tr><tr><td>Stable Diffusion [49]</td><td>0.592 ±0.0007</td><td>0.598 ±0.0006</td><td>-</td><td>0.703 ±0.0003</td><td>0.704±0.0004</td><td>-</td></tr><tr><td>GILL (ours)</td><td>0.581 ±0.0005</td><td>0.612 ±0.0011</td><td>0.641 ±0.0011</td><td>0.702 ±0.0004</td><td>0.696 ±0.0008</td><td>0.693 ±0.0008</td></tr></table>
|
| 140 |
+
|
| 141 |
+
Table 2: Results on contextual image generation on VisDial [16] (averaged over 5 random seeds). Our model can process longer sequences of dialogue-like text to generate more relevant images.
|
| 142 |
+
|
| 143 |
+
<table><tr><td></td><td colspan="3">CLIP Similarity (↑)</td><td colspan="3">LPIPS (↓)</td></tr><tr><td>Model</td><td>1 round</td><td>5 rounds</td><td>10 rounds</td><td>1 round</td><td>5 rounds</td><td>10 rounds</td></tr><tr><td>GLIDE [38]</td><td>0.562</td><td>0.595</td><td>0.587</td><td>0.800</td><td>0.794</td><td>0.799</td></tr><tr><td>Stable Diffusion [49]</td><td>0.552 ±0.0015</td><td>0.629 ±0.0015</td><td>0.622 ±0.0012</td><td>0.742 ±0.0010</td><td>0.722 ±0.0012</td><td>0.723 ±0.0008</td></tr><tr><td>GILL (ours)</td><td>0.528 ±0.0014</td><td>0.621 ±0.0009</td><td>0.645 ±0.0010</td><td>0.742 ±0.0022</td><td>0.718±0.0028</td><td>0.714±0.0006</td></tr></table>
|
| 144 |
+
|
| 145 |
+
Generating from Visual Stories VIST [28] is a dataset for sequential vision-and-language tasks, with examples of sequences of 5 images and text that constitute a story, as shown in Fig. 5. Similar to [31], we test the models on generating the last image in the sequence, conditioned on different inputs:
|
| 146 |
+
|
| 147 |
+
1. 1 caption: Input consists of the last text description. This is similar to standard text-toimage generation, where a model conditions on a single caption to generate an image.
|
| 148 |
+
2. 5 captions: Input consists of all text from the entire story sequence. This tests the ability of models to process longer and temporally dependent text descriptions.
|
| 149 |
+
3. 5 captions, 4 images: Lastly, we test models with inputs of all images and texts preceding the last image (i.e., sequenced as “<text1><img1>...<text4><img4><text5>”). This tests the ability of models to effectively process multimodal context in image generation. A novel feature of GILL is its ability to process interleaved image-text inputs, which most existing text-to-image generation models are unable to handle.
|
| 150 |
+
|
| 151 |
+
We report results on VIST in Tab. 1, comparing GILL against text-to-image generation baselines (including Stable Diffusion (SD) [49], which we use as our generation backbone $G _ { \psi }$ ). With a single story caption input to both models, the performance is comparable, with SD achieving a slightly better CLIP Similarity score, and both models achieving similar LPIPS. However, when all 5 story captions are provided as input, our model outperforms SD, improving CLIP Similarity from 0.598 to 0.612, and LPIPS from 0.704 to 0.696. Interestingly, when further provided with the full multimodal context (the preceding 5 captions and 4 images), our model improves substantially, attaining a CLIP Similarity of 0.641 and LPIPS of 0.693. In contrast, SD is unable to handle interleaved image-text inputs without significant modifications. We also show several qualitative examples in Fig. 5. We find that GILL is generally more sensitive to input context compared to SD. GILL can also condition on image inputs, enabling it to use visual context to produce more relevant images.
|
| 152 |
+
|
| 153 |
+
We highlight that both models use the same image generation backbone, and the primary difference is in their text encoders. GILL is able to better handle long text inputs and multimodal context, which we attribute to the stronger LLM encoder coupled with our GILLMapper model.
|
| 154 |
+
|
| 155 |
+
Generating from Visual Dialogue We also test our model on the VisDial [16] dataset. VisDial examples contain a sequence of question and answer (Q&A) pairs about a particular image, simulating dialogue between two people who are discussing an image. Examples contain up to 10 rounds of Q&A dialogue pairs. Similar to VIST, we evaluate the ability of models to accurately synthesize the image being described, provided with increasing amounts of the Q&A dialogue context as input. This experiment tests the ability of our approach to (1) generalize to dialogue-like text (as our approach is only finetuned on image caption data), and (2) process long text sequences.
|
| 156 |
+
|
| 157 |
+
Our results are presented in Tab. 2. Similar to the VIST evaluations, we find that with shorter length inputs, SD outperforms our model. However, when the input context is increased, our model gradually improves, and can synthesize images that are more similar to the groundtruth image. When the full 10 rounds of dialogue are provided, GILL significantly outperforms SD, improving over both CLIP
|
| 158 |
+
|
| 159 |
+
Table 3: Image generation performance on CC3M [52] and VIST [28] with different text mapping networks.
|
| 160 |
+
|
| 161 |
+
<table><tr><td></td><td>CC3M</td><td>VIST</td></tr><tr><td>Model</td><td>FID (↓)</td><td>CLIP Sim (↑)</td></tr><tr><td>Stable Diffusion [49]</td><td>13.94</td><td>0.598</td></tr><tr><td>Ours + Linear</td><td>15.50</td><td>0.500</td></tr><tr><td>Ours + 3-layerMLP</td><td>15.33</td><td>0.502</td></tr><tr><td>Ours + Transformer Encoder</td><td>16.30</td><td>0.605</td></tr><tr><td>Ours + GILLMapper</td><td>15.31</td><td>0.641</td></tr></table>
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
Figure 6: Performance of GILL on VIST generation.
|
| 165 |
+
|
| 166 |
+
Table 4: GILL image generation results on CC3M [52] with different number of image tokens $( r )$ .
|
| 167 |
+
|
| 168 |
+
<table><tr><td rowspan="2">r</td><td>CC3M</td><td>VIST</td></tr><tr><td>FID (↓)</td><td>CLIP Sim (↑)</td></tr><tr><td>1</td><td>15.93</td><td>0.631</td></tr><tr><td>2</td><td>15.32</td><td>0.629</td></tr><tr><td>4</td><td>15.32</td><td>0.642</td></tr><tr><td>8</td><td>15.31</td><td>0.641</td></tr></table>
|
| 169 |
+
|
| 170 |
+
Table 5: Contextual image retrieval on VIST (5 captions, 4 images as input). † indicates results from [31].
|
| 171 |
+
|
| 172 |
+
<table><tr><td rowspan="2">Model</td><td colspan="3">VIST Recall@ k (↑)</td></tr><tr><td>R@1</td><td>R@5</td><td>R@10</td></tr><tr><td>CLIP ViT-L [43]t</td><td>8.8</td><td>22.3</td><td>29.8</td></tr><tr><td>FROMAGe [31]t</td><td>18.2</td><td>42.7</td><td>51.8</td></tr><tr><td>GILL (Ours)</td><td>20.3</td><td>45.0</td><td>53.7</td></tr></table>
|
| 173 |
+
|
| 174 |
+
Similarity (0.622 to 0.645) and LPIPS (0.723 to 0.714). These results further highlight the efficacy of our model on handling long dialogue-like text inputs.
|
| 175 |
+
|
| 176 |
+
# 4.2 Qualitative Results
|
| 177 |
+
|
| 178 |
+
Finally, one of the more compelling applications of GILL is perhaps its ability to generalize to many different tasks, due to the LLM pretraining and freezing. We showcase several of these capabilities in Fig. 5. In many examples, we observed that GILL is able to outperform retrieval models such as FROMAGe [31] on examples where FROMAGe is unable to retrieve relevant images. GILL is also generally more sensitive to input context compared to Stable Diffusion [49], and can condition on image inputs, in addition to text, to generate more visually and semantically relevant image outputs.
|
| 179 |
+
|
| 180 |
+
# 5 Analysis
|
| 181 |
+
|
| 182 |
+
Contextual Image Retrieval In addition to generation, GILL is capable of image retrieval conditioned on image-text inputs. We run GILL on the VIST retrieval evaluation from [31]. We find that GILL performs comparably or better compared to prior approaches (Tab. 5). This shows that that the image generation objective does not cause image retrieval performance to deteriorate.
|
| 183 |
+
|
| 184 |
+
The Effect of Context GILL leverages an LLM backbone, which allows it to inherit some of the LLM’s capabilities, such as improved sensitivity to long inputs. Fig. 6 shows that the performance of GILL generally improves with increasing input contexts on VIST [28]. In particular, when 2 captions and 1 image are provided as context, the model significantly outperforms the model with 5 text-only captions, highlighting the value of multimodal context over unimodal context.
|
| 185 |
+
|
| 186 |
+
Generation-Only Objective We investigate the effect of removing the retrieval loss (Eq. 4) from the training objective. On VIST (5 captions, 4 images), this ablated model achieves CLIP similarity of 0.636 and LPIPS of 0.694, which are comparable to scores of the original model (0.641 and 0.693 respectively). This suggests that the retrieval loss is not necessary for strong performance, although such a model would only be able to generate images and text and not retrieve images. These results also suggest that GILL is not bottlenecked by including the retrieval objective, and that it has sufficient capacity to perform both generation and retrieval.
|
| 187 |
+
|
| 188 |
+
GILLMapper Module As described in Sec. 3.2, we propose the GILLMapper module, a lightweight transformer model that conditions on [IMG] embeddings and $q$ learnt embedding vectors. The output maps the LM embeddings into the input space of a text-to-image generation model, enabling image synthesis. We run several baselines to compare effectiveness, comparing our proposed model against (1) a linear layer, (2) a multilayer perceptron (MLP) with LeakyReLU activations, and (3) a 4-layer bidirectional transformer encoder. All models are conditioned on the $r$ [IMG] token embeddings from the LLM. Our results are presented in Tab. 3. GILLMapper is substantially better than these baseline models at learning the mapping from the frozen LLM to the Stable Diffusion generation model, as measured by Fréchet Inception Distance (FID) [25] on the CC3M validation set (which is a measure of image realism), and CLIP Similarity on VIST. On the VIST evaluation (which is out of distribution from CC3M), the other baselines perform significantly worse than GILLMapper, suggesting that they cannot generalize to longer sequences containing multiple images and texts.
|
| 189 |
+
|
| 190 |
+
Number of [IMG] Tokens We experiment with varying the number of [IMG] tokens, $r$ (Tab. 4). As $r$ increases, generation generally improves, plateauing around $r = 4$ . We observe that lower values of $r$ appear to result in worse results, as the inputs to GILLMapper are shorter and less expressive.
|
| 191 |
+
|
| 192 |
+
# 6 Conclusion
|
| 193 |
+
|
| 194 |
+
We proposed a method of mapping text-only LLMs to strong visual models. This enables them to learn to process arbitrarily interleaved image-and-text inputs, and output generated text, retrieved images, and generated images. We show that it is possible to efficiently learn a mapping between the embeddings of a frozen pretrained LLM and a frozen pretrained image generation model, and that doing so effectively boosts image generation for tasks that require stronger language context dependence. Finally, we also showcased several compelling qualitative results on a variety of multimodal tasks. Our approach is modular, and can benefit from stronger LLMs or visual models released in the future. Scaling up the LLM backbone, image generation backbone, or visual processing model, are promising directions that will likely induce even stronger vision-and-language capabilities.
|
| 195 |
+
|
| 196 |
+
# Acknowledgements
|
| 197 |
+
|
| 198 |
+
This work was partially supported by a gift from Cisco Systems, and by ONR N000142312368 and DARPA/AFRL FA87502321015. We thank Wendy Kua for help with the figures. We thank Jared Fernandez, Yutong He, Saujas Vaduguru, Yonatan Bisk, and our anonymous reviewers for feedback and helpful discussions on previous versions of this paper.
|
| 199 |
+
|
| 200 |
+
# References
|
| 201 |
+
|
| 202 |
+
[1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467, 2016.
|
| 203 |
+
[2] Armen Aghajanyan, Bernie Huang, Candace Ross, Vladimir Karpukhin, Hu Xu, Naman Goyal, Dmytro Okhonko, Mandar Joshi, Gargi Ghosh, Mike Lewis, et al. Cm3: A causal masked multimodal model of the internet. arXiv preprint arXiv:2201.07520, 2022.
|
| 204 |
+
[3] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, et al. Do as i can, not as i say: Grounding language in robotic affordances. CoRL, 2023.
|
| 205 |
+
[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022.
|
| 206 |
+
[5] Satanjeev Banerjee and Alon Lavie. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, 2005.
|
| 207 |
+
[6] Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610–623, 2021.
|
| 208 |
+
[7] Amanda Bertsch, Uri Alon, Graham Neubig, and Matthew R Gormley. Unlimiformer: Longrange transformers with unlimited length input. arXiv preprint arXiv:2305.01625, 2023.
|
| 209 |
+
[8] Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe. Multimodal datasets: misogyny, pornography, and malignant stereotypes. arXiv preprint arXiv:2110.01963, 2021.
|
| 210 |
+
[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020.
|
| 211 |
+
[10] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020.
|
| 212 |
+
[11] Stephanie CY Chan, Ishita Dasgupta, Junkyung Kim, Dharshan Kumaran, Andrew K Lampinen, and Felix Hill. Transformers generalize differently from information stored in context vs in weights. NeurIPS MemARI Workshop, 2022.
|
| 213 |
+
[12] Keshigeyan Chandrasegaran, Ngoc-Trung Tran, Alexander Binder, and Ngai-Man Cheung. Discovering transferable forensic features for cnn-generated images detection. In ECCV, 2022.
|
| 214 |
+
[13] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In CVPR, 2022.
|
| 215 |
+
[14] Wenhu Chen, Hexiang Hu, Chitwan Saharia, and William W Cohen. Re-imagen: Retrievalaugmented text-to-image generator. ICLR, 2023.
|
| 216 |
+
[15] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $9 0 \% *$ chatgpt quality, March 2023.
|
| 217 |
+
[16] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In CVPR, 2017.
|
| 218 |
+
[17] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. NeurIPS, 2021.
|
| 219 |
+
[18] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.
|
| 220 |
+
[19] Constantin Eichenberg, Sidney Black, Samuel Weinbach, Letitia Parcalabescu, and Anette Frank. Magma–multimodal augmentation of generative models through adapter-based finetuning. EMNLP, 2022.
|
| 221 |
+
[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021.
|
| 222 |
+
[21] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In ECCV, 2022.
|
| 223 |
+
[22] Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. EMNLP, 2020.
|
| 224 |
+
[23] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 2020.
|
| 225 |
+
[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017.
|
| 226 |
+
[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017.
|
| 227 |
+
[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020.
|
| 228 |
+
[27] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. ICLR, 2020.
|
| 229 |
+
[28] Ting-Hao Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, et al. Visual storytelling. In NAACL-HLT, 2016.
|
| 230 |
+
[29] Gabriel Ilharco, Rowan Zellers, Ali Farhadi, and Hannaneh Hajishirzi. Probing contextual language models for common ground with visual representations. NAACL, 2021.
|
| 231 |
+
[30] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ICLR, 2015.
|
| 232 |
+
[31] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. ICML, 2023.
|
| 233 |
+
[32] Nathan Lambert, Louis Castricato, Leandro von Werra, and Alex Havrilla. Illustrating reinforcement learning from human feedback (rlhf). Hugging Face Blog, 2022. https://huggingface.co/blog/rlhf.
|
| 234 |
+
[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. ICML, 2023.
|
| 235 |
+
[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. ECCV, 2014.
|
| 236 |
+
[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023.
|
| 237 |
+
[36] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022.
|
| 238 |
+
[37] Xiyang Luo, Ruohan Zhan, Huiwen Chang, Feng Yang, and Peyman Milanfar. Distortion agnostic deep watermarking. In CVPR, 2020.
|
| 239 |
+
[38] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. ICML, 2022.
|
| 240 |
+
[39] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018.
|
| 241 |
+
[40] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. NeurIPS, 2022.
|
| 242 |
+
[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002.
|
| 243 |
+
[42] Can Qin, Ning Yu, Chen Xing, Shu Zhang, Zeyuan Chen, Stefano Ermon, Yun Fu, Caiming Xiong, and Ran Xu. Gluegen: Plug and play multi-modal encoders for x-to-image generation. arXiv preprint arXiv:2303.10056, 2023.
|
| 244 |
+
[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021.
|
| 245 |
+
[44] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022.
|
| 246 |
+
[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021.
|
| 247 |
+
[46] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019.
|
| 248 |
+
[47] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In ICML, 2016.
|
| 249 |
+
[48] Machel Reid, Yutaro Yamada, and Shixiang Shane Gu. Can wikipedia help offline reinforcement learning? arXiv preprint arXiv:2201.12122, 2022.
|
| 250 |
+
[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022.
|
| 251 |
+
[50] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022.
|
| 252 |
+
[51] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021.
|
| 253 |
+
[52] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018.
|
| 254 |
+
[53] Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. Efficient transformers: A survey. ACM Computing Surveys, 55(6):1–28, 2022.
|
| 255 |
+
[54] Yi Tay, Mostafa Dehghani, Vinh Q Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, and Donald Metzler. Unifying language learning paradigms. arXiv preprint arXiv:2205.05131, 2022.
|
| 256 |
+
[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.
|
| 257 |
+
[56] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurIPS, 2021.
|
| 258 |
+
[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017.
|
| 259 |
+
[58] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. ICLR, 2022.
|
| 260 |
+
[59] Yuhuai Wu, Markus N Rabe, DeLesley Hutchins, and Christian Szegedy. Memorizing transformers. ICLR, 2022.
|
| 261 |
+
[60] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In CVPR, 2018.
|
| 262 |
+
[61] Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian. Doc: Improving long story coherence with detailed outline control. arXiv preprint arXiv:2212.10077, 2022.
|
| 263 |
+
[62] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. ICML, 2023.
|
| 264 |
+
|
| 265 |
+
[63] Haoxuan You, Mandy Guo, Zhecan Wang, Kai-Wei Chang, Jason Baldridge, and Jiahui Yu. Cobit: A contrastive bi-directional image-text generation model. arXiv preprint arXiv:2303.13455, 2023.
|
| 266 |
+
|
| 267 |
+
[64] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022.
|
| 268 |
+
|
| 269 |
+
[65] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022.
|
| 270 |
+
|
| 271 |
+
[66] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to-image generation. In CVPR, 2021.
|
| 272 |
+
|
| 273 |
+
[67] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N. Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV, 2017.
|
| 274 |
+
|
| 275 |
+
[68] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018.
|
| 276 |
+
|
| 277 |
+
[69] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.
|
| 278 |
+
|
| 279 |
+
[70] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023.
|
| 280 |
+
|
| 281 |
+
[71] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022.
|
| 282 |
+
|
| 283 |
+
[72] Yufan Zhou, Chunyuan Li, Changyou Chen, Jianfeng Gao, and Jinhui Xu. Lafite2: Few-shot text-to-image generation. arXiv preprint arXiv:2210.14124, 2022.
|
| 284 |
+
|
| 285 |
+
[73] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuxiang Gu, Jinhui Xu, and Tong Sun. Lafite: Towards language-free training for text-to-image generation. CVPR, 2022.
|
| 286 |
+
|
| 287 |
+
# A Limitations
|
| 288 |
+
|
| 289 |
+
GILL relies on an LLM backbone for many of its capabilities. As such, it also inherits many of the limitations that are typical of LLMs. One limitation is the potential for hallucinations [6], where the model generates content that is false or not relevant to the input data. Another limitation of the model in generating text is in repetitions and neural text degeneration [27], where the model generates the same content multiple times. We also observed that the OPT-6.7B model also does not always consistently generate coherent dialogue text.
|
| 290 |
+
|
| 291 |
+
These limitations may be addressed by techniques that address hallucinations and degenerations in text-only LLMs, or by using improved LLMs that are less prone to these issues. In GILL, we used a 6.7B model. In the future, it will be valuable to scale up the approach with even larger LMs, or those trained with improved objectives [54], instruction finetuning [58] or human feedback [40]. Depending on downstream applications, using models trained explicitly on dialogue data [15] may also be helpful for dialogue capabilities (e.g., deploying multimodal chatbots).
|
| 292 |
+
|
| 293 |
+
With regards to the visual models, another limitation of our approach is in its limited visual processing. At the moment, we use only $k = 4$ visual vectors to represent each input image (due to computational constraints), which may not capture all the relevant visual information needed for downstream tasks. These vectors are produced by a frozen pre-trained visual encoder, and so the visual information in the vectors is heavily constrained by the pre-training task. As a result, the model may not always process images correctly or in enough detail to produce accurate or high-quality results. However, this limitation can potentially be addressed in the future by scaling up the visual model, using models with varied pre-training objectives that encode more visual information while still being mappable to the hidden space of the LLM, or using more sophisticated visual mappings [4, 33] that can capture a richer set of visual features. Similarly, we observed during inference that our model sometimes does not generate relevant images for certain types of prompts. We attribute this to our finetuning dataset being CC3M, which is relatively small compared to modern large scale image-text datasets [51]. It is likely that training GILLMapper on an even larger corpus of text data will improve its alignment to the image generation backbone.
|
| 294 |
+
|
| 295 |
+
One of the advantages of our model is that it is modular, and can benefit from stronger visual and language models released in the future. It is likely that it will also benefit from stronger text-to-image generation backbones, or through finetuning the generation backbone rather than just the GILLMapper module. We leave such scaling explorations for future work.
|
| 296 |
+
|
| 297 |
+
# B Broader Impact
|
| 298 |
+
|
| 299 |
+
AI Assistants Recent advances in dialogue based chatbots have sparked interest in using LLMs for interactive conversational applications. GILL is a multimodal language model capable of processing image and text inputs, and producing image and text outputs. These capabilities may enable a wider range of applications. For example, AI assistants which can produce image and text outputs would be able to answer a wider range of queries, providing visual content when necessary to illustrate certain points. Concrete applications may include creative endeavors (e.g., iteratively refining a generated image with instructions), answering questions that benefit from visual outputs (e.g., describing food items), and more. Scaling GILL and refining it with methods such as reinforcement learning from human feedback (RLHF) [32] are promising directions to improve the capabilities of multimodal AI assistant systems.
|
| 300 |
+
|
| 301 |
+
Disinformation and Harms Aside from the technical limitations detailed in Sec. A, there are broader societal issues that should be considered with the development of generative models of text and images. LLMs have the potential to generate plausible sounding (but false) text [22, 6], propagating disinformation at scale. As GILL uses an LLM backbone, it is also susceptible to these potential issues. Furthermore, as multimodal generative models which can also produce image content, models such as GILL also introduce potential issues with producing even more convincing disinformation through interleaving text with realistic generated images. As GILL makes use of an image generation backbone, it is also susceptible to the risks that typical text-to-image generation models introduce, such as generating false images of real people. These harms may possibly be mitigated by introducing watermarking into generated images [37, 70], or by deploying systems to detect generated images [12].
|
| 302 |
+
|
| 303 |
+
Bias and Safety GILL makes use of pretrained LLMs and multimodal models (such as CLIP [43] and Stable Diffusion [49]), which are trained on large, noisy, Internet-scraped data (such as LAION400M [51]). Due to their curation process, these datasets often contain undesired biases, malignant stereotypes (see [8] for a comprehensive discussion on large scaled multimodal datasets). One advantage of GILL is that it is efficient to train and completely modular, allowing its components (i.e., the LLM, visual encoder, or image generator) to be swapped out for other pretrained models (for example, models which have been further calibrated to reduce unintended biases).
|
| 304 |
+
|
| 305 |
+
Intended Uses GILL is a research prototype which showcases possible capabilities of multimodal language models which can both process and produce image and text outputs. Due to the limitations described above, GILL is not in its current state intended for deployment in practical applications, especially in high risk or sensitive domains without further analysis. At its current model scale (a 6.7B parameter LLM), GILL also lacks many of the abilities of larger language models [9], and applications would likely benefit from increased scaling of the LLM and visual models.
|
| 306 |
+
|
| 307 |
+
# C Deciding to Generate or Retrieve
|
| 308 |
+
|
| 309 |
+
As detailed in Sec. 3.3, we evaluate several models on the annotated PartiPrompts [65] dataset. Each prompt is annotated with one of two labels: “ret” or “gen”, indicating whether image retrieval or image generation produces a more appropriate image for the corresponding prompt. For example, the prompt “a portrait of a statue of the Egyptian god Anubis wearing aviator goggles, white t-shirt and leather jacket, flying over the city of Mars.” is labeled as “gen”, as there are (understandably) no appropriate images in the CC3M retrieval set, and generation produces a more relevant output. In contrast, “the geyser Old Faithful” is labeled as “ret,” as there are very relevant candidate images available for this prompt. We evaluate several models for making this decision on the validation set (Tab. 6), evaluating using F1 score given the class imbalance of the dataset (201 “gen”, 110 “ret” in the validation set labels):
|
| 310 |
+
|
| 311 |
+
1. Baselines: We measure the F1 score of several baseline methods, which provide a lower bound for how well data-driven approaches can do. We find that always retrieving an image, always generating an image, or simply deciding randomly (with a prior proportional to class frequencies) achieve F1 scores of 0.267, 0.389, and 0.451 respectively.
|
| 312 |
+
|
| 313 |
+
2. Heuristic: We also consider a simple heuristic which considers the maximum cosine similarity of the retrieval embedding against the entire image candidate set (i.e., the training set of CC3M). We run a grid search from 0 to 1 for possible threshold values. Whenever the maximum cosine similarity is above a threshold, we return “ret” and “gen” otherwise. This achieves an F1 of $0 . 2 6 1 - 0 . 5 5 9$ , depending on the threshold used (a threshold of 0.5 gives F1 of 0.261).
|
| 314 |
+
|
| 315 |
+
3. Linear classifier: Lastly, we train a linear classifier that takes as input the outputs of the LLM for the [IMG] tokens and the maximum cosine similarity. This classifier is trained with the binary cross-entropy loss over the training set of PartiPrompts annotations. This linear classifier achieves an F1 score of between 0.393 – 0.552, depending on the probability threshold used (a threshold of 0.5 gives an F1 score of 0.547).
|
| 316 |
+
|
| 317 |
+
We use the linear classifier in our final model, as it requires less hyperparameter tuning compared to the heuristic baseline, and performs comparably on quantitative metrics. During generation of qualitative samples (Fig. 7 and Fig. 5 in the main paper), we observed that the linear classifier generally performed well for many prompts, and decided correctly whether to retrieve or generate.
|
| 318 |
+
|
| 319 |
+
# D Qualitative Results
|
| 320 |
+
|
| 321 |
+
We present further qualitative samples in Fig. 7. We find that GILL is able to process complex text prompts more effectively than Stable Diffusion for many examples in PartiPrompts [65]. On VisDial [16] dialogue inputs, GILL is able to generate more relevant outputs (as measured against groundtruth images). We attribute these improved results to the stronger text representations of the LLM, and the effectiveness of our GILLMapper network.
|
| 322 |
+
|
| 323 |
+
Table 6: Results on PartiPrompts for classifying retrieval or generation.
|
| 324 |
+
|
| 325 |
+
<table><tr><td>Method</td><td>F1</td></tr><tr><td>Always retrieve</td><td>0.267</td></tr><tr><td>Always generate</td><td>0.389</td></tr><tr><td>Random</td><td>0.451</td></tr><tr><td>Heuristic</td><td>0.261- 0.559</td></tr><tr><td>Linear classifier</td><td>0.393 - 0.552</td></tr><tr><td>Human performance</td><td>0.851</td></tr></table>
|
| 326 |
+
|
| 327 |
+
Table 7: Results on image captioning on MS-COCO (2017) [34] and VQA [24]. For captioning, we report BLEU [41] and METEOR [5] scores. For VQA, we report the accuracy after applying the normalization described in the VQA repo (https://github.com/GT-Vision-Lab/VQA). indicates our reimplementation.
|
| 328 |
+
|
| 329 |
+
<table><tr><td rowspan="2">Model</td><td colspan="2">Captioning</td><td>VQA</td></tr><tr><td>BLEU-4</td><td>METEOR</td><td>O-shot</td></tr><tr><td>Frozen† [56]</td><td></td><td></td><td>0.2553</td></tr><tr><td>MAGMA [19]</td><td></td><td></td><td>0.2835</td></tr><tr><td>FROMAGe [31]</td><td>0.1023</td><td>0.2873</td><td>0.2851</td></tr><tr><td>Ours</td><td>0.1059</td><td>0.2529</td><td>0.3178</td></tr></table>
|
| 330 |
+
|
| 331 |
+
# E Other Evaluations
|
| 332 |
+
|
| 333 |
+
# E.1 Comparison to Prior Multimodal LMs
|
| 334 |
+
|
| 335 |
+
We ran evaluations on VQAv2 [24] and image captioning on MS-COCO [34]. The results are presented in Tab. 7. We found that GILL is comparable to models trained with similar compute and data. On VQAv2, we achieve a zero-shot val accuracy of 0.3178, which is slightly better than prior approaches of similar model sizes and compute: FROMAGe [31] achieves a zero-shot accuracy of 0.2851, Frozen [56] achieves 0.2553, and MAGMA [19] achieves 0.2835. For image captioning on the MS-COCO (2017) validation set, GILL achieves a BLEU $@ 4$ of 0.1059 and METEOR of 0.2529, which is comparable to FROMAGe (BLEU $@ 4$ of 0.1023 and METEOR of 0.2873). GILL is also capable of a wider set of tasks (e.g., generating interleaved image and text outputs) compared to these models.
|
| 336 |
+
|
| 337 |
+
We note that these scores are lower than SOTA models, as they are usually much larger and trained with significantly more compute and data (e.g., Flamingo [4] uses 23,040 TPU days, BLIP-2 [33] uses 144 GPU days, while ours uses 4 GPU days). Scaling up GILL to similar data and parameter scales to further push its capabilities is an exciting avenue for future work.
|
| 338 |
+
|
| 339 |
+
# E.2 Increasing Context on VisDial
|
| 340 |
+
|
| 341 |
+
GILL leverages an LLM backbone, which allows it to inherit some of the LLM’s capabilities, such as improved sensitivity to long input contexts. In the main paper, we showed that GILL can better condition on longer image and text inputs to generate more relevant images for VIST [28]. We run a similar experiment on Visual Dialogue [16], varying the number of dialogue rounds provided as input context to GILL and Stable Diffusion (SD) [49].
|
| 342 |
+
|
| 343 |
+
The results are presented in Fig. 8. We find that when longer text context is provided to both models, the performance of generating relevant images steadily improves. Interestingly, SD performance plateaus after 6 rounds of dialogue, while GILL continues to improve, outperforming SD when 7 or more rounds of dialogue are provided. These results showcase the improved sensitivity of our model to conditioning on long, dialogue-like text. Despite both approaches using the same image generation backbone, GILL is able to better make use of longer dialogue-text inputs (despite being only finetuned on image caption data).
|
| 344 |
+
|
| 345 |
+
|
| 346 |
+
|
| 347 |
+
# Comparison Against Stable Diffusion
|
| 348 |
+
|
| 349 |
+
|
| 350 |
+
Our model can process long, dialogue-like text inputs to generate more relevant images compared to non-LLM based text-to-image generation models.
|
| 351 |
+
|
| 352 |
+
# Visual Dialogue
|
| 353 |
+
|
| 354 |
+
|
| 355 |
+
Figure 7: Further qualitative samples from GILL. It is more sensitive to text inputs due to its LLM backbone, and better at processing complex text prompts.
|
| 356 |
+
|
| 357 |
+
|
| 358 |
+
Figure 8: Performance of our model and Stable Diffusion [49] with increasing context for generating VisDial [16] images. Our model is able to better process long dialoguelike text descriptions.
|
| 359 |
+
|
| 360 |
+
Table 8: Zero-shot FID [25] on the MS-COCO [34] (2014) validation set. 30,000 random samples are used to evaluate all models.
|
| 361 |
+
|
| 362 |
+
<table><tr><td>Model</td><td>FID (↓)</td></tr><tr><td>GLIDE [38]</td><td>12.24</td></tr><tr><td>Make-A-Scene [21] DALL-E 2 [44]</td><td>11.84</td></tr><tr><td>LAFITE2 [72]</td><td>10.39</td></tr><tr><td></td><td>8.42</td></tr><tr><td>Imagen [50]</td><td>7.27</td></tr><tr><td>Parti [65]</td><td>7.23</td></tr><tr><td>Re-Imagen [14]</td><td>6.88</td></tr><tr><td>SD [49] v1.5</td><td>9.22</td></tr><tr><td>GILL (ours)</td><td>12.2</td></tr></table>
|
| 363 |
+
|
| 364 |
+
# E.3 Image Generation
|
| 365 |
+
|
| 366 |
+
In addition to our evaluations on VIST [28] and VisDial [16], we also run evaluations on our model’s ability to generate images from MS-COCO [34] captions (Tab. 8). We generate images using 30,000 randomly sampled captions from the MS-COCO (2014) validation set, which is the standard evaluation of text-to-image generation models. We report zero-shot FID scores [25] of our model, Stable Diffusion [49] v1.5 (which we use as our backbone image generator), and several other approaches in Tab. 8. For our generation results and SD results, we use a classifier-free guidance scaling factor of 3.0 and 250 DDIM inference steps. On MS-COCO, our approach achieves a worse FID score than SD (9.22 to 12.2). This is likely because this task does not benefit as much from the LLM backbone, which has not been trained on as many image captions as SD (which exclusively trains on caption-like data). These numbers will likely improve further by finetuning GILL on even more text data (including image captions), which will allow our model to align more closely to the input space of the SD image generator.
|
| 367 |
+
|
| 368 |
+
# E.4 Inference Speed
|
| 369 |
+
|
| 370 |
+
One concern for deploying LLMs is the inference throughput and speed. We benchmark the inference performance of GILL on a single A6000 GPU. Generating text has the same throughput as a regular LM of the same size (i.e., that of OPT 6.7B). The main increase in inference time occurs when the model produces [IMG] tokens. For a batch size of 1, if the model decides to retrieve images, the additional inference overhead is minimal (less than 0.001s on average) as image retrieval is fast, requiring just a single matrix multiplication followed by a max operation. However, if GILL generates an [IMG] token, it takes 3.5s on average for the Stable Diffusion backbone to generate a corresponding image.
|
| 371 |
+
|
| 372 |
+
Overall, GILL’s inference speed is bottlenecked by the frequency of image generation, which is dependent on the application domain. In the case of generating dialogue-like text, we observed that images are usually generated or retrieved once or twice in a natural conversation. Amortized over a long conversation, it does not lead to a significant increase compared to a text-only LLM, though exact numbers would depend on the application.
|
| 373 |
+
|
| 374 |
+
# F Human Annotation on PartiPrompts
|
| 375 |
+
|
| 376 |
+
In Sec. 3.3 of the main paper, we described the process of annotating PartiPrompts [65] with perexample labels to retrieve or generate. The interface shown to human annotators is shown in Fig. 9. Annotators are tasked to determine which of two anonymized images are (1) more relevant to the provided prompt, and (2) more realistic. We randomize the order of the two images as well (i.e., the output of the retrieval model shows up $50 \%$ of the time as Image A).
|
| 377 |
+
|
| 378 |
+
|
| 379 |
+
Figure 9: User interface shown to human annotators for annotating PartiPrompts [65] examples.
|
| 380 |
+
|
| 381 |
+
We show each example to 5 independent human annotators. For determining whether to label a particular example as “ret” or “gen”, we take the majority vote of the 5 annotators on the image relevance question (“Is image A or image B more relevant to the above caption?”), and only keep the examples with an inter-annotator agreement of at least $4 / 5$ . This results in approximately 900 examples remaining (out of the 1,632 examples in PartiPrompts). Our annotations will be publicly released to facilitate future evaluations on this task.
|
| 382 |
+
|
| 383 |
+
We conducted evaluations on the Amazon Mechanical Turk platform with human annotators located in the US and Canada. Annotators were paid at an estimated hourly rate of 15 USD per hour. In total, we spent approximately 326 USD to collect these annotations.
|
parse/dev/Uczck6TlSZ/Uczck6TlSZ_content_list.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Uczck6TlSZ/Uczck6TlSZ_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/Uczck6TlSZ/Uczck6TlSZ_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/VD-AYtP0dve/VD-AYtP0dve_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/WqiZJGNkjn/WqiZJGNkjn.md
ADDED
|
@@ -0,0 +1,204 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MotionGPT: Human Motion as a Foreign Language
|
| 2 |
+
|
| 3 |
+
Biao Jiang1,2∗ Xin Chen2∗ Wen Liu2 Jingyi ${ \bf { Y } } { \bf { u } } ^ { 3 }$ Gang $\mathbf { Y } \mathbf { u } ^ { 2 }$ Tao Chen1† 1Fudan University 2Tencent 3ShanghaiTech University https://github.com/OpenMotionLab/MotionGPT
|
| 4 |
+
|
| 5 |
+
# Abstract
|
| 6 |
+
|
| 7 |
+
Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multimodal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models,motionlanguage pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this “motion vocabulary”, we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based questionand-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between.
|
| 8 |
+
|
| 9 |
+
# 1 Introduction
|
| 10 |
+
|
| 11 |
+
Recent years have witnessed a significant breakthrough in pre-trained large language models such as GPT [34, 35, 3, 27], BERT [7], and T5 [36, 5], which lead to the convergence of language [59, 47], image [33, 50, 20], mesh [55, 26] and mutlimodal [8] modeling. Nevertheless, a general pre-trained model for human motion and language has yet to emerge. This pre-trained motion-language model, capable of supporting numerous motion-relevant tasks through prompts, should benefit diverse fields like gaming, robotics, virtual assistant, and human behavior analysis.
|
| 12 |
+
|
| 13 |
+
Previous research on human motion has explored various tasks, including motion generation [29, 10, 46, 52, 57], motion captioning [9, 11], and motion prediction [56, 61, 24]. Recent text-to-motion works[46, 58, 30, 52] have attempted to employ pre-trained language-relevant models [7, 33]. For instance, MDM [46] learns a motion diffusion model with conditional text tokens from CLIP [33], while MLD [52] integrates motion latent space to improve the efficiency of motion diffusion process. On the other hand, MotionCLIP [45] and TM2T [11] concentrate on modeling the coupled relationship between motion and text description. However, the above approaches treat motion and language as separate modalities, which often require strictly paired motion and text data. Moreover, since the supervisions are task-specific, they can hardly generalize effectively to unseen tasks or data, as they lack a comprehensive understanding of the relationship between motion and language. We thus focus on building a pre-trained motion-language model, which can generalize to various tasks and learn in-depth motion-language correlation knowledge from more feasible motion and language data.
|
| 14 |
+
|
| 15 |
+
Two challenges are crucial and need to be solved for pre-training a promising motion-language model. The first is modeling the relation between language and motion, and the second is building a uniform multi-task framework that can generalize to new tasks. Fortunately, human motion exhibits a semantic coupling similar to human language, often interpreted as a form of body language. Building upon this observation,we follow vision-language pre-training from BEiT-3 [50] to treat human motion as a specific foreign language. By integrating motion and language data together and encoding them within a single vocabulary, the relationship between motion and language becomes more apparent. Therefore, with recent significantly larger-scale language data and models, the motion-language pretraining has great potential to improve the performance on motion tasks. Meanwhile, this pre-training on language enables textual instructions like prompts in InstructGPT [27] and makes the model more versatile and user-friendly for various motion tasks.
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
Figure 1: MotionGPT can address diverse motion-relevant tasks uniformly given different instructions. We provide the results on text-to-motion (the upper left), motion captioning (the bottom left), motion completion (the upper right), and the language question-to-answer (the bottom right). The left to right of motion represents the time order. Blue motion denotes the input, and yellow is the generation.
|
| 19 |
+
|
| 20 |
+
In this work, we propose a uniform motion-language framework, namely MotionGPT, that leverages the strong language generation and zero-shot transfer abilities of pre-trained language models for doing human motion-related tasks. To enable MotionGPT to comprehend and generate human-like motions, we first learn a motion-specific vector quantized variational autoencoder (VQ-VAE) model to construct “motion vocabulary”, akin to English vocabulary and then convert raw motion data into a sequence of motion tokens. These tokens are then processed by a pre-trained language model [36, 5] that learns the underlying grammar and syntax of the motion language, as well as its relationship with the corresponding textual descriptions. To effectively integrate language and motion in MotionGPT, we design a two-stage training scheme. We first pre-train the language model on the raw motion dataset to learn the basic grammar and syntax of the motion language. For prompt tuning, we fine-tune the language model on an instruction dataset, which contains both textual descriptions and motion data, to learn the correlation between the two modalities. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performance on text-to-motion, motion-to-text, motion prediction, and motion in-between.
|
| 21 |
+
|
| 22 |
+
We summarize our contributions as follows: (1) We propose a uniform motion-language generative pre-trained model, MotionGPT, which treats human motion as a foreign language, introduces natural language models into motion-relevant generation, and performs diverse motion tasks with a single model. (2) We introduce a motion-language training scheme with instruction tuning, to learn from task feedback and produce promising results through prompts. (3) We propose a general motion benchmark for multi-task evaluation, wherein MotionGPT achieves competitive performance across diverse tasks, including text-to-motion, motion-to-text, motion prediction, and motion in-between, with all available codes and data.
|
| 23 |
+
|
| 24 |
+
# 2 Related Work
|
| 25 |
+
|
| 26 |
+
Human Motion Synthesis involves generating diverse and realistic human-like motion using multimodal inputs, such as text [10, 30, 58, 46, 11, 1, 17], action [29, 12, 46, 52], and incomplete motion [56, 61, 24, 46]. Text-to-motion is one of the most important motion generation tasks, due to the userfriendly and convenient language input. MDM [46] proposes a diffusion-based generative model [14] separately trained on several motion tasks. MLD [52] advances the latent diffusion model [43, 38] to generate motions based on different conditional inputs. T2M-GPT [57] investigates a generative framework based on VQ-VAE and Generative Pre-trained Transformer (GPT) for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 61, 24] or motion in-between [46], which generates the intermediate motion while the first and last parts are fixed. Although they show promising results in various human motion tasks, most above methods are limited in using a single model to handle multiple tasks. We thus propose a uniform approach that treats human motion as a foreign language, and leverages the strong language generation and zero-shot transfer abilities of pre-trained language models
|
| 27 |
+
|
| 28 |
+
Table 1: Comparison of recent state-of-the-art methods on diverse motion-relevant tasks. Random Motion and Random Caption represent unconstrained generation of motions and motion descriptions.
|
| 29 |
+
|
| 30 |
+
<table><tr><td>Methods</td><td></td><td></td><td></td><td></td><td></td><td>Text-to-MotionMotion-to-TextMotion PredictionMotion In-betweenRandom MotionRandom Description</td></tr><tr><td>T2M-GPT[57]</td><td></td><td>xx<x</td><td></td><td></td><td></td><td></td></tr><tr><td>MLD [52]</td><td>：</td><td></td><td></td><td></td><td></td><td>xxx></td></tr><tr><td>TM2T[11]</td><td></td><td></td><td><xx</td><td>xxx</td><td>><x</td><td></td></tr><tr><td>MDM [46]</td><td></td><td></td><td></td><td></td><td>√</td><td>X</td></tr><tr><td>MotionDiffuse[58]</td><td></td><td>X</td><td></td><td></td><td><</td><td>X</td></tr><tr><td>MotionGPT (Ours)</td><td>√</td><td>√</td><td>√</td><td>√</td><td>√</td><td>√</td></tr></table>
|
| 31 |
+
|
| 32 |
+
Human Motion Captioning. To describe human motion with natural languages, [44] learns the mapping from motions to language relying on two statistical models. Furthermore, recurrent networks have also been used in [54, 32]. More recently, TM2T [11] proposed a new motion representation that compresses motions into a short sequence of discrete variables, then uses a neural translation network to build mappings between two modalities. While previous research like TM2T [11] incorporated captioning modules into their training pipeline for motion generation, these approaches are constrained to bidirectional translation between text and motion within one uniform framework.
|
| 33 |
+
|
| 34 |
+
Language Models and Multi-Modal. Large-scale language models (LLMs) [7, 6, 36, 3, 59, 47], enabled by extensive datasets and model size, have demonstrated impressive comprehension and generation capabilities, elevating natural language processing to new heights. BERT [7] pre-trains deep bidirectional language representations that can support downstream tasks. T5 [36] introduced a unified framework that converts all text-based language problems into a text-to-text format. More recent research [51, 2, 27, 5] find that by fine-tuning pre-trained models using input-output pairs consisting of instructions and coupled answers, the performance of pre-trained models can be further improved. FLAN [5] presents an instruction-tuning technique that surpasses the performance of non-tuned models in unseen tasks. Recently, the wave of multi-modal models [20, 15, 19] is intriguing to process text along with other modalities, such as images [20, 15, 8], audio [13, 8], and videos [53]. CLIP [33] further learns a semantic latent representation that couples images with corresponding language descriptions. Despite the success of language models in various vision-language tasks, the development of multi-modal language models that can handle human motion is still limited.
|
| 35 |
+
|
| 36 |
+
Motion Language Pre-training. Existing text-to-motion generation methods [10, 30, 46, 11, 1, 17] can be characterized as caption-to-motion, where the models take in a pure text description of the desired motion. While these methods can generate motions from textual descriptions, they are often limited in supporting instructions from users like InstructGPT [27]. In other words, they do not allow users to provide context-specific instructions for certain applications. MotionCLIP [45] utilizes the language and visual understanding of CLIP [33] to align its latent space with a motion auto-encoder. Meanwhile, many language models, such as T5[36] and InstructGPT [27], have been developed to address diverse language processing tasks, including translation, question answering, and classification. These models are typically designed to map a given text input to a target output, such as a translation or answer. However, while these models have shown remarkable performance in language tasks, they have not been widely applied to motion tasks. Therefore, we propose MotionGPT to enable the effective integration of natural language models with human motion tasks, providing a unified solution for motion synthesis problems.
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
Figure 2: Method overview: MotionGPT consists of a motion tokenizer $\nu$ (Sec. 3.1) and a motionaware language model (Sec. 3.2). Combining Motion Tokens learned by $\nu$ and Text Tokens by text tokenizer, we then learn motion and language jointly utilizing language model as backbone.
|
| 40 |
+
|
| 41 |
+
# 3 Method
|
| 42 |
+
|
| 43 |
+
To involve large language data and models in the motion generation tasks, we propose a unified motion-language framework named MotionGPT. As illustrated in Fig. 2, MotionGPT consists of a motion tokenizer responsible for converting raw motion data into discrete motion tokens (Sec. 3.1), as well as a motion-aware language model that learns to understand the motion tokens from large language pre-training models by corresponding textual descriptions (Sec. 3.2). To address motionrelevant tasks, we introduce a three-stage training scheme (Sec. 3.3) of MotionGPT for the training of motion tokenizer, motion-language pre-training, and instruction tuning.
|
| 44 |
+
|
| 45 |
+
We first propose the motion tokenizer consisting of a motion encoder $\mathcal { E }$ and a motion decoder $\mathcal { D }$ , to encode a $M$ frame motion $m ^ { 1 : M } = \{ x ^ { i } \} _ { i = 1 } ^ { M }$ into $L$ motion tokens $z ^ { 1 : L } = \{ z ^ { i } \} _ { i = 1 } ^ { L } , L = M / l$ , and decode $z ^ { 1 : L }$ back into the motion $\hat { m } ^ { 1 : M } = \bar { \mathcal { D } } ( z ^ { 1 : L } ) = \mathcal { D } ( \mathcal { E } ( m ^ { 1 : M } ) )$ , where $l$ denotes the temporal downsampling rate on motion length. Then, given an $N$ length sentence $\boldsymbol { w ^ { 1 : N } } = \{ w ^ { i } \} _ { i = 1 } ^ { N }$ describing a motion-related question or demand, MotionGPT aims to generate its answer as $L$ length tokens $\hat { x } ^ { 1 : L } = \{ \hat { x } ^ { i } \} _ { i = 1 } ^ { L }$ . It could be the human motion tokens $\hat { x } _ { m } ^ { 1 : L }$ or the text tokens $\hat { x } _ { t } ^ { 1 : L }$ , which results in a motion $\hat { m } ^ { 1 : M }$ or a sentence $\hat { w } ^ { 1 : L }$ like a description of the given motion.
|
| 46 |
+
|
| 47 |
+
# 3.1 Motion Tokenizer
|
| 48 |
+
|
| 49 |
+
To represent motion in discrete tokens, we pre-train a 3D human motion tokenizer $\nu$ based on the Vector Quantized Variational Autoencoders (VQ-VAE) architecture used in [48, 42, 11, 57]. Our motion tokenizer consists of an encoder $\mathcal { E }$ and a decoder $\mathcal { D }$ . The encoder generates discrete motion tokens with high informative density, while the decoder is able to reconstruct the motion tokens into motion sequences $\hat { m } ^ { 1 : M }$ . This approach enables us to efficiently represent motion as a language, facilitating the integration of motion and language for various motion-related tasks.
|
| 50 |
+
|
| 51 |
+
Specifically, the motion encoder $\mathcal { E }$ first applies 1D convolutions to given frame-wise motion features $\stackrel { \bullet } { m } ^ { 1 : M }$ along the time dimension, to obtain latent vectors $\hat { z } ^ { 1 : L } = \mathcal { E } ( m ^ { 1 : M } )$ . Next, we transform $\hat { z }$ of codebooconsists of ntries laten $z$ t rough discrete quantization. Tbedding vectors, each of dimen len able codebookThe process of $Z = \{ z ^ { i } \} _ { i = 1 } ^ { K } \subset \mathbb { R } ^ { d }$ $K$ $d$ $Q ( \cdot )$ $b$ $b _ { k }$ $Z$
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
\begin{array} { r } { z _ { i } = Q ( \hat { z } ^ { i } ) : = \arg \operatorname* { m i n } _ { z _ { k } \in Z } \left\| \hat { z } _ { i } - z _ { k } \right\| _ { 2 } . } \end{array}
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
After quantization, the motion decoder $D$ project $z ^ { 1 : L } = \{ z ^ { i } \} _ { i = 1 } ^ { L }$ back to raw motion space as the motion $\hat { m } ^ { 1 : M }$ with $M$ frames. To train this motion tokenizer, we follow [11, 57] to utilize three distinct loss functions for training and optimizing the motion tokenizer: $\mathcal { L } _ { \mathcal { V } } = \mathcal { L } _ { r } + \mathcal { L } _ { e } + \mathcal { L } _ { c }$ , where the reconstruction loss $\mathcal { L } _ { r }$ , the embedding loss $\mathcal { L } _ { e }$ , and the commitment loss $\mathcal { L } _ { c }$ . To further improve the generated motion quality, we follow [57] to utilize L1 smooth loss and velocity regularization in the reconstruction loss, as well as exponential moving average (EMA) and codebook reset techniques [37] to enhance codebook utilization during training. We provide more details about the architecture and the training of our motion tokenizer in the supplement.
|
| 58 |
+
|
| 59 |
+
# 3.2 Motion-aware Language Model
|
| 60 |
+
|
| 61 |
+
Employing this motion tokenizer, a human motion $m ^ { 1 : M }$ can be mapped to a sequence of motion tokens $\tilde { z } ^ { 1 : \overline { { L } } }$ , allowing for joint representation with similar vocabulary embedding in language models [18, 36, 27]. By combining them in the unified vocabulary, we then learn motion and language jointly. We first represent motion tokens $z ^ { 1 : L }$ as a sequence of indices $s ^ { 1 : L } = \{ s ^ { i } \} _ { i = 1 } ^ { L }$ , where $\bar { s } ^ { i }$ corresponds to the index number of motion tokens $z ^ { 1 : L }$ . On the other hand, previous language models, such as T5 [36], encode text as WordPiece tokens. They utilized a vocabulary of $K _ { t }$ word pieces and trained the SentencePiece [18] model on a mixture of language datasets.
|
| 62 |
+
|
| 63 |
+
Most previous text-to-motion [11, 52, 57] or motion-to-text [11] approaches employ different modules the same way. To achieve this, we combine the original text vocabulary to handle text and motion individually, while we aim to model text and human motion together and in $V _ { t } = \{ v _ { t } ^ { i } \} _ { i = 1 } ^ { K _ { t } }$ with motion vocabulary $V _ { m } = \{ v _ { m } ^ { i } \} _ { i = 1 } ^ { K _ { m } }$ , which is order-preserving to our motion codebook $Z$ . Moreover, $V _ { m }$ includes several special tokens like boundary indicators, for example, $\cdot$ as the start and end of the motion. Thus, we employ a new unified text-motion vocabulary $V = \{ V _ { t } , V _ { m } \}$ , and can formulate diverse motion-related tasks in a general format, where both input "words" and output "words" are from the same $V$ . These "words" can represent natural language, human motion, or even a mixture of two, depending on the specific task to be solved. Therefore, our MotionGPT allows for the flexible representation and generation of diverse motion-related outputs within a single model.
|
| 64 |
+
|
| 65 |
+
To address the conditioned generation task, we employ a transformer-based model based on the architecture proposed in [36], which effectively maps the input sequences to the output. Our source input consists of a sequence of tokens $\ = X _ { s } \ \dot { = } \ \{ \dot { x _ { s } } ^ { i } \} _ { i = 1 } ^ { N }$ , where $x _ { s } ~ \in ~ V$ and $N$ represents the input length. Similarly, the target output is $X _ { t } ~ = ~ \{ x _ { t } ^ { ~ i } \} _ { i = 1 } ^ { L }$ , where $x _ { t } ~ \in ~ V$ and $L$ denotes the output length. As shown in Fig. 2, the source tokens are fed into the transformer encoder, and the subsequent decoder predicts the probability distribution of the potential next token at each step $\begin{array} { r } { p _ { \theta } ( x _ { t } \mid \hat { x _ { s } } ) = \prod _ { i } p _ { \theta } \left( \hat { x } _ { t } ^ { i } \mid x _ { t } ^ { < i } , x _ { s } \right) } \end{array}$ in an autoregressive manner. Therefore, during the training process, the objective is to maximize the log-likelihood of the data distribution:
|
| 66 |
+
|
| 67 |
+
$$
|
| 68 |
+
\mathcal { L } _ { L M } = - \sum _ { i = 0 } ^ { L _ { t } - 1 } \log p _ { \theta } \left( x _ { t } ^ { i } \mid x _ { t } ^ { < i } , x _ { s } \right) .
|
| 69 |
+
$$
|
| 70 |
+
|
| 71 |
+
By optimizing this objective, MotionGPT learns to capture the underlying patterns and relationships from the data distribution, facilitating the accurate and meaningful generation of the target "words". During the inference process, the target tokens are sampled recursively from the predicted distribution $p _ { \theta } \left( \hat { x _ { t } _ { t } } ^ { i } \mid \hat { x _ { t } } ^ { < i } , x _ { s } \right)$ until the end token (i.e., ${ < } / { \mathrm { s } } > ,$ ). This sampling strategy enables the generation of the target sequence in a step-by-step manner, where each token is probabilistically determined based on the previously generated tokens and the given source input.
|
| 72 |
+
|
| 73 |
+
# 3.3 Training Strategy
|
| 74 |
+
|
| 75 |
+
Since T5s have only been exposed to language data, represented within a text vocabulary $V _ { t }$ , we thus bridge motion and language and enable this language model to comprehend human motion concepts, by learning the motion vocabulary $V _ { m }$ . As shown in Fig. 3, our training scheme includes three stages: (1) Training of motion tokenizer, which focuses on learning the motion codebook to represent human motion as discrete tokens. (2) Motion-language pre-training stage, which includes unsupervised and supervised objectives to learn the relationship between motion and language. (3) Instruction tuning stage, which tunes the model based on prompt-based instructions for different motion-relevant tasks.
|
| 76 |
+
|
| 77 |
+
Training of Motion Tokenizer. We first learn the motion tokenizer using the objective defined in Equation 3.1. This training process allows any human motion sequence $\hat { x } ^ { \top : L }$ to be represented as a sequence of motion tokens, enabling seamless integration with textual information. Once optimized, the motion tokenizer remains unchanged throughout the subsequent stages of the pipeline.
|
| 78 |
+
|
| 79 |
+
Motion-language Pre-training Stage. The T5 models [36, 5] are trained and fine-tuned on natural language datasets with instruction-based phrasing [5, 27]. We continue to pre-train this model using a mixture of language and motions data in both unsupervised and supervised manners: 1) To generalize to various downstream tasks like [7, 35, 36, 27], we follow [36] to design an objective, where a certain percentage $( 1 5 \% )$ of tokens in the input tokens $X _ { s }$ are randomly replaced with a special sentinel token. On the other side, the corresponding target sequence is constructed by extracting the dropped-out spans of tokens, delimited by the same sentinel tokens used in the input sequence, along with an additional sentinel token to indicate the end of the target sequence. 2) We then learn the motion-language relation by the supervision of paired text-motion datasets [10, 31]. We train MotionGPT on the supervised motion-language translation, where the input is either a human motion or a text description.After unsupervised and supervised training processes, we aim to equip our model with the understanding of text and motion relationships.
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
Figure 3: Training Scheme. We introduce three training steps for our MotionGPT (Sec. 3.3): First $\nu$ learn a codebook for discrete motion representation. Then we train language using a mixture of language and motion data to learn the semantic coupling between text and motion. Finally, we fine-tune the model in a multi-task text-motion dataset with instructions.
|
| 83 |
+
|
| 84 |
+
Instruction Tuning Stage. We construct a multi-task text-motion dataset by formulating it as instructions, building upon the foundation of existing text-to-motion datasets such as HumanML3D [10] and KIT [31]. Specifically, we define 15 core motion tasks, such as motion generation with text, motion captioning, motion prediction, and others. For each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. For example, an instruction prompt for motion generation task could be “Can you generate a motion sequence that depicts ‘a person emulates the motions of a waltz dance’?”. Similarly, for the motion captioning task, the instruction prompt could be “Provide an accurate caption describing the motion of <motion_tokens>”, where <motion_tokens> represents a sequence of motion tokens generated by our motion tokenizer. We have demonstrated the efficacy of instruction tuning in Sec. 4.3, which leads to improvement across various tasks and enhances the model performance for unseen tasks or prompts. More examples of prompts are provided in the supplements.
|
| 85 |
+
|
| 86 |
+
# 4 Experiments
|
| 87 |
+
|
| 88 |
+
Extensive comparisons evaluate the performance of our MotionGPTs across multiple motion-relevant tasks and datasets. Details of the dataset settings, evaluation metrics, and implementation specifics (Sec. 4.1) are provided. We first present a uniform benchmark by comparing our approach with other SOTAs across various tasks (Sec. 4.2). Then, we evaluate each specific comparison on text-to-motion (Sec. 4.2), motion-to-text (Sec. 4.2), motion prediction and motion in-between (Sec. 4.2). The supplements include more qualitative results, user studies, and further implementation details.
|
| 89 |
+
|
| 90 |
+
# 4.1 Experimental Setup
|
| 91 |
+
|
| 92 |
+
Datasets. General motion synthesis can support diverse task settings, and thus previous datasets and a modified benchmark are utilized to evaluate MotionGPT. The study primarily focuses on two text-to-motion datasets: HumanML3D [10] and KIT [31]. The KIT dataset provides 6,353 textual descriptions corresponding to 3,911 motion sequences, while the HumanML3D dataset [10] is a more recent dataset that contains 14,616 motion sequences obtained from AMASS [25], along with 44,970 sequence-level textual descriptions. To evaluate MotionGPT as a uniform framework on tasks, such as motion prediction and motion completion (in-between), we utilize the motion sequences available in HumanML3D, which is also a subset of the larger AMASS dataset. Following the previous works [10, 52, 46], we adopt the same motion representation for fair comparisons, which combines joint velocities, positions, and rotations. By using this consistent representation, MotionGPT enables the availability to support further studies in the field. (cf. supplement for the benchmark details.)
|
| 93 |
+
|
| 94 |
+
Table 2: Comparison of four motion-related tasks on HumanML3D [10] dataset. The evaluation metrics are computed using the encoder introduced in [10]. The empty columns of previous methods indicate that they can not handle the task. The arrows $( )$ indicate that closer to Real is desirable. Bold and underline indicate the best and the second best result on text-to-motion task.
|
| 95 |
+
|
| 96 |
+
<table><tr><td rowspan="2">Methods</td><td colspan="3">Text-to-Motion</td><td colspan="3">Motion-to-Text</td><td colspan="2">Motion Prediction</td><td colspan="2">Motion In-between</td></tr><tr><td>R TOP1↑</td><td>FID↓</td><td>DIV→</td><td>R TOP3↑</td><td>Bleu@4↑</td><td>Cider↑</td><td>FID↓</td><td>DIV→</td><td>FID↓</td><td>DIV→</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.002±.000</td><td>9.503±.065</td><td>0.828</td><td>-</td><td>-</td><td>0.002</td><td>9.503</td><td>0.002</td><td>9.503</td></tr><tr><td>MLD [52]</td><td>0.481±.003</td><td>0.473±.013</td><td>9.724±.082</td><td>1</td><td></td><td></td><td></td><td>=</td><td>-</td><td></td></tr><tr><td>T2M-GPT[57]</td><td>0.491±.003</td><td>0.116±.004</td><td>9.761±.081</td><td>-</td><td></td><td>-</td><td>2.056</td><td>8.635</td><td>-</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.424±.017</td><td>1.501±.003</td><td>8.589±.076</td><td>0.823</td><td>7.00</td><td>16.8</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MDM [46]</td><td>0.320±005</td><td>0.544±.044</td><td>9.559±.086</td><td>-</td><td>-</td><td>-</td><td>6.031</td><td>7.813</td><td>2.698</td><td>8.420</td></tr><tr><td>MotionGPT(Ours)</td><td>0.492±.003</td><td>0.232±.008</td><td>9.528±.071</td><td>0.827</td><td>12.47</td><td>29.2</td><td>0.905</td><td>8.972</td><td>0.214</td><td>9.560</td></tr></table>
|
| 97 |
+
|
| 98 |
+
Evaluation Metrics are summarized as four parts. (1) Motion quality: Frechet Inception Distance (FID) is our primary metric based on a feature extractor [10] to evaluate the distance of feature distributions between the generated and real motions. For motion completion, we utilize metrics used in motion prediction studies [56, 61, 24], such as Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the accuracy of the predicted motion. (2) Generation diversity: We utilize the Diversity (DIV) metric to assess the motions diversity, which calculates the variance through features extracted from the motions [10]. MultiModality (MM) measures the diversity of generated motions within the same text description of motion. (3) Text matching: Based on the feature space from [10], the motion-retrieval precision (R Precision) evaluates the accuracy of matching between texts and motions using Top 1/2/3 retrieval accuracy. Multi-modal Distance (MM Dist) measures the distance between motions and texts. (4) Linguistic quality: We follow 1] utilizing linguistic metrics from natural language studies, including BLUE [28], Rouge [23], Cider [49], and BertScore [60] to evaluate the quality of generated motion captions.
|
| 99 |
+
|
| 100 |
+
Implementation Details. We set the codebook of motion tokenizer as $K \in \mathbb { R } ^ { 5 1 2 \times 5 1 2 }$ for most comparisons. The motion encoder $\mathcal { E }$ incorporates a temporal downsampling rate $l$ of 4. We utilize T5 [36] as the underlying architecture for our language model, with a baseline model consisting of 12 layers in both the transformer encoder and decoder. The feed-forward networks have an output dimensionality of $d _ { \mathrm { f f } } = 3 0 7 2$ , and the attention mechanisms employ an inner dimensionality of $d _ { \mathrm { k v } } = 6 4$ . The remaining sub-layers and embeddings have a dimensionality of $d _ { \mathrm { m o d e l } } = 7 6 8$ Moreover, all our models employ the AdamW optimizer for training. The motion tokenizers are trained utilizing a $1 0 ^ { - 4 }$ learning rate and a 256 mini-batch size, while our language models have a $2 \times 1 0 ^ { - 4 }$ learning rate for the pre-train stage, $1 0 ^ { - 4 }$ for the instruction tuning stage, and a 16 mini-batch size for both stages. The motion tokenizer undergoes 150K iterations of training, while the language model undergoes 300K iterations during the pre-train stage and another 300K iterations during the instruction tuning stage. Small and Base models are trained on 8 Tesla V100 GPUs while Large models are trianed on 64 Tesla V100 GPUs.
|
| 101 |
+
|
| 102 |
+
# 4.2 Comparisons on Motion-relevant Tasks
|
| 103 |
+
|
| 104 |
+
Comparisons on Multiple Tasks. By introducing a uniform framework that treats human motion as a foreign language, we open up the exploration of diverse motion-relevant tasks. We employ a 220M pre-trained Flan-T5-Base[36, 5] model as our backbone and fine-tune the model through the pre-training and instruction tuning stage (Sec. 3.3) for all following comparisons. As shown in Tab. 2, we evaluate MotionGPT against state-of-the-art methods on key tasks such as text-conditioned motion generation [52, 57, 11, 46], motion captioning [11], motion prediction [46], and motion in-between[46]. While we leverage existing results from previous works or benchmarks for textto-motion and motion-to-text tasks, we re-implement the motion diffusion models [46] for motion prediction and evaluate it under the same metrics and settings. Please note that some methods are designed for specific tasks, and thus some metrics are empty for tasks they cannot handle. The results presented in Tab. 2 demonstrate that our MotionGPT achieves competitive performance across all evaluated tasks, highlighting its capability to address diverse motion tasks within a single model.
|
| 105 |
+
|
| 106 |
+
Table 3: Comparison of text-to-motion on HumanML3D [10]. The empty MModality indicates Real motion is deterministic. These methods are sorted by FID. $P r e$ -trained and Fine-tuned indicate uniform motion-language pre-training and specific fine-tuning on this task. $( c f$ . Tab. 2 for notations.)
|
| 107 |
+
|
| 108 |
+
<table><tr><td rowspan="2">Methods</td><td colspan="3">RPrecision↑</td><td rowspan="2">FID↓</td><td rowspan="2">MMDist↓</td><td rowspan="2">Diversity→</td><td rowspan="2">MModality↑</td></tr><tr><td>Top1</td><td>Top2</td><td>Top3</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.424±.003</td><td>0.618±.003</td><td>0.729±.002</td><td>1.501±.017</td><td>3.467±.011</td><td>8.589±.076</td><td>2.424±.093</td></tr><tr><td>T2M[10]</td><td>0.457±.002</td><td>0.639±.003</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MotionDiffuse [58]</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>MDM [46]</td><td>0.320±.005</td><td>0.498±.004</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MLD [52]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>T2M-GPT [57]</td><td>0.491±.003</td><td>0.680±.003</td><td>0.775±.002</td><td>0.116±.004</td><td>3.118±.011</td><td>9.761±.081</td><td>1.856±.011</td></tr><tr><td>MotionGPT (Pre-trained)</td><td>0.435±.003</td><td>0.607±.002</td><td>0.700±.002</td><td>0.160±.008</td><td>3.700±.009</td><td>9.411±.081</td><td>3.437±.091</td></tr><tr><td>MotionGPT (Fine-tuned)</td><td>0.492±.003</td><td>0.681±.003</td><td>0.778±.002</td><td>0.232±.008</td><td>3.096±.008</td><td>9.528±.071</td><td>2.008±.084</td></tr></table>
|
| 109 |
+
|
| 110 |
+
Table 4: Comparison of motion captioning on HumanML3D [10]. The evaluation metrics follow [11], while we use the ground truth texts without pre-processing for linguistic metrics calculation.
|
| 111 |
+
|
| 112 |
+
<table><tr><td rowspan="2">Methods</td><td colspan="2">RPrecision↑</td><td rowspan="2">MMDist↓</td><td rowspan="2">Lengthavg↑</td><td rowspan="2">Bleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑</td><td rowspan="2"></td><td rowspan="2"></td><td rowspan="2"></td><td rowspan="2"></td></tr><tr><td>Top1</td><td>Top3</td></tr><tr><td>Real</td><td>0.523</td><td>0.828</td><td>2.901</td><td>12.75</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.516</td><td>0.823</td><td>2.935</td><td>10.67</td><td>48.9</td><td>7.00</td><td>38.1</td><td>16.8</td><td>32.2</td></tr><tr><td>MotionGPT (Ours)</td><td>0.543</td><td>0.827</td><td>2.821</td><td>13.04</td><td>48.2</td><td>12.47</td><td>37.4</td><td>29.2</td><td>32.4</td></tr></table>
|
| 113 |
+
|
| 114 |
+
Comparisons on Text-to-Motion. The text-to-motion task involves generating human motion sequences based on a given text input. We evaluate the proposed the MotionGPT model as the pre-trained MotionGPT, the same one in Tab. 2, as well as fine-tuned it on text-to-motion task. We compare our MotionGPTs with other SOTAs [11, 10, 46, 52, 57] and evaluate the performance on both HumanML3D and KIT datasets using suggested metrics [10]. The results are computed with a $9 5 \%$ confidence interval, obtained from 20 repeated runs. The majority of the reported results are taken directly from their own papers or the benchmark presented in [10]. Tab. 3 summarizes the comparison results, where MotionGPT achieves competitive performance on most metrics.
|
| 115 |
+
|
| 116 |
+
Comparisons on Motion-to-Text. The motion-to-text task involves generating a text description based on a given human motion sequence. We compare the pre-trained MotionGPT with recent work TM2T [11]. We evaluate the performance on the HumanML3D using the suggested metrics from [11]. Additionally, we measure the average numbers of words Length $\mathbf { a v } \mathbf { g }$ for further comparisons. Please note that the reported results in [11] are evaluated with pre-processed ground truth text, which ignores the grammatical tense and plural forms of words. In Tab. 4, we directly use the ground truth text descriptions for a more accurate assessment. This comparison shows that MotionGPT overperforms recent work on text descriptions of given motions.
|
| 117 |
+
|
| 118 |
+
Comparisons on Motion Prediction and In-between. We summarize motion prediction and in-between together as general motion completion. To evaluate the motion completion capability of MotionGPT, we employ part of the AMASS dataset [25], a motion-only dataset. For motion prediction task, we only input around the first $20 \%$ of the motion sequence as conditions. For in-between, we mask about $50 \%$ motion randomly for completion. We also fine-tune MotionGPT specifically for this task and employ FID, ADE, and FDE as metrics like Sec. 4.1. Furthermore, we evaluate MDM [46] on motion prediction by utilizing their provided model, which also supports motion in-between through masked motion “in-painting”. The real motion data is used as one of our baselines. Tab. 5 reports that our MotionGPT has the best motion completion quality and diversity.
|
| 119 |
+
|
| 120 |
+
Table 5: Comparison of motion prediction and motion in-between on part of AMASSS [25] dataset using motion data only. FID indicates motion quality and Diversity (DIV) for motion diversity within each condition. ADE and FDE are joints distance between generation and ground truth.
|
| 121 |
+
|
| 122 |
+
<table><tr><td rowspan="3">Methods</td><td colspan="4">Motion Prediction</td><td colspan="3">Motion In-between</td></tr><tr><td>FID↓</td><td>Diversity↑</td><td>ADE↓</td><td>FDE↓</td><td>FID↓</td><td>Diversity↑</td><td>ADE↓</td></tr><tr><td>Real</td><td>0.002</td><td>9.503</td><td>-</td><td>-</td><td>0.002</td><td>9.503</td><td>-</td></tr><tr><td>MDM[46]</td><td>6.031</td><td>7.813</td><td>5.446</td><td>8.561</td><td>2.698</td><td>8.420</td><td>3.787</td></tr><tr><td>T2M-GPT[57]</td><td>2.056</td><td>8.635</td><td>6.161</td><td>8.302</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MotionGPT (Ours)</td><td>0.905</td><td>8.972</td><td>4.745</td><td>6.040</td><td>0.214</td><td>9.560</td><td>3.762</td></tr></table>
|
| 123 |
+
|
| 124 |
+
<table><tr><td rowspan="2">Size</td><td rowspan="2">Instruction Tuning</td><td colspan="3">Text-to-Motion</td><td colspan="3">Motion-to-Text</td><td colspan="2">Motion Prediction</td><td colspan="2">Motion In-between</td></tr><tr><td>R TOP3↑</td><td>FID←</td><td>DIV→</td><td>MMDist↓</td><td>Bleu@4↑</td><td>Cider↑</td><td>FID↓</td><td>DIV→</td><td>FID↓</td><td>DIV→</td></tr><tr><td>Real</td><td>-</td><td>0.797</td><td>0.002</td><td>9.503</td><td>2.901</td><td>-</td><td>1</td><td>0.002</td><td>9.503</td><td>0.002</td><td>9.503</td></tr><tr><td>Small</td><td></td><td>0.706</td><td>0.727</td><td>9.264</td><td>2.748</td><td>12.02</td><td>24.9</td><td>-</td><td>-</td><td>1</td><td>-</td></tr><tr><td>Small</td><td>√</td><td>0.663</td><td>0.336</td><td>9.239</td><td>2.931</td><td>10.54</td><td>24.3</td><td>0.954</td><td>8.727</td><td>0.326</td><td>9.618</td></tr><tr><td>Base</td><td></td><td>0.722</td><td>0.365</td><td>9.407</td><td>2.821</td><td>12.47</td><td>29.2</td><td></td><td>-</td><td></td><td></td></tr><tr><td>Base</td><td>√</td><td>0.700</td><td>0.160</td><td>9.411</td><td>3.019</td><td>11.42</td><td>28.2</td><td>0.905</td><td>8.972</td><td>0.214</td><td>9.560</td></tr><tr><td>Large</td><td></td><td>0.694</td><td>0.234</td><td>9.310</td><td>2.776</td><td>12.44</td><td>28.5</td><td>1</td><td>-</td><td>1</td><td>1</td></tr><tr><td>Large</td><td>√</td><td>0.708</td><td>0.159</td><td>9.301</td><td>3.011</td><td>11.71</td><td>29.1</td><td>0.556</td><td>8.975</td><td>0.223</td><td>9.358</td></tr></table>
|
| 125 |
+
|
| 126 |
+
Table 6: Evaluation of instruction tuning and different model sizes of MotionGPTs in four motion tasks on HumanML3D [10] dataset. $( c f$ . Tab. 2 for metrics details)
|
| 127 |
+
|
| 128 |
+
# 4.3 Ablation Studies
|
| 129 |
+
|
| 130 |
+
MotionGPT employs T5 [36] as the motion-aware language backbone model, and we train these models with pre-training and then instruction tuning. Thus, both model size and training strategy influence the performance of MotionGPTs. We here evaluate them on the typical motion tasks. More detailed ablation studies are provided in the supplements.
|
| 131 |
+
|
| 132 |
+
Model Sizes. We evaluate the performance of models with different sizes across four motion tasks. Besides the base 220M MotionGPT in Sec. 4.1, we now evaluate 60M, 220M, and 770M MotionGPTs. Tab. 6 demonstrates that the 220M base model has achieved remarkable performance compared to the smaller 60M model. However, the larger model size of current Motions does not yield significant improvements and, in few cases, even leads to worse results, as observed in the motion in-between task. We believe this could be caused by the small amount of current motion datasets. HumanML3D only includes $1 5 \mathrm { k }$ motion sequences, much smaller than even billions of language and image data.
|
| 133 |
+
|
| 134 |
+
Effectiveness of Instruction Tuning. We evaluate the impact of our instruction tuning strategy on different model sizes. The results in Tab. 6 demonstrate that instruction tuning enhances the versatility of MotionGPT, enabling more motion tasks like motion completion and improving the motion performance of the text-to-motion task. However, for pure text-generation tasks, the model performance is downgraded, likely due to the pair amount of textual descriptions and coupled motions.
|
| 135 |
+
|
| 136 |
+
# 5 Disscusion
|
| 137 |
+
|
| 138 |
+
As the first trial, to our best knowledge, exploring human motion generation using language models, the proposed MotionGPT still owns limitations as follows. MotionGPT only utilizes motion on articulated human bodies, while many other works focus on faces [16, 4], hands [39, 22, 21] and even animal [40, 62] motion. Besides, our method is also restricted to multiple humans without modeling human-object, or human-environment interactions [41]. It is interesting to model the human interaction scenarios in a motion-language framework and generate controllable motions [41].
|
| 139 |
+
|
| 140 |
+
We summarize the proposed MotionGPT as a uniform motion-language framework to generate plausible human motion and natural language descriptions through prompt-based instructions. Compared to the compatible motion diffusion methods [52, 46], our MotionGPT produces competitive results on motion generation, motion captioning, motion prediction, and motion in-between using only one pre-trained generative model. With the advancement of large language data and models [36, 5], MotionGPT is also capable of addressing natural question-to-answer tasks. Extensive experiments on various human motion-relevant tasks demonstrate the effectiveness and extendibility of MotionGPT.
|
| 141 |
+
|
| 142 |
+
References
|
| 143 |
+
[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019.
|
| 144 |
+
[2] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Promptsource: An integrated development environment and repository for natural language prompts. arXiv preprint arXiv:2202.01279, 2022.
|
| 145 |
+
[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.
|
| 146 |
+
[4] Xuan Cao, Zhang Chen, Anpei Chen, Xin Chen, Shiying Li, and Jingyi Yu. Sparse photometric 3d face reconstruction guided by morphable models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4635–4644, 2018.
|
| 147 |
+
[5] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.
|
| 148 |
+
[6] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019.
|
| 149 |
+
[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
|
| 150 |
+
[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665, 2023.
|
| 151 |
+
[9] Yusuke Goutsu and Tetsunari Inamura. Linguistic descriptions of human motion with generative adversarial seq2seq learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4281–4287. IEEE, 2021.
|
| 152 |
+
[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152–5161, June 2022.
|
| 153 |
+
[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022.
|
| 154 |
+
[12] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020.
|
| 155 |
+
[13] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976–980. IEEE, 2022.
|
| 156 |
+
[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020.
|
| 157 |
+
[15] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023.
|
| 158 |
+
[16] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1–12, 2017.
|
| 159 |
+
[17] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022.
|
| 160 |
+
[18] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018.
|
| 161 |
+
[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023.
|
| 162 |
+
[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888–12900. PMLR, 2022.
|
| 163 |
+
[21] Yuwei Li, Minye Wu, Yuyao Zhang, Lan Xu, and Jingyi Yu. Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893, 2021.
|
| 164 |
+
[22] Yuwei Li, Longwen Zhang, Zesong Qiu, Yingwenqi Jiang, Nianyi Li, Yuexin Ma, Yuyao Zhang, Lan Xu, and Jingyi Yu. Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG), 41(4):1–16, 2022.
|
| 165 |
+
[23] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004.
|
| 166 |
+
[24] Hengbo Ma, Jiachen Li, Ramtin Hosseini, Masayoshi Tomizuka, and Chiho Choi. Multi-objective diverse human motion prediction with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8161–8171, 2022.
|
| 167 |
+
[25] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019.
|
| 168 |
+
[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1–8, 2022.
|
| 169 |
+
[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022.
|
| 170 |
+
[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002.
|
| 171 |
+
[29] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021.
|
| 172 |
+
[30] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022.
|
| 173 |
+
[31] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big Data, 4(4):236–252, dec 2016.
|
| 174 |
+
[32] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13–26, 2018.
|
| 175 |
+
[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021.
|
| 176 |
+
[34] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018.
|
| 177 |
+
[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019.
|
| 178 |
+
[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020.
|
| 179 |
+
[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019.
|
| 180 |
+
[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 181 |
+
[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022.
|
| 182 |
+
[40] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876–3884, June 2022.
|
| 183 |
+
[41] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023.
|
| 184 |
+
[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050–11059, 2022.
|
| 185 |
+
[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.
|
| 186 |
+
[44] Wataru Takano and Yoshihiko Nakamura. Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research, 34(10):1314–1328, 2015.
|
| 187 |
+
[45] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022.
|
| 188 |
+
[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022.
|
| 189 |
+
[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.
|
| 190 |
+
[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017.
|
| 191 |
+
[49] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566–4575, 2015.
|
| 192 |
+
[50] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022.
|
| 193 |
+
[51] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021.
|
| 194 |
+
[52] Chen Xin, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023.
|
| 195 |
+
[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021.
|
| 196 |
+
[54] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441–3448, 2018.
|
| 197 |
+
[55] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part III, pages 173–191. Springer, 2022.
|
| 198 |
+
[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16, pages 346–364. Springer, 2020.
|
| 199 |
+
[57] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.
|
| 200 |
+
[58] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022.
|
| 201 |
+
[59] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022.
|
| 202 |
+
[60] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019.
|
| 203 |
+
[61] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372–3382, 2021.
|
| 204 |
+
[62] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3955–3963. IEEE Computer Society, 2018.
|
parse/dev/WqiZJGNkjn/WqiZJGNkjn_content_list.json
ADDED
|
@@ -0,0 +1,842 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "MotionGPT: Human Motion as a Foreign Language ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
183,
|
| 8 |
+
102,
|
| 9 |
+
813,
|
| 10 |
+
126
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Biao Jiang1,2∗ Xin Chen2∗ Wen Liu2 Jingyi ${ \\bf { Y } } { \\bf { u } } ^ { 3 }$ Gang $\\mathbf { Y } \\mathbf { u } ^ { 2 }$ Tao Chen1† 1Fudan University 2Tencent 3ShanghaiTech University https://github.com/OpenMotionLab/MotionGPT ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
184,
|
| 19 |
+
150,
|
| 20 |
+
823,
|
| 21 |
+
195
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
231,
|
| 32 |
+
535,
|
| 33 |
+
247
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "Though the advancement of pre-trained large language models unfolds, the exploration of building a unified model for language and other multimodal data, such as motion, remains challenging and untouched so far. Fortunately, human motion displays a semantic coupling akin to human language, often perceived as a form of body language. By fusing language data with large-scale motion models,motionlanguage pre-training that can enhance the performance of motion-related tasks becomes feasible. Driven by this insight, we propose MotionGPT, a unified, versatile, and user-friendly motion-language model to handle multiple motion-relevant tasks. Specifically, we employ the discrete vector quantization for human motion and transfer 3D motion into motion tokens, similar to the generation process of word tokens. Building upon this “motion vocabulary”, we perform language modeling on both motion and text in a unified manner, treating human motion as a specific language. Moreover, inspired by prompt learning, we pre-train MotionGPT with a mixture of motion-language data and fine-tune it on prompt-based questionand-answer tasks. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performances on multiple motion tasks including text-driven motion generation, motion captioning, motion prediction, and motion in-between. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
232,
|
| 42 |
+
263,
|
| 43 |
+
766,
|
| 44 |
+
498
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
174,
|
| 54 |
+
525,
|
| 55 |
+
310,
|
| 56 |
+
542
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "Recent years have witnessed a significant breakthrough in pre-trained large language models such as GPT [34, 35, 3, 27], BERT [7], and T5 [36, 5], which lead to the convergence of language [59, 47], image [33, 50, 20], mesh [55, 26] and mutlimodal [8] modeling. Nevertheless, a general pre-trained model for human motion and language has yet to emerge. This pre-trained motion-language model, capable of supporting numerous motion-relevant tasks through prompts, should benefit diverse fields like gaming, robotics, virtual assistant, and human behavior analysis. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
556,
|
| 66 |
+
825,
|
| 67 |
+
640
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Previous research on human motion has explored various tasks, including motion generation [29, 10, 46, 52, 57], motion captioning [9, 11], and motion prediction [56, 61, 24]. Recent text-to-motion works[46, 58, 30, 52] have attempted to employ pre-trained language-relevant models [7, 33]. For instance, MDM [46] learns a motion diffusion model with conditional text tokens from CLIP [33], while MLD [52] integrates motion latent space to improve the efficiency of motion diffusion process. On the other hand, MotionCLIP [45] and TM2T [11] concentrate on modeling the coupled relationship between motion and text description. However, the above approaches treat motion and language as separate modalities, which often require strictly paired motion and text data. Moreover, since the supervisions are task-specific, they can hardly generalize effectively to unseen tasks or data, as they lack a comprehensive understanding of the relationship between motion and language. We thus focus on building a pre-trained motion-language model, which can generalize to various tasks and learn in-depth motion-language correlation knowledge from more feasible motion and language data. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
647,
|
| 77 |
+
825,
|
| 78 |
+
813
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "Two challenges are crucial and need to be solved for pre-training a promising motion-language model. The first is modeling the relation between language and motion, and the second is building a uniform multi-task framework that can generalize to new tasks. Fortunately, human motion exhibits a semantic coupling similar to human language, often interpreted as a form of body language. Building upon this observation,we follow vision-language pre-training from BEiT-3 [50] to treat human motion as a specific foreign language. By integrating motion and language data together and encoding them within a single vocabulary, the relationship between motion and language becomes more apparent. Therefore, with recent significantly larger-scale language data and models, the motion-language pretraining has great potential to improve the performance on motion tasks. Meanwhile, this pre-training on language enables textual instructions like prompts in InstructGPT [27] and makes the model more versatile and user-friendly for various motion tasks. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
176,
|
| 87 |
+
819,
|
| 88 |
+
825,
|
| 89 |
+
861
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/9737c527cb383ed27a763417e2c1d6f6a049534388d2b3fee34d43bcdf275ced.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: MotionGPT can address diverse motion-relevant tasks uniformly given different instructions. We provide the results on text-to-motion (the upper left), motion captioning (the bottom left), motion completion (the upper right), and the language question-to-answer (the bottom right). The left to right of motion represents the time order. Blue motion denotes the input, and yellow is the generation. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
176,
|
| 102 |
+
89,
|
| 103 |
+
821,
|
| 104 |
+
311
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
388,
|
| 114 |
+
825,
|
| 115 |
+
500
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "In this work, we propose a uniform motion-language framework, namely MotionGPT, that leverages the strong language generation and zero-shot transfer abilities of pre-trained language models for doing human motion-related tasks. To enable MotionGPT to comprehend and generate human-like motions, we first learn a motion-specific vector quantized variational autoencoder (VQ-VAE) model to construct “motion vocabulary”, akin to English vocabulary and then convert raw motion data into a sequence of motion tokens. These tokens are then processed by a pre-trained language model [36, 5] that learns the underlying grammar and syntax of the motion language, as well as its relationship with the corresponding textual descriptions. To effectively integrate language and motion in MotionGPT, we design a two-stage training scheme. We first pre-train the language model on the raw motion dataset to learn the basic grammar and syntax of the motion language. For prompt tuning, we fine-tune the language model on an instruction dataset, which contains both textual descriptions and motion data, to learn the correlation between the two modalities. Extensive experiments demonstrate that MotionGPT achieves state-of-the-art performance on text-to-motion, motion-to-text, motion prediction, and motion in-between. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
507,
|
| 125 |
+
825,
|
| 126 |
+
699
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "We summarize our contributions as follows: (1) We propose a uniform motion-language generative pre-trained model, MotionGPT, which treats human motion as a foreign language, introduces natural language models into motion-relevant generation, and performs diverse motion tasks with a single model. (2) We introduce a motion-language training scheme with instruction tuning, to learn from task feedback and produce promising results through prompts. (3) We propose a general motion benchmark for multi-task evaluation, wherein MotionGPT achieves competitive performance across diverse tasks, including text-to-motion, motion-to-text, motion prediction, and motion in-between, with all available codes and data. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
705,
|
| 136 |
+
825,
|
| 137 |
+
816
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "2 Related Work ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
174,
|
| 147 |
+
837,
|
| 148 |
+
321,
|
| 149 |
+
854
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "Human Motion Synthesis involves generating diverse and realistic human-like motion using multimodal inputs, such as text [10, 30, 58, 46, 11, 1, 17], action [29, 12, 46, 52], and incomplete motion [56, 61, 24, 46]. Text-to-motion is one of the most important motion generation tasks, due to the userfriendly and convenient language input. MDM [46] proposes a diffusion-based generative model [14] separately trained on several motion tasks. MLD [52] advances the latent diffusion model [43, 38] to generate motions based on different conditional inputs. T2M-GPT [57] investigates a generative framework based on VQ-VAE and Generative Pre-trained Transformer (GPT) for motion generation. Motion completion task generates motion conditioning on partial motions, such as classical motion prediction [56, 61, 24] or motion in-between [46], which generates the intermediate motion while the first and last parts are fixed. Although they show promising results in various human motion tasks, most above methods are limited in using a single model to handle multiple tasks. We thus propose a uniform approach that treats human motion as a foreign language, and leverages the strong language generation and zero-shot transfer abilities of pre-trained language models ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
176,
|
| 158 |
+
869,
|
| 159 |
+
825,
|
| 160 |
+
911
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 1
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "table",
|
| 166 |
+
"img_path": "images/889a67c1cf3a409e7e28d4101d676811c9c58a960dc3e7edbdfdb3bf8ac2bedf.jpg",
|
| 167 |
+
"table_caption": [
|
| 168 |
+
"Table 1: Comparison of recent state-of-the-art methods on diverse motion-relevant tasks. Random Motion and Random Caption represent unconstrained generation of motions and motion descriptions. "
|
| 169 |
+
],
|
| 170 |
+
"table_footnote": [],
|
| 171 |
+
"table_body": "<table><tr><td>Methods</td><td></td><td></td><td></td><td></td><td></td><td>Text-to-MotionMotion-to-TextMotion PredictionMotion In-betweenRandom MotionRandom Description</td></tr><tr><td>T2M-GPT[57]</td><td></td><td>xx<x</td><td></td><td></td><td></td><td></td></tr><tr><td>MLD [52]</td><td>：</td><td></td><td></td><td></td><td></td><td>xxx></td></tr><tr><td>TM2T[11]</td><td></td><td></td><td><xx</td><td>xxx</td><td>><x</td><td></td></tr><tr><td>MDM [46]</td><td></td><td></td><td></td><td></td><td>√</td><td>X</td></tr><tr><td>MotionDiffuse[58]</td><td></td><td>X</td><td></td><td></td><td><</td><td>X</td></tr><tr><td>MotionGPT (Ours)</td><td>√</td><td>√</td><td>√</td><td>√</td><td>√</td><td>√</td></tr></table>",
|
| 172 |
+
"bbox": [
|
| 173 |
+
173,
|
| 174 |
+
88,
|
| 175 |
+
825,
|
| 176 |
+
184
|
| 177 |
+
],
|
| 178 |
+
"page_idx": 2
|
| 179 |
+
},
|
| 180 |
+
{
|
| 181 |
+
"type": "text",
|
| 182 |
+
"text": "",
|
| 183 |
+
"bbox": [
|
| 184 |
+
174,
|
| 185 |
+
272,
|
| 186 |
+
825,
|
| 187 |
+
411
|
| 188 |
+
],
|
| 189 |
+
"page_idx": 2
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"type": "text",
|
| 193 |
+
"text": "Human Motion Captioning. To describe human motion with natural languages, [44] learns the mapping from motions to language relying on two statistical models. Furthermore, recurrent networks have also been used in [54, 32]. More recently, TM2T [11] proposed a new motion representation that compresses motions into a short sequence of discrete variables, then uses a neural translation network to build mappings between two modalities. While previous research like TM2T [11] incorporated captioning modules into their training pipeline for motion generation, these approaches are constrained to bidirectional translation between text and motion within one uniform framework. ",
|
| 194 |
+
"bbox": [
|
| 195 |
+
174,
|
| 196 |
+
441,
|
| 197 |
+
825,
|
| 198 |
+
540
|
| 199 |
+
],
|
| 200 |
+
"page_idx": 2
|
| 201 |
+
},
|
| 202 |
+
{
|
| 203 |
+
"type": "text",
|
| 204 |
+
"text": "Language Models and Multi-Modal. Large-scale language models (LLMs) [7, 6, 36, 3, 59, 47], enabled by extensive datasets and model size, have demonstrated impressive comprehension and generation capabilities, elevating natural language processing to new heights. BERT [7] pre-trains deep bidirectional language representations that can support downstream tasks. T5 [36] introduced a unified framework that converts all text-based language problems into a text-to-text format. More recent research [51, 2, 27, 5] find that by fine-tuning pre-trained models using input-output pairs consisting of instructions and coupled answers, the performance of pre-trained models can be further improved. FLAN [5] presents an instruction-tuning technique that surpasses the performance of non-tuned models in unseen tasks. Recently, the wave of multi-modal models [20, 15, 19] is intriguing to process text along with other modalities, such as images [20, 15, 8], audio [13, 8], and videos [53]. CLIP [33] further learns a semantic latent representation that couples images with corresponding language descriptions. Despite the success of language models in various vision-language tasks, the development of multi-modal language models that can handle human motion is still limited. ",
|
| 205 |
+
"bbox": [
|
| 206 |
+
174,
|
| 207 |
+
545,
|
| 208 |
+
825,
|
| 209 |
+
726
|
| 210 |
+
],
|
| 211 |
+
"page_idx": 2
|
| 212 |
+
},
|
| 213 |
+
{
|
| 214 |
+
"type": "text",
|
| 215 |
+
"text": "Motion Language Pre-training. Existing text-to-motion generation methods [10, 30, 46, 11, 1, 17] can be characterized as caption-to-motion, where the models take in a pure text description of the desired motion. While these methods can generate motions from textual descriptions, they are often limited in supporting instructions from users like InstructGPT [27]. In other words, they do not allow users to provide context-specific instructions for certain applications. MotionCLIP [45] utilizes the language and visual understanding of CLIP [33] to align its latent space with a motion auto-encoder. Meanwhile, many language models, such as T5[36] and InstructGPT [27], have been developed to address diverse language processing tasks, including translation, question answering, and classification. These models are typically designed to map a given text input to a target output, such as a translation or answer. However, while these models have shown remarkable performance in language tasks, they have not been widely applied to motion tasks. Therefore, we propose MotionGPT to enable the effective integration of natural language models with human motion tasks, providing a unified solution for motion synthesis problems. ",
|
| 216 |
+
"bbox": [
|
| 217 |
+
174,
|
| 218 |
+
731,
|
| 219 |
+
825,
|
| 220 |
+
911
|
| 221 |
+
],
|
| 222 |
+
"page_idx": 2
|
| 223 |
+
},
|
| 224 |
+
{
|
| 225 |
+
"type": "image",
|
| 226 |
+
"img_path": "images/4ffce89095268f0e8c697a2bd20bc606f20925501ca5be46a40a3036f7cabd2b.jpg",
|
| 227 |
+
"image_caption": [
|
| 228 |
+
"Figure 2: Method overview: MotionGPT consists of a motion tokenizer $\\nu$ (Sec. 3.1) and a motionaware language model (Sec. 3.2). Combining Motion Tokens learned by $\\nu$ and Text Tokens by text tokenizer, we then learn motion and language jointly utilizing language model as backbone. "
|
| 229 |
+
],
|
| 230 |
+
"image_footnote": [],
|
| 231 |
+
"bbox": [
|
| 232 |
+
174,
|
| 233 |
+
89,
|
| 234 |
+
820,
|
| 235 |
+
280
|
| 236 |
+
],
|
| 237 |
+
"page_idx": 3
|
| 238 |
+
},
|
| 239 |
+
{
|
| 240 |
+
"type": "text",
|
| 241 |
+
"text": "3 Method ",
|
| 242 |
+
"text_level": 1,
|
| 243 |
+
"bbox": [
|
| 244 |
+
174,
|
| 245 |
+
358,
|
| 246 |
+
271,
|
| 247 |
+
376
|
| 248 |
+
],
|
| 249 |
+
"page_idx": 3
|
| 250 |
+
},
|
| 251 |
+
{
|
| 252 |
+
"type": "text",
|
| 253 |
+
"text": "To involve large language data and models in the motion generation tasks, we propose a unified motion-language framework named MotionGPT. As illustrated in Fig. 2, MotionGPT consists of a motion tokenizer responsible for converting raw motion data into discrete motion tokens (Sec. 3.1), as well as a motion-aware language model that learns to understand the motion tokens from large language pre-training models by corresponding textual descriptions (Sec. 3.2). To address motionrelevant tasks, we introduce a three-stage training scheme (Sec. 3.3) of MotionGPT for the training of motion tokenizer, motion-language pre-training, and instruction tuning. ",
|
| 254 |
+
"bbox": [
|
| 255 |
+
173,
|
| 256 |
+
388,
|
| 257 |
+
825,
|
| 258 |
+
487
|
| 259 |
+
],
|
| 260 |
+
"page_idx": 3
|
| 261 |
+
},
|
| 262 |
+
{
|
| 263 |
+
"type": "text",
|
| 264 |
+
"text": "We first propose the motion tokenizer consisting of a motion encoder $\\mathcal { E }$ and a motion decoder $\\mathcal { D }$ , to encode a $M$ frame motion $m ^ { 1 : M } = \\{ x ^ { i } \\} _ { i = 1 } ^ { M }$ into $L$ motion tokens $z ^ { 1 : L } = \\{ z ^ { i } \\} _ { i = 1 } ^ { L } , L = M / l$ , and decode $z ^ { 1 : L }$ back into the motion $\\hat { m } ^ { 1 : M } = \\bar { \\mathcal { D } } ( z ^ { 1 : L } ) = \\mathcal { D } ( \\mathcal { E } ( m ^ { 1 : M } ) )$ , where $l$ denotes the temporal downsampling rate on motion length. Then, given an $N$ length sentence $\\boldsymbol { w ^ { 1 : N } } = \\{ w ^ { i } \\} _ { i = 1 } ^ { N }$ describing a motion-related question or demand, MotionGPT aims to generate its answer as $L$ length tokens $\\hat { x } ^ { 1 : L } = \\{ \\hat { x } ^ { i } \\} _ { i = 1 } ^ { L }$ . It could be the human motion tokens $\\hat { x } _ { m } ^ { 1 : L }$ or the text tokens $\\hat { x } _ { t } ^ { 1 : L }$ , which results in a motion $\\hat { m } ^ { 1 : M }$ or a sentence $\\hat { w } ^ { 1 : L }$ like a description of the given motion. ",
|
| 265 |
+
"bbox": [
|
| 266 |
+
173,
|
| 267 |
+
492,
|
| 268 |
+
825,
|
| 269 |
+
593
|
| 270 |
+
],
|
| 271 |
+
"page_idx": 3
|
| 272 |
+
},
|
| 273 |
+
{
|
| 274 |
+
"type": "text",
|
| 275 |
+
"text": "3.1 Motion Tokenizer ",
|
| 276 |
+
"text_level": 1,
|
| 277 |
+
"bbox": [
|
| 278 |
+
174,
|
| 279 |
+
608,
|
| 280 |
+
336,
|
| 281 |
+
623
|
| 282 |
+
],
|
| 283 |
+
"page_idx": 3
|
| 284 |
+
},
|
| 285 |
+
{
|
| 286 |
+
"type": "text",
|
| 287 |
+
"text": "To represent motion in discrete tokens, we pre-train a 3D human motion tokenizer $\\nu$ based on the Vector Quantized Variational Autoencoders (VQ-VAE) architecture used in [48, 42, 11, 57]. Our motion tokenizer consists of an encoder $\\mathcal { E }$ and a decoder $\\mathcal { D }$ . The encoder generates discrete motion tokens with high informative density, while the decoder is able to reconstruct the motion tokens into motion sequences $\\hat { m } ^ { 1 : M }$ . This approach enables us to efficiently represent motion as a language, facilitating the integration of motion and language for various motion-related tasks. ",
|
| 288 |
+
"bbox": [
|
| 289 |
+
174,
|
| 290 |
+
633,
|
| 291 |
+
825,
|
| 292 |
+
718
|
| 293 |
+
],
|
| 294 |
+
"page_idx": 3
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"type": "text",
|
| 298 |
+
"text": "Specifically, the motion encoder $\\mathcal { E }$ first applies 1D convolutions to given frame-wise motion features $\\stackrel { \\bullet } { m } ^ { 1 : M }$ along the time dimension, to obtain latent vectors $\\hat { z } ^ { 1 : L } = \\mathcal { E } ( m ^ { 1 : M } )$ . Next, we transform $\\hat { z }$ of codebooconsists of ntries laten $z$ t rough discrete quantization. Tbedding vectors, each of dimen len able codebookThe process of $Z = \\{ z ^ { i } \\} _ { i = 1 } ^ { K } \\subset \\mathbb { R } ^ { d }$ $K$ $d$ $Q ( \\cdot )$ $b$ $b _ { k }$ $Z$ ",
|
| 299 |
+
"bbox": [
|
| 300 |
+
173,
|
| 301 |
+
723,
|
| 302 |
+
825,
|
| 303 |
+
794
|
| 304 |
+
],
|
| 305 |
+
"page_idx": 3
|
| 306 |
+
},
|
| 307 |
+
{
|
| 308 |
+
"type": "equation",
|
| 309 |
+
"img_path": "images/55b00358c1c9649e29b96be50e67272585ae898a92f84272c07b78221ebae10c.jpg",
|
| 310 |
+
"text": "$$\n\\begin{array} { r } { z _ { i } = Q ( \\hat { z } ^ { i } ) : = \\arg \\operatorname* { m i n } _ { z _ { k } \\in Z } \\left\\| \\hat { z } _ { i } - z _ { k } \\right\\| _ { 2 } . } \\end{array}\n$$",
|
| 311 |
+
"text_format": "latex",
|
| 312 |
+
"bbox": [
|
| 313 |
+
362,
|
| 314 |
+
799,
|
| 315 |
+
635,
|
| 316 |
+
819
|
| 317 |
+
],
|
| 318 |
+
"page_idx": 3
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"type": "text",
|
| 322 |
+
"text": "After quantization, the motion decoder $D$ project $z ^ { 1 : L } = \\{ z ^ { i } \\} _ { i = 1 } ^ { L }$ back to raw motion space as the motion $\\hat { m } ^ { 1 : M }$ with $M$ frames. To train this motion tokenizer, we follow [11, 57] to utilize three distinct loss functions for training and optimizing the motion tokenizer: $\\mathcal { L } _ { \\mathcal { V } } = \\mathcal { L } _ { r } + \\mathcal { L } _ { e } + \\mathcal { L } _ { c }$ , where the reconstruction loss $\\mathcal { L } _ { r }$ , the embedding loss $\\mathcal { L } _ { e }$ , and the commitment loss $\\mathcal { L } _ { c }$ . To further improve the generated motion quality, we follow [57] to utilize L1 smooth loss and velocity regularization in the reconstruction loss, as well as exponential moving average (EMA) and codebook reset techniques [37] to enhance codebook utilization during training. We provide more details about the architecture and the training of our motion tokenizer in the supplement. ",
|
| 323 |
+
"bbox": [
|
| 324 |
+
174,
|
| 325 |
+
825,
|
| 326 |
+
825,
|
| 327 |
+
912
|
| 328 |
+
],
|
| 329 |
+
"page_idx": 3
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"type": "text",
|
| 333 |
+
"text": "",
|
| 334 |
+
"bbox": [
|
| 335 |
+
173,
|
| 336 |
+
90,
|
| 337 |
+
823,
|
| 338 |
+
119
|
| 339 |
+
],
|
| 340 |
+
"page_idx": 4
|
| 341 |
+
},
|
| 342 |
+
{
|
| 343 |
+
"type": "text",
|
| 344 |
+
"text": "3.2 Motion-aware Language Model ",
|
| 345 |
+
"text_level": 1,
|
| 346 |
+
"bbox": [
|
| 347 |
+
176,
|
| 348 |
+
135,
|
| 349 |
+
433,
|
| 350 |
+
150
|
| 351 |
+
],
|
| 352 |
+
"page_idx": 4
|
| 353 |
+
},
|
| 354 |
+
{
|
| 355 |
+
"type": "text",
|
| 356 |
+
"text": "Employing this motion tokenizer, a human motion $m ^ { 1 : M }$ can be mapped to a sequence of motion tokens $\\tilde { z } ^ { 1 : \\overline { { L } } }$ , allowing for joint representation with similar vocabulary embedding in language models [18, 36, 27]. By combining them in the unified vocabulary, we then learn motion and language jointly. We first represent motion tokens $z ^ { 1 : L }$ as a sequence of indices $s ^ { 1 : L } = \\{ s ^ { i } \\} _ { i = 1 } ^ { L }$ , where $\\bar { s } ^ { i }$ corresponds to the index number of motion tokens $z ^ { 1 : L }$ . On the other hand, previous language models, such as T5 [36], encode text as WordPiece tokens. They utilized a vocabulary of $K _ { t }$ word pieces and trained the SentencePiece [18] model on a mixture of language datasets. ",
|
| 357 |
+
"bbox": [
|
| 358 |
+
173,
|
| 359 |
+
160,
|
| 360 |
+
825,
|
| 361 |
+
260
|
| 362 |
+
],
|
| 363 |
+
"page_idx": 4
|
| 364 |
+
},
|
| 365 |
+
{
|
| 366 |
+
"type": "text",
|
| 367 |
+
"text": "Most previous text-to-motion [11, 52, 57] or motion-to-text [11] approaches employ different modules the same way. To achieve this, we combine the original text vocabulary to handle text and motion individually, while we aim to model text and human motion together and in $V _ { t } = \\{ v _ { t } ^ { i } \\} _ { i = 1 } ^ { K _ { t } }$ with motion vocabulary $V _ { m } = \\{ v _ { m } ^ { i } \\} _ { i = 1 } ^ { K _ { m } }$ , which is order-preserving to our motion codebook $Z$ . Moreover, $V _ { m }$ includes several special tokens like boundary indicators, for example, $\\cdot$ as the start and end of the motion. Thus, we employ a new unified text-motion vocabulary $V = \\{ V _ { t } , V _ { m } \\}$ , and can formulate diverse motion-related tasks in a general format, where both input \"words\" and output \"words\" are from the same $V$ . These \"words\" can represent natural language, human motion, or even a mixture of two, depending on the specific task to be solved. Therefore, our MotionGPT allows for the flexible representation and generation of diverse motion-related outputs within a single model. ",
|
| 368 |
+
"bbox": [
|
| 369 |
+
173,
|
| 370 |
+
265,
|
| 371 |
+
825,
|
| 372 |
+
409
|
| 373 |
+
],
|
| 374 |
+
"page_idx": 4
|
| 375 |
+
},
|
| 376 |
+
{
|
| 377 |
+
"type": "text",
|
| 378 |
+
"text": "To address the conditioned generation task, we employ a transformer-based model based on the architecture proposed in [36], which effectively maps the input sequences to the output. Our source input consists of a sequence of tokens $\\ = X _ { s } \\ \\dot { = } \\ \\{ \\dot { x _ { s } } ^ { i } \\} _ { i = 1 } ^ { N }$ , where $x _ { s } ~ \\in ~ V$ and $N$ represents the input length. Similarly, the target output is $X _ { t } ~ = ~ \\{ x _ { t } ^ { ~ i } \\} _ { i = 1 } ^ { L }$ , where $x _ { t } ~ \\in ~ V$ and $L$ denotes the output length. As shown in Fig. 2, the source tokens are fed into the transformer encoder, and the subsequent decoder predicts the probability distribution of the potential next token at each step $\\begin{array} { r } { p _ { \\theta } ( x _ { t } \\mid \\hat { x _ { s } } ) = \\prod _ { i } p _ { \\theta } \\left( \\hat { x } _ { t } ^ { i } \\mid x _ { t } ^ { < i } , x _ { s } \\right) } \\end{array}$ in an autoregressive manner. Therefore, during the training process, the objective is to maximize the log-likelihood of the data distribution: ",
|
| 379 |
+
"bbox": [
|
| 380 |
+
173,
|
| 381 |
+
414,
|
| 382 |
+
825,
|
| 383 |
+
527
|
| 384 |
+
],
|
| 385 |
+
"page_idx": 4
|
| 386 |
+
},
|
| 387 |
+
{
|
| 388 |
+
"type": "equation",
|
| 389 |
+
"img_path": "images/0aa0c7f92a5fb2ddbe06e35380c8718a78118d0e781bfb73f4bea9460c5db9dc.jpg",
|
| 390 |
+
"text": "$$\n\\mathcal { L } _ { L M } = - \\sum _ { i = 0 } ^ { L _ { t } - 1 } \\log p _ { \\theta } \\left( x _ { t } ^ { i } \\mid x _ { t } ^ { < i } , x _ { s } \\right) .\n$$",
|
| 391 |
+
"text_format": "latex",
|
| 392 |
+
"bbox": [
|
| 393 |
+
372,
|
| 394 |
+
530,
|
| 395 |
+
624,
|
| 396 |
+
574
|
| 397 |
+
],
|
| 398 |
+
"page_idx": 4
|
| 399 |
+
},
|
| 400 |
+
{
|
| 401 |
+
"type": "text",
|
| 402 |
+
"text": "By optimizing this objective, MotionGPT learns to capture the underlying patterns and relationships from the data distribution, facilitating the accurate and meaningful generation of the target \"words\". During the inference process, the target tokens are sampled recursively from the predicted distribution $p _ { \\theta } \\left( \\hat { x _ { t } _ { t } } ^ { i } \\mid \\hat { x _ { t } } ^ { < i } , x _ { s } \\right)$ until the end token (i.e., ${ < } / { \\mathrm { s } } > ,$ ). This sampling strategy enables the generation of the target sequence in a step-by-step manner, where each token is probabilistically determined based on the previously generated tokens and the given source input. ",
|
| 403 |
+
"bbox": [
|
| 404 |
+
174,
|
| 405 |
+
577,
|
| 406 |
+
825,
|
| 407 |
+
662
|
| 408 |
+
],
|
| 409 |
+
"page_idx": 4
|
| 410 |
+
},
|
| 411 |
+
{
|
| 412 |
+
"type": "text",
|
| 413 |
+
"text": "3.3 Training Strategy ",
|
| 414 |
+
"text_level": 1,
|
| 415 |
+
"bbox": [
|
| 416 |
+
174,
|
| 417 |
+
678,
|
| 418 |
+
336,
|
| 419 |
+
694
|
| 420 |
+
],
|
| 421 |
+
"page_idx": 4
|
| 422 |
+
},
|
| 423 |
+
{
|
| 424 |
+
"type": "text",
|
| 425 |
+
"text": "Since T5s have only been exposed to language data, represented within a text vocabulary $V _ { t }$ , we thus bridge motion and language and enable this language model to comprehend human motion concepts, by learning the motion vocabulary $V _ { m }$ . As shown in Fig. 3, our training scheme includes three stages: (1) Training of motion tokenizer, which focuses on learning the motion codebook to represent human motion as discrete tokens. (2) Motion-language pre-training stage, which includes unsupervised and supervised objectives to learn the relationship between motion and language. (3) Instruction tuning stage, which tunes the model based on prompt-based instructions for different motion-relevant tasks. ",
|
| 426 |
+
"bbox": [
|
| 427 |
+
173,
|
| 428 |
+
704,
|
| 429 |
+
825,
|
| 430 |
+
801
|
| 431 |
+
],
|
| 432 |
+
"page_idx": 4
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"type": "text",
|
| 436 |
+
"text": "Training of Motion Tokenizer. We first learn the motion tokenizer using the objective defined in Equation 3.1. This training process allows any human motion sequence $\\hat { x } ^ { \\top : L }$ to be represented as a sequence of motion tokens, enabling seamless integration with textual information. Once optimized, the motion tokenizer remains unchanged throughout the subsequent stages of the pipeline. ",
|
| 437 |
+
"bbox": [
|
| 438 |
+
174,
|
| 439 |
+
806,
|
| 440 |
+
825,
|
| 441 |
+
863
|
| 442 |
+
],
|
| 443 |
+
"page_idx": 4
|
| 444 |
+
},
|
| 445 |
+
{
|
| 446 |
+
"type": "text",
|
| 447 |
+
"text": "Motion-language Pre-training Stage. The T5 models [36, 5] are trained and fine-tuned on natural language datasets with instruction-based phrasing [5, 27]. We continue to pre-train this model using a mixture of language and motions data in both unsupervised and supervised manners: 1) To generalize to various downstream tasks like [7, 35, 36, 27], we follow [36] to design an objective, where a certain percentage $( 1 5 \\% )$ of tokens in the input tokens $X _ { s }$ are randomly replaced with a special sentinel token. On the other side, the corresponding target sequence is constructed by extracting the dropped-out spans of tokens, delimited by the same sentinel tokens used in the input sequence, along with an additional sentinel token to indicate the end of the target sequence. 2) We then learn the motion-language relation by the supervision of paired text-motion datasets [10, 31]. We train MotionGPT on the supervised motion-language translation, where the input is either a human motion or a text description.After unsupervised and supervised training processes, we aim to equip our model with the understanding of text and motion relationships. ",
|
| 448 |
+
"bbox": [
|
| 449 |
+
174,
|
| 450 |
+
869,
|
| 451 |
+
825,
|
| 452 |
+
911
|
| 453 |
+
],
|
| 454 |
+
"page_idx": 4
|
| 455 |
+
},
|
| 456 |
+
{
|
| 457 |
+
"type": "image",
|
| 458 |
+
"img_path": "images/7c5af9943a80fd851373e50d6bfdfc3874e25d6ee27159e1f5656fc26f2e717e.jpg",
|
| 459 |
+
"image_caption": [
|
| 460 |
+
"Figure 3: Training Scheme. We introduce three training steps for our MotionGPT (Sec. 3.3): First $\\nu$ learn a codebook for discrete motion representation. Then we train language using a mixture of language and motion data to learn the semantic coupling between text and motion. Finally, we fine-tune the model in a multi-task text-motion dataset with instructions. "
|
| 461 |
+
],
|
| 462 |
+
"image_footnote": [],
|
| 463 |
+
"bbox": [
|
| 464 |
+
173,
|
| 465 |
+
88,
|
| 466 |
+
825,
|
| 467 |
+
314
|
| 468 |
+
],
|
| 469 |
+
"page_idx": 5
|
| 470 |
+
},
|
| 471 |
+
{
|
| 472 |
+
"type": "text",
|
| 473 |
+
"text": "",
|
| 474 |
+
"bbox": [
|
| 475 |
+
173,
|
| 476 |
+
380,
|
| 477 |
+
825,
|
| 478 |
+
505
|
| 479 |
+
],
|
| 480 |
+
"page_idx": 5
|
| 481 |
+
},
|
| 482 |
+
{
|
| 483 |
+
"type": "text",
|
| 484 |
+
"text": "Instruction Tuning Stage. We construct a multi-task text-motion dataset by formulating it as instructions, building upon the foundation of existing text-to-motion datasets such as HumanML3D [10] and KIT [31]. Specifically, we define 15 core motion tasks, such as motion generation with text, motion captioning, motion prediction, and others. For each task, we compose dozens of different instruction templates, resulting in more than one thousand different tasks, each having a unique instruction prompt. For example, an instruction prompt for motion generation task could be “Can you generate a motion sequence that depicts ‘a person emulates the motions of a waltz dance’?”. Similarly, for the motion captioning task, the instruction prompt could be “Provide an accurate caption describing the motion of <motion_tokens>”, where <motion_tokens> represents a sequence of motion tokens generated by our motion tokenizer. We have demonstrated the efficacy of instruction tuning in Sec. 4.3, which leads to improvement across various tasks and enhances the model performance for unseen tasks or prompts. More examples of prompts are provided in the supplements. ",
|
| 485 |
+
"bbox": [
|
| 486 |
+
174,
|
| 487 |
+
511,
|
| 488 |
+
825,
|
| 489 |
+
678
|
| 490 |
+
],
|
| 491 |
+
"page_idx": 5
|
| 492 |
+
},
|
| 493 |
+
{
|
| 494 |
+
"type": "text",
|
| 495 |
+
"text": "4 Experiments ",
|
| 496 |
+
"text_level": 1,
|
| 497 |
+
"bbox": [
|
| 498 |
+
174,
|
| 499 |
+
696,
|
| 500 |
+
313,
|
| 501 |
+
714
|
| 502 |
+
],
|
| 503 |
+
"page_idx": 5
|
| 504 |
+
},
|
| 505 |
+
{
|
| 506 |
+
"type": "text",
|
| 507 |
+
"text": "Extensive comparisons evaluate the performance of our MotionGPTs across multiple motion-relevant tasks and datasets. Details of the dataset settings, evaluation metrics, and implementation specifics (Sec. 4.1) are provided. We first present a uniform benchmark by comparing our approach with other SOTAs across various tasks (Sec. 4.2). Then, we evaluate each specific comparison on text-to-motion (Sec. 4.2), motion-to-text (Sec. 4.2), motion prediction and motion in-between (Sec. 4.2). The supplements include more qualitative results, user studies, and further implementation details. ",
|
| 508 |
+
"bbox": [
|
| 509 |
+
174,
|
| 510 |
+
717,
|
| 511 |
+
825,
|
| 512 |
+
799
|
| 513 |
+
],
|
| 514 |
+
"page_idx": 5
|
| 515 |
+
},
|
| 516 |
+
{
|
| 517 |
+
"type": "text",
|
| 518 |
+
"text": "4.1 Experimental Setup ",
|
| 519 |
+
"text_level": 1,
|
| 520 |
+
"bbox": [
|
| 521 |
+
176,
|
| 522 |
+
815,
|
| 523 |
+
351,
|
| 524 |
+
832
|
| 525 |
+
],
|
| 526 |
+
"page_idx": 5
|
| 527 |
+
},
|
| 528 |
+
{
|
| 529 |
+
"type": "text",
|
| 530 |
+
"text": "Datasets. General motion synthesis can support diverse task settings, and thus previous datasets and a modified benchmark are utilized to evaluate MotionGPT. The study primarily focuses on two text-to-motion datasets: HumanML3D [10] and KIT [31]. The KIT dataset provides 6,353 textual descriptions corresponding to 3,911 motion sequences, while the HumanML3D dataset [10] is a more recent dataset that contains 14,616 motion sequences obtained from AMASS [25], along with 44,970 sequence-level textual descriptions. To evaluate MotionGPT as a uniform framework on tasks, such as motion prediction and motion completion (in-between), we utilize the motion sequences available in HumanML3D, which is also a subset of the larger AMASS dataset. Following the previous works [10, 52, 46], we adopt the same motion representation for fair comparisons, which combines joint velocities, positions, and rotations. By using this consistent representation, MotionGPT enables the availability to support further studies in the field. (cf. supplement for the benchmark details.) ",
|
| 531 |
+
"bbox": [
|
| 532 |
+
174,
|
| 533 |
+
842,
|
| 534 |
+
825,
|
| 535 |
+
911
|
| 536 |
+
],
|
| 537 |
+
"page_idx": 5
|
| 538 |
+
},
|
| 539 |
+
{
|
| 540 |
+
"type": "table",
|
| 541 |
+
"img_path": "images/bd73c831e2dc9475af7b8d56c5e586d847e1266cd7b3186d0e9e112df53e9d9d.jpg",
|
| 542 |
+
"table_caption": [
|
| 543 |
+
"Table 2: Comparison of four motion-related tasks on HumanML3D [10] dataset. The evaluation metrics are computed using the encoder introduced in [10]. The empty columns of previous methods indicate that they can not handle the task. The arrows $( )$ indicate that closer to Real is desirable. Bold and underline indicate the best and the second best result on text-to-motion task. "
|
| 544 |
+
],
|
| 545 |
+
"table_footnote": [],
|
| 546 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">Text-to-Motion</td><td colspan=\"3\">Motion-to-Text</td><td colspan=\"2\">Motion Prediction</td><td colspan=\"2\">Motion In-between</td></tr><tr><td>R TOP1↑</td><td>FID↓</td><td>DIV→</td><td>R TOP3↑</td><td>Bleu@4↑</td><td>Cider↑</td><td>FID↓</td><td>DIV→</td><td>FID↓</td><td>DIV→</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.002±.000</td><td>9.503±.065</td><td>0.828</td><td>-</td><td>-</td><td>0.002</td><td>9.503</td><td>0.002</td><td>9.503</td></tr><tr><td>MLD [52]</td><td>0.481±.003</td><td>0.473±.013</td><td>9.724±.082</td><td>1</td><td></td><td></td><td></td><td>=</td><td>-</td><td></td></tr><tr><td>T2M-GPT[57]</td><td>0.491±.003</td><td>0.116±.004</td><td>9.761±.081</td><td>-</td><td></td><td>-</td><td>2.056</td><td>8.635</td><td>-</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.424±.017</td><td>1.501±.003</td><td>8.589±.076</td><td>0.823</td><td>7.00</td><td>16.8</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MDM [46]</td><td>0.320±005</td><td>0.544±.044</td><td>9.559±.086</td><td>-</td><td>-</td><td>-</td><td>6.031</td><td>7.813</td><td>2.698</td><td>8.420</td></tr><tr><td>MotionGPT(Ours)</td><td>0.492±.003</td><td>0.232±.008</td><td>9.528±.071</td><td>0.827</td><td>12.47</td><td>29.2</td><td>0.905</td><td>8.972</td><td>0.214</td><td>9.560</td></tr></table>",
|
| 547 |
+
"bbox": [
|
| 548 |
+
173,
|
| 549 |
+
88,
|
| 550 |
+
825,
|
| 551 |
+
194
|
| 552 |
+
],
|
| 553 |
+
"page_idx": 6
|
| 554 |
+
},
|
| 555 |
+
{
|
| 556 |
+
"type": "text",
|
| 557 |
+
"text": "",
|
| 558 |
+
"bbox": [
|
| 559 |
+
174,
|
| 560 |
+
287,
|
| 561 |
+
825,
|
| 562 |
+
372
|
| 563 |
+
],
|
| 564 |
+
"page_idx": 6
|
| 565 |
+
},
|
| 566 |
+
{
|
| 567 |
+
"type": "text",
|
| 568 |
+
"text": "Evaluation Metrics are summarized as four parts. (1) Motion quality: Frechet Inception Distance (FID) is our primary metric based on a feature extractor [10] to evaluate the distance of feature distributions between the generated and real motions. For motion completion, we utilize metrics used in motion prediction studies [56, 61, 24], such as Average Displacement Error (ADE) and Final Displacement Error (FDE), to evaluate the accuracy of the predicted motion. (2) Generation diversity: We utilize the Diversity (DIV) metric to assess the motions diversity, which calculates the variance through features extracted from the motions [10]. MultiModality (MM) measures the diversity of generated motions within the same text description of motion. (3) Text matching: Based on the feature space from [10], the motion-retrieval precision (R Precision) evaluates the accuracy of matching between texts and motions using Top 1/2/3 retrieval accuracy. Multi-modal Distance (MM Dist) measures the distance between motions and texts. (4) Linguistic quality: We follow 1] utilizing linguistic metrics from natural language studies, including BLUE [28], Rouge [23], Cider [49], and BertScore [60] to evaluate the quality of generated motion captions. ",
|
| 569 |
+
"bbox": [
|
| 570 |
+
173,
|
| 571 |
+
378,
|
| 572 |
+
825,
|
| 573 |
+
558
|
| 574 |
+
],
|
| 575 |
+
"page_idx": 6
|
| 576 |
+
},
|
| 577 |
+
{
|
| 578 |
+
"type": "text",
|
| 579 |
+
"text": "Implementation Details. We set the codebook of motion tokenizer as $K \\in \\mathbb { R } ^ { 5 1 2 \\times 5 1 2 }$ for most comparisons. The motion encoder $\\mathcal { E }$ incorporates a temporal downsampling rate $l$ of 4. We utilize T5 [36] as the underlying architecture for our language model, with a baseline model consisting of 12 layers in both the transformer encoder and decoder. The feed-forward networks have an output dimensionality of $d _ { \\mathrm { f f } } = 3 0 7 2$ , and the attention mechanisms employ an inner dimensionality of $d _ { \\mathrm { k v } } = 6 4$ . The remaining sub-layers and embeddings have a dimensionality of $d _ { \\mathrm { m o d e l } } = 7 6 8$ Moreover, all our models employ the AdamW optimizer for training. The motion tokenizers are trained utilizing a $1 0 ^ { - 4 }$ learning rate and a 256 mini-batch size, while our language models have a $2 \\times 1 0 ^ { - 4 }$ learning rate for the pre-train stage, $1 0 ^ { - 4 }$ for the instruction tuning stage, and a 16 mini-batch size for both stages. The motion tokenizer undergoes 150K iterations of training, while the language model undergoes 300K iterations during the pre-train stage and another 300K iterations during the instruction tuning stage. Small and Base models are trained on 8 Tesla V100 GPUs while Large models are trianed on 64 Tesla V100 GPUs. ",
|
| 580 |
+
"bbox": [
|
| 581 |
+
173,
|
| 582 |
+
564,
|
| 583 |
+
825,
|
| 584 |
+
742
|
| 585 |
+
],
|
| 586 |
+
"page_idx": 6
|
| 587 |
+
},
|
| 588 |
+
{
|
| 589 |
+
"type": "text",
|
| 590 |
+
"text": "4.2 Comparisons on Motion-relevant Tasks ",
|
| 591 |
+
"text_level": 1,
|
| 592 |
+
"bbox": [
|
| 593 |
+
176,
|
| 594 |
+
761,
|
| 595 |
+
486,
|
| 596 |
+
775
|
| 597 |
+
],
|
| 598 |
+
"page_idx": 6
|
| 599 |
+
},
|
| 600 |
+
{
|
| 601 |
+
"type": "text",
|
| 602 |
+
"text": "Comparisons on Multiple Tasks. By introducing a uniform framework that treats human motion as a foreign language, we open up the exploration of diverse motion-relevant tasks. We employ a 220M pre-trained Flan-T5-Base[36, 5] model as our backbone and fine-tune the model through the pre-training and instruction tuning stage (Sec. 3.3) for all following comparisons. As shown in Tab. 2, we evaluate MotionGPT against state-of-the-art methods on key tasks such as text-conditioned motion generation [52, 57, 11, 46], motion captioning [11], motion prediction [46], and motion in-between[46]. While we leverage existing results from previous works or benchmarks for textto-motion and motion-to-text tasks, we re-implement the motion diffusion models [46] for motion prediction and evaluate it under the same metrics and settings. Please note that some methods are designed for specific tasks, and thus some metrics are empty for tasks they cannot handle. The results presented in Tab. 2 demonstrate that our MotionGPT achieves competitive performance across all evaluated tasks, highlighting its capability to address diverse motion tasks within a single model. ",
|
| 603 |
+
"bbox": [
|
| 604 |
+
174,
|
| 605 |
+
786,
|
| 606 |
+
825,
|
| 607 |
+
911
|
| 608 |
+
],
|
| 609 |
+
"page_idx": 6
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"type": "table",
|
| 613 |
+
"img_path": "images/5c9137490dcb540d89abfd13e20c421752f2c3c898d943e8af326e89001154e4.jpg",
|
| 614 |
+
"table_caption": [
|
| 615 |
+
"Table 3: Comparison of text-to-motion on HumanML3D [10]. The empty MModality indicates Real motion is deterministic. These methods are sorted by FID. $P r e$ -trained and Fine-tuned indicate uniform motion-language pre-training and specific fine-tuning on this task. $( c f$ . Tab. 2 for notations.) "
|
| 616 |
+
],
|
| 617 |
+
"table_footnote": [],
|
| 618 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"3\">RPrecision↑</td><td rowspan=\"2\">FID↓</td><td rowspan=\"2\">MMDist↓</td><td rowspan=\"2\">Diversity→</td><td rowspan=\"2\">MModality↑</td></tr><tr><td>Top1</td><td>Top2</td><td>Top3</td></tr><tr><td>Real</td><td>0.511±.003</td><td>0.703±.003</td><td>0.797±.002</td><td>0.002±.000</td><td>2.974±.008</td><td>9.503±.065</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.424±.003</td><td>0.618±.003</td><td>0.729±.002</td><td>1.501±.017</td><td>3.467±.011</td><td>8.589±.076</td><td>2.424±.093</td></tr><tr><td>T2M[10]</td><td>0.457±.002</td><td>0.639±.003</td><td>0.740±.003</td><td>1.067±.002</td><td>3.340±.008</td><td>9.188±.002</td><td>2.090±.083</td></tr><tr><td>MotionDiffuse [58]</td><td>0.491±.001</td><td>0.681±.001</td><td>0.782±.001</td><td>0.630±.001</td><td>3.113±.001</td><td>9.410±.049</td><td>1.553±.042</td></tr><tr><td>MDM [46]</td><td>0.320±.005</td><td>0.498±.004</td><td>0.611±.007</td><td>0.544±.044</td><td>5.566±.027</td><td>9.559±.086</td><td>2.799±.072</td></tr><tr><td>MLD [52]</td><td>0.481±.003</td><td>0.673±.003</td><td>0.772±.002</td><td>0.473±.013</td><td>3.196±.010</td><td>9.724±.082</td><td>2.413±.079</td></tr><tr><td>T2M-GPT [57]</td><td>0.491±.003</td><td>0.680±.003</td><td>0.775±.002</td><td>0.116±.004</td><td>3.118±.011</td><td>9.761±.081</td><td>1.856±.011</td></tr><tr><td>MotionGPT (Pre-trained)</td><td>0.435±.003</td><td>0.607±.002</td><td>0.700±.002</td><td>0.160±.008</td><td>3.700±.009</td><td>9.411±.081</td><td>3.437±.091</td></tr><tr><td>MotionGPT (Fine-tuned)</td><td>0.492±.003</td><td>0.681±.003</td><td>0.778±.002</td><td>0.232±.008</td><td>3.096±.008</td><td>9.528±.071</td><td>2.008±.084</td></tr></table>",
|
| 619 |
+
"bbox": [
|
| 620 |
+
173,
|
| 621 |
+
101,
|
| 622 |
+
825,
|
| 623 |
+
251
|
| 624 |
+
],
|
| 625 |
+
"page_idx": 7
|
| 626 |
+
},
|
| 627 |
+
{
|
| 628 |
+
"type": "table",
|
| 629 |
+
"img_path": "images/6f2750c6c5675112100be79ae6a2a64a11d1b5cc4912a605ab55b446e64f48b0.jpg",
|
| 630 |
+
"table_caption": [
|
| 631 |
+
"Table 4: Comparison of motion captioning on HumanML3D [10]. The evaluation metrics follow [11], while we use the ground truth texts without pre-processing for linguistic metrics calculation. "
|
| 632 |
+
],
|
| 633 |
+
"table_footnote": [],
|
| 634 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td colspan=\"2\">RPrecision↑</td><td rowspan=\"2\">MMDist↓</td><td rowspan=\"2\">Lengthavg↑</td><td rowspan=\"2\">Bleu@1↑Bleu@4↑Rouge↑Cider↑BertScore↑</td><td rowspan=\"2\"></td><td rowspan=\"2\"></td><td rowspan=\"2\"></td><td rowspan=\"2\"></td></tr><tr><td>Top1</td><td>Top3</td></tr><tr><td>Real</td><td>0.523</td><td>0.828</td><td>2.901</td><td>12.75</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>TM2T[11]</td><td>0.516</td><td>0.823</td><td>2.935</td><td>10.67</td><td>48.9</td><td>7.00</td><td>38.1</td><td>16.8</td><td>32.2</td></tr><tr><td>MotionGPT (Ours)</td><td>0.543</td><td>0.827</td><td>2.821</td><td>13.04</td><td>48.2</td><td>12.47</td><td>37.4</td><td>29.2</td><td>32.4</td></tr></table>",
|
| 635 |
+
"bbox": [
|
| 636 |
+
173,
|
| 637 |
+
344,
|
| 638 |
+
825,
|
| 639 |
+
422
|
| 640 |
+
],
|
| 641 |
+
"page_idx": 7
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"type": "text",
|
| 645 |
+
"text": "",
|
| 646 |
+
"bbox": [
|
| 647 |
+
178,
|
| 648 |
+
502,
|
| 649 |
+
821,
|
| 650 |
+
545
|
| 651 |
+
],
|
| 652 |
+
"page_idx": 7
|
| 653 |
+
},
|
| 654 |
+
{
|
| 655 |
+
"type": "text",
|
| 656 |
+
"text": "Comparisons on Text-to-Motion. The text-to-motion task involves generating human motion sequences based on a given text input. We evaluate the proposed the MotionGPT model as the pre-trained MotionGPT, the same one in Tab. 2, as well as fine-tuned it on text-to-motion task. We compare our MotionGPTs with other SOTAs [11, 10, 46, 52, 57] and evaluate the performance on both HumanML3D and KIT datasets using suggested metrics [10]. The results are computed with a $9 5 \\%$ confidence interval, obtained from 20 repeated runs. The majority of the reported results are taken directly from their own papers or the benchmark presented in [10]. Tab. 3 summarizes the comparison results, where MotionGPT achieves competitive performance on most metrics. ",
|
| 657 |
+
"bbox": [
|
| 658 |
+
174,
|
| 659 |
+
550,
|
| 660 |
+
825,
|
| 661 |
+
662
|
| 662 |
+
],
|
| 663 |
+
"page_idx": 7
|
| 664 |
+
},
|
| 665 |
+
{
|
| 666 |
+
"type": "text",
|
| 667 |
+
"text": "Comparisons on Motion-to-Text. The motion-to-text task involves generating a text description based on a given human motion sequence. We compare the pre-trained MotionGPT with recent work TM2T [11]. We evaluate the performance on the HumanML3D using the suggested metrics from [11]. Additionally, we measure the average numbers of words Length $\\mathbf { a v } \\mathbf { g }$ for further comparisons. Please note that the reported results in [11] are evaluated with pre-processed ground truth text, which ignores the grammatical tense and plural forms of words. In Tab. 4, we directly use the ground truth text descriptions for a more accurate assessment. This comparison shows that MotionGPT overperforms recent work on text descriptions of given motions. ",
|
| 668 |
+
"bbox": [
|
| 669 |
+
173,
|
| 670 |
+
667,
|
| 671 |
+
825,
|
| 672 |
+
781
|
| 673 |
+
],
|
| 674 |
+
"page_idx": 7
|
| 675 |
+
},
|
| 676 |
+
{
|
| 677 |
+
"type": "text",
|
| 678 |
+
"text": "Comparisons on Motion Prediction and In-between. We summarize motion prediction and in-between together as general motion completion. To evaluate the motion completion capability of MotionGPT, we employ part of the AMASS dataset [25], a motion-only dataset. For motion prediction task, we only input around the first $20 \\%$ of the motion sequence as conditions. For in-between, we mask about $50 \\%$ motion randomly for completion. We also fine-tune MotionGPT specifically for this task and employ FID, ADE, and FDE as metrics like Sec. 4.1. Furthermore, we evaluate MDM [46] on motion prediction by utilizing their provided model, which also supports motion in-between through masked motion “in-painting”. The real motion data is used as one of our baselines. Tab. 5 reports that our MotionGPT has the best motion completion quality and diversity. ",
|
| 679 |
+
"bbox": [
|
| 680 |
+
174,
|
| 681 |
+
786,
|
| 682 |
+
825,
|
| 683 |
+
911
|
| 684 |
+
],
|
| 685 |
+
"page_idx": 7
|
| 686 |
+
},
|
| 687 |
+
{
|
| 688 |
+
"type": "table",
|
| 689 |
+
"img_path": "images/27e05cccf1d5223ffe65c9e9844239d7bfb3eac6c57eddca6c31866b0d23a7c1.jpg",
|
| 690 |
+
"table_caption": [
|
| 691 |
+
"Table 5: Comparison of motion prediction and motion in-between on part of AMASSS [25] dataset using motion data only. FID indicates motion quality and Diversity (DIV) for motion diversity within each condition. ADE and FDE are joints distance between generation and ground truth. "
|
| 692 |
+
],
|
| 693 |
+
"table_footnote": [],
|
| 694 |
+
"table_body": "<table><tr><td rowspan=\"3\">Methods</td><td colspan=\"4\">Motion Prediction</td><td colspan=\"3\">Motion In-between</td></tr><tr><td>FID↓</td><td>Diversity↑</td><td>ADE↓</td><td>FDE↓</td><td>FID↓</td><td>Diversity↑</td><td>ADE↓</td></tr><tr><td>Real</td><td>0.002</td><td>9.503</td><td>-</td><td>-</td><td>0.002</td><td>9.503</td><td>-</td></tr><tr><td>MDM[46]</td><td>6.031</td><td>7.813</td><td>5.446</td><td>8.561</td><td>2.698</td><td>8.420</td><td>3.787</td></tr><tr><td>T2M-GPT[57]</td><td>2.056</td><td>8.635</td><td>6.161</td><td>8.302</td><td>-</td><td>-</td><td>-</td></tr><tr><td>MotionGPT (Ours)</td><td>0.905</td><td>8.972</td><td>4.745</td><td>6.040</td><td>0.214</td><td>9.560</td><td>3.762</td></tr></table>",
|
| 695 |
+
"bbox": [
|
| 696 |
+
235,
|
| 697 |
+
88,
|
| 698 |
+
758,
|
| 699 |
+
178
|
| 700 |
+
],
|
| 701 |
+
"page_idx": 8
|
| 702 |
+
},
|
| 703 |
+
{
|
| 704 |
+
"type": "table",
|
| 705 |
+
"img_path": "images/3f03a6565272e34f8c35ff3cdeb186e6e03173d62074499a7a371d6ae0d265e7.jpg",
|
| 706 |
+
"table_caption": [],
|
| 707 |
+
"table_footnote": [
|
| 708 |
+
"Table 6: Evaluation of instruction tuning and different model sizes of MotionGPTs in four motion tasks on HumanML3D [10] dataset. $( c f$ . Tab. 2 for metrics details) "
|
| 709 |
+
],
|
| 710 |
+
"table_body": "<table><tr><td rowspan=\"2\">Size</td><td rowspan=\"2\">Instruction Tuning</td><td colspan=\"3\">Text-to-Motion</td><td colspan=\"3\">Motion-to-Text</td><td colspan=\"2\">Motion Prediction</td><td colspan=\"2\">Motion In-between</td></tr><tr><td>R TOP3↑</td><td>FID←</td><td>DIV→</td><td>MMDist↓</td><td>Bleu@4↑</td><td>Cider↑</td><td>FID↓</td><td>DIV→</td><td>FID↓</td><td>DIV→</td></tr><tr><td>Real</td><td>-</td><td>0.797</td><td>0.002</td><td>9.503</td><td>2.901</td><td>-</td><td>1</td><td>0.002</td><td>9.503</td><td>0.002</td><td>9.503</td></tr><tr><td>Small</td><td></td><td>0.706</td><td>0.727</td><td>9.264</td><td>2.748</td><td>12.02</td><td>24.9</td><td>-</td><td>-</td><td>1</td><td>-</td></tr><tr><td>Small</td><td>√</td><td>0.663</td><td>0.336</td><td>9.239</td><td>2.931</td><td>10.54</td><td>24.3</td><td>0.954</td><td>8.727</td><td>0.326</td><td>9.618</td></tr><tr><td>Base</td><td></td><td>0.722</td><td>0.365</td><td>9.407</td><td>2.821</td><td>12.47</td><td>29.2</td><td></td><td>-</td><td></td><td></td></tr><tr><td>Base</td><td>√</td><td>0.700</td><td>0.160</td><td>9.411</td><td>3.019</td><td>11.42</td><td>28.2</td><td>0.905</td><td>8.972</td><td>0.214</td><td>9.560</td></tr><tr><td>Large</td><td></td><td>0.694</td><td>0.234</td><td>9.310</td><td>2.776</td><td>12.44</td><td>28.5</td><td>1</td><td>-</td><td>1</td><td>1</td></tr><tr><td>Large</td><td>√</td><td>0.708</td><td>0.159</td><td>9.301</td><td>3.011</td><td>11.71</td><td>29.1</td><td>0.556</td><td>8.975</td><td>0.223</td><td>9.358</td></tr></table>",
|
| 711 |
+
"bbox": [
|
| 712 |
+
173,
|
| 713 |
+
232,
|
| 714 |
+
825,
|
| 715 |
+
352
|
| 716 |
+
],
|
| 717 |
+
"page_idx": 8
|
| 718 |
+
},
|
| 719 |
+
{
|
| 720 |
+
"type": "text",
|
| 721 |
+
"text": "4.3 Ablation Studies ",
|
| 722 |
+
"text_level": 1,
|
| 723 |
+
"bbox": [
|
| 724 |
+
174,
|
| 725 |
+
400,
|
| 726 |
+
328,
|
| 727 |
+
415
|
| 728 |
+
],
|
| 729 |
+
"page_idx": 8
|
| 730 |
+
},
|
| 731 |
+
{
|
| 732 |
+
"type": "text",
|
| 733 |
+
"text": "MotionGPT employs T5 [36] as the motion-aware language backbone model, and we train these models with pre-training and then instruction tuning. Thus, both model size and training strategy influence the performance of MotionGPTs. We here evaluate them on the typical motion tasks. More detailed ablation studies are provided in the supplements. ",
|
| 734 |
+
"bbox": [
|
| 735 |
+
174,
|
| 736 |
+
425,
|
| 737 |
+
825,
|
| 738 |
+
481
|
| 739 |
+
],
|
| 740 |
+
"page_idx": 8
|
| 741 |
+
},
|
| 742 |
+
{
|
| 743 |
+
"type": "text",
|
| 744 |
+
"text": "Model Sizes. We evaluate the performance of models with different sizes across four motion tasks. Besides the base 220M MotionGPT in Sec. 4.1, we now evaluate 60M, 220M, and 770M MotionGPTs. Tab. 6 demonstrates that the 220M base model has achieved remarkable performance compared to the smaller 60M model. However, the larger model size of current Motions does not yield significant improvements and, in few cases, even leads to worse results, as observed in the motion in-between task. We believe this could be caused by the small amount of current motion datasets. HumanML3D only includes $1 5 \\mathrm { k }$ motion sequences, much smaller than even billions of language and image data. ",
|
| 745 |
+
"bbox": [
|
| 746 |
+
174,
|
| 747 |
+
487,
|
| 748 |
+
825,
|
| 749 |
+
584
|
| 750 |
+
],
|
| 751 |
+
"page_idx": 8
|
| 752 |
+
},
|
| 753 |
+
{
|
| 754 |
+
"type": "text",
|
| 755 |
+
"text": "Effectiveness of Instruction Tuning. We evaluate the impact of our instruction tuning strategy on different model sizes. The results in Tab. 6 demonstrate that instruction tuning enhances the versatility of MotionGPT, enabling more motion tasks like motion completion and improving the motion performance of the text-to-motion task. However, for pure text-generation tasks, the model performance is downgraded, likely due to the pair amount of textual descriptions and coupled motions. ",
|
| 756 |
+
"bbox": [
|
| 757 |
+
174,
|
| 758 |
+
590,
|
| 759 |
+
825,
|
| 760 |
+
660
|
| 761 |
+
],
|
| 762 |
+
"page_idx": 8
|
| 763 |
+
},
|
| 764 |
+
{
|
| 765 |
+
"type": "text",
|
| 766 |
+
"text": "5 Disscusion ",
|
| 767 |
+
"text_level": 1,
|
| 768 |
+
"bbox": [
|
| 769 |
+
173,
|
| 770 |
+
679,
|
| 771 |
+
292,
|
| 772 |
+
695
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 8
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "text",
|
| 778 |
+
"text": "As the first trial, to our best knowledge, exploring human motion generation using language models, the proposed MotionGPT still owns limitations as follows. MotionGPT only utilizes motion on articulated human bodies, while many other works focus on faces [16, 4], hands [39, 22, 21] and even animal [40, 62] motion. Besides, our method is also restricted to multiple humans without modeling human-object, or human-environment interactions [41]. It is interesting to model the human interaction scenarios in a motion-language framework and generate controllable motions [41]. ",
|
| 779 |
+
"bbox": [
|
| 780 |
+
174,
|
| 781 |
+
710,
|
| 782 |
+
825,
|
| 783 |
+
794
|
| 784 |
+
],
|
| 785 |
+
"page_idx": 8
|
| 786 |
+
},
|
| 787 |
+
{
|
| 788 |
+
"type": "text",
|
| 789 |
+
"text": "We summarize the proposed MotionGPT as a uniform motion-language framework to generate plausible human motion and natural language descriptions through prompt-based instructions. Compared to the compatible motion diffusion methods [52, 46], our MotionGPT produces competitive results on motion generation, motion captioning, motion prediction, and motion in-between using only one pre-trained generative model. With the advancement of large language data and models [36, 5], MotionGPT is also capable of addressing natural question-to-answer tasks. Extensive experiments on various human motion-relevant tasks demonstrate the effectiveness and extendibility of MotionGPT. ",
|
| 790 |
+
"bbox": [
|
| 791 |
+
174,
|
| 792 |
+
800,
|
| 793 |
+
825,
|
| 794 |
+
897
|
| 795 |
+
],
|
| 796 |
+
"page_idx": 8
|
| 797 |
+
},
|
| 798 |
+
{
|
| 799 |
+
"type": "text",
|
| 800 |
+
"text": "References \n[1] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019. \n[2] Stephen H Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, et al. Promptsource: An integrated development environment and repository for natural language prompts. arXiv preprint arXiv:2202.01279, 2022. \n[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020. \n[4] Xuan Cao, Zhang Chen, Anpei Chen, Xin Chen, Shiying Li, and Jingyi Yu. Sparse photometric 3d face reconstruction guided by morphable models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4635–4644, 2018. \n[5] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. \n[6] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. \n[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. \n[8] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. arXiv preprint arXiv:2305.05665, 2023. \n[9] Yusuke Goutsu and Tetsunari Inamura. Linguistic descriptions of human motion with generative adversarial seq2seq learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4281–4287. IEEE, 2021. \n[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5152–5161, June 2022. \n[11] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In ECCV, 2022. \n[12] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021–2029, 2020. \n[13] Andrey Guzhov, Federico Raue, Jörn Hees, and Andreas Dengel. Audioclip: Extending clip to image, text and audio. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 976–980. IEEE, 2022. \n[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020. \n[15] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. \n[16] Tero Karras, Timo Aila, Samuli Laine, Antti Herva, and Jaakko Lehtinen. Audio-driven facial animation by joint end-to-end learning of pose and emotion. ACM Transactions on Graphics (TOG), 36(4):1–12, 2017. \n[17] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Free-form language-based motion synthesis & editing. arXiv preprint arXiv:2209.00349, 2022. \n[18] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. \n[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. \n[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888–12900. PMLR, 2022. \n[21] Yuwei Li, Minye Wu, Yuyao Zhang, Lan Xu, and Jingyi Yu. Piano: A parametric hand bone model from magnetic resonance imaging. arXiv preprint arXiv:2106.10893, 2021. \n[22] Yuwei Li, Longwen Zhang, Zesong Qiu, Yingwenqi Jiang, Nianyi Li, Yuexin Ma, Yuyao Zhang, Lan Xu, and Jingyi Yu. Nimble: a non-rigid hand model with bones and muscles. ACM Transactions on Graphics (TOG), 41(4):1–16, 2022. \n[23] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81, 2004. \n[24] Hengbo Ma, Jiachen Li, Ramtin Hosseini, Masayoshi Tomizuka, and Chiho Choi. Multi-objective diverse human motion prediction with knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8161–8171, 2022. \n[25] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. \n[26] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1–8, 2022. \n[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730–27744, 2022. \n[28] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. \n[29] Mathis Petrovich, Michael J. Black, and Gül Varol. Action-conditioned 3D human motion synthesis with transformer VAE. In International Conference on Computer Vision (ICCV), 2021. \n[30] Mathis Petrovich, Michael J. Black, and Gül Varol. TEMOS: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision (ECCV), 2022. \n[31] Matthias Plappert, Christian Mandery, and Tamim Asfour. The kit motion-language dataset. Big Data, 4(4):236–252, dec 2016. \n[32] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13–26, 2018. \n[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. \n[34] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. \n[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. \n[36] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020. \n[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. \n[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. \n[39] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. \n[40] Nadine Rueegg, Silvia Zuffi, Konrad Schindler, and Michael J. Black. BARC: Learning to regress 3D dog shape from images by exploiting breed information. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3876–3884, June 2022. \n[41] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. arXiv preprint arXiv:2303.01418, 2023. \n[42] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050–11059, 2022. \n[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. \n[44] Wataru Takano and Yoshihiko Nakamura. Statistical mutual conversion between whole body motion primitives and linguistic sentences for human motions. The International Journal of Robotics Research, 34(10):1314–1328, 2015. \n[45] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. \n[46] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. \n[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. \n[48] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. \n[49] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566–4575, 2015. \n[50] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. \n[51] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. \n[52] Chen Xin, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. \n[53] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. \n[54] Tatsuro Yamada, Hiroyuki Matsunaga, and Tetsuya Ogata. Paired recurrent autoencoders for bidirectional translation between robot actions and linguistic descriptions. IEEE Robotics and Automation Letters, 3(4):3441–3448, 2018. \n[55] Kim Youwang, Kim Ji-Yeon, and Tae-Hyun Oh. Clip-actor: Text-driven recommendation and stylization for animating human meshes. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part III, pages 173–191. Springer, 2022. \n[56] Ye Yuan and Kris Kitani. Dlow: Diversifying latent flows for diverse human motion prediction. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16, pages 346–364. Springer, 2020. \n[57] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. \n[58] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. \n[59] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. \n[60] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019. \n[61] Yan Zhang, Michael J Black, and Siyu Tang. We are more than our joints: Predicting how 3d bodies move. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3372–3382, 2021. \n[62] Silvia Zuffi, Angjoo Kanazawa, and Michael J. Black. Lions and tigers and bears: Capturing non-rigid, 3D, articulated shape from images. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3955–3963. IEEE Computer Society, 2018. ",
|
| 801 |
+
"bbox": [
|
| 802 |
+
173,
|
| 803 |
+
39,
|
| 804 |
+
828,
|
| 805 |
+
922
|
| 806 |
+
],
|
| 807 |
+
"page_idx": 9
|
| 808 |
+
},
|
| 809 |
+
{
|
| 810 |
+
"type": "text",
|
| 811 |
+
"text": "",
|
| 812 |
+
"bbox": [
|
| 813 |
+
171,
|
| 814 |
+
60,
|
| 815 |
+
828,
|
| 816 |
+
914
|
| 817 |
+
],
|
| 818 |
+
"page_idx": 10
|
| 819 |
+
},
|
| 820 |
+
{
|
| 821 |
+
"type": "text",
|
| 822 |
+
"text": "",
|
| 823 |
+
"bbox": [
|
| 824 |
+
171,
|
| 825 |
+
73,
|
| 826 |
+
828,
|
| 827 |
+
917
|
| 828 |
+
],
|
| 829 |
+
"page_idx": 11
|
| 830 |
+
},
|
| 831 |
+
{
|
| 832 |
+
"type": "text",
|
| 833 |
+
"text": "",
|
| 834 |
+
"bbox": [
|
| 835 |
+
171,
|
| 836 |
+
92,
|
| 837 |
+
828,
|
| 838 |
+
310
|
| 839 |
+
],
|
| 840 |
+
"page_idx": 12
|
| 841 |
+
}
|
| 842 |
+
]
|
parse/dev/WqiZJGNkjn/WqiZJGNkjn_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/WqiZJGNkjn/WqiZJGNkjn_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/XA4ru9mfxTP/XA4ru9mfxTP.md
ADDED
|
@@ -0,0 +1,283 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Unifying Voxel-based Representation with Transformer for 3D Object Detection
|
| 2 |
+
|
| 3 |
+
Yanwei Li1 Yilun Chen1 Xiaojuan Qi2 Zeming Li3 Jian Sun3 Jiaya Jia1,4
|
| 4 |
+
|
| 5 |
+
The Chinese University of Hong Kong1 The University of Hong Kong2 MEGVII Technology3 SmartMore4
|
| 6 |
+
|
| 7 |
+
# Abstract
|
| 8 |
+
|
| 9 |
+
In this work, we present a unified framework for multi-modality 3D object detection, named UVTR. The proposed method aims to unify multi-modality representations in the voxel space for accurate and robust single- or cross-modality 3D detection. To this end, the modality-specific space is first designed to represent different inputs in the voxel feature space. Different from previous work, our approach preserves the voxel space without height compression to alleviate semantic ambiguity and enable spatial connections. To make full use of the inputs from different sensors, the cross-modality interaction is then proposed, including knowledge transfer and modality fusion. In this way, geometry-aware expressions in point clouds and context-rich features in images are well utilized for better performance and robustness. The transformer decoder is applied to efficiently sample features from the unified space with learnable positions, which facilitates object-level interactions. In general, UVTR presents an early attempt to represent different modalities in a unified framework. It surpasses previous work in single- or multi-modality entries. The proposed method achieves leading performance in the nuScenes test set for both object detection and the following object tracking task. Code is made publicly available at https://github.com/dvlab-research/UVTR.
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Detecting 3D objects with multi-modality sensors (i.e., LiDAR and camera) is regarded as a fundamental task in real-world scenes. For accurate object detection, data from different modalities are utilized to provide complementary knowledge, like accurate positions from point clouds and rich context from images. Toward this purpose, a unified representation is essential to facilitate knowledge transfer and feature fusion across modalities. However, due to the lack of accurate depth from cameras, images can not be naturally represented in voxel space like that of point clouds.
|
| 14 |
+
|
| 15 |
+
In the unified progress, several representations have been studied that can be roughly separated into input- and feature-level streams. For the first one, multi-modality data is aligned at the beginning of network. In particular, pseudo point clouds in Figure 1a are transformed from image aided by predicted depth [1, 2], while the range-view image in Figure 1b is projected from point clouds [3, 4]. Because of inaccurate depth in pseudo point clouds and collapsed 3D geometry in range-view images, the spatial structure of data is damaged, which brings inferior results. For feature-level method, a typical approach is to transform image features as frustum and then compress to BEV space [5, 6], like that in Figure 1c. However, due to ray-like trajectories in the frustum [7], height compression at each position aggregates features from various objects and thus introduces semantic ambiguity. Meanwhile, other implicit manners in contemporary work [8, 9, 10] can hardly support explicit feature interactions in 3D space and restrict further knowledge transfer. Therefore, a more unified representation is desired to bridge modality gap and facilitate interactions from multiple aspects.
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
Figure 1: Toy example of methods for unified representation. Compared with others, the proposed manner in 1d constructs the voxel space by sampling features from the image plane and represents multi-modalities uniformly without height-level compression in 1c that brings semantic ambiguity.
|
| 19 |
+
|
| 20 |
+
In this paper, we present a simple yet effective framework to unify the voxel-based representation with transformer, called UVTR. In particular, features from images and point clouds are represented and interacted in the explicit voxel-based space. For images, we construct the voxel space by sampling features from the image plane according to predicted depth scores and geometric constraints, as briefly depicted in Figure 1d. For point clouds, the accurate position naturally allows us to associate features with voxels. Then, voxel encoder is introduced for spatial interaction that establishes the relationship among adjacent features. In this way, cross-modality interaction is naturally conducted with features in each voxel space. For object-level interaction, deformable transformer [11] is adopted as the decoder that samples specific feature for each object query with position $( x , y , z )$ in the unified voxel space, as illustrated in Figure 1d. Meanwhile, the introduction of 3D query position efficiently alleviates the semantic ambiguity brought by height compression in BEV space as analysed before.
|
| 21 |
+
|
| 22 |
+
Compared with previous and even concurrent studies [8, 9], more key advances can be achieved with the proposed framework. First, the explicit voxel-based representation supports spatial interaction in 3D space and multi-frame scenes that bring significant improvements. Second, the proposed unified manner facilitates cross-modality learning and can be naturally applied for knowledge transfer and feature fusion, which further boosts the performance. Finally, data augmentation for both modalities can be directly synchronized in the voxel space without the complex aligning process [12, 7].
|
| 23 |
+
|
| 24 |
+
The overall framework, called UVTR, can be easily instantiated and improved with various imageor voxel-based backbones for single- and multi-modality 3D object detection. Extensive empirical studies are conducted in Section 4 to reveal the effect of each component. The proposed UVTR attains leading performance in various settings. For detection, it achieves $6 9 . 7 \%$ , $5 5 . 1 \%$ , and $7 1 . 1 \%$ NDS on nuScenes test set with point clouds, images, and multi-modality inputs, respectively. Given naive association strategy, UVTR also achieves strong tracking results with $6 7 . 0 \%$ , ${ \mathfrak { s } } 1 . 9 \%$ , and $7 0 . 1 \%$ AMOTA on LiDAR-based, camera-based, and multi-modality setting, respectively.
|
| 25 |
+
|
| 26 |
+
# 2 Related Work
|
| 27 |
+
|
| 28 |
+
LiDAR-based 3D Detection. With point clouds captured from LiDAR, traditional methods process the irregular input and generate 3D boxes with different representations, e.g., point, voxel, and range view. Point-based detectors usually aggregate features from raw point clouds with set abstraction [13] and then predict box proposals [14, 15, 16]. For voxel-based methods, point clouds are transformed into regular grids and processed with 3D sparse convolutions [17, 18] or 2D convolutions [19, 20, 21] directly. Final predictions are usually generated on top of the bird-eye view (BEV) space with the flatted height axis [22, 23, 24]. There are also studies [3, 4] that project point clouds to range view and process them like images. However, due to the collapsed 3D geometry in range-view images, the relationship in point clouds cannot be fully explored. In this work, we follow the voxel-based pipeline but keep the fine-grained voxel space without height compression, as shown in Figure 2.
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
Figure 2: The framework of UVTR with multi-modality input. Given single- or multi-frame images and point clouds, we first process them in individual backbone and convert to modality-specific space $\mathbf { V } _ { I }$ and $\mathbf { V } _ { P }$ , where view transform is utilized for that of image. In voxel encoder, features are spatially interacted, and knowledge transfer is easily supported during training. Single- or multi-modality features are selected via modality switch according to different settings. Finally, transformer decoder is utilized for prediction by sampling features from the unified space $\mathbf { V } _ { U }$ with learnable positions.
|
| 32 |
+
|
| 33 |
+
Camera-based 3D Detection. Camera-based methods perform 3D detection on single- or multi-view images. With monocular image, previous approaches try to predict 3D boxes based on image features directly [25, 26, 27] or utilize the middle representation [1, 2, 5]. For multi-view input, image features are usually optimized in the constructed 3D geometry volume [28, 29]. Most recently, multi-view features are projected and merged in the frustum feature space with the aid of predicted depth [6]. Following the LiDAR-based paradigm, the frustum feature is collapsed to the BEV space, as briefly introduced in Figure 1c. However, the accuracy of the predicted depth map is much inferior to that of LiDAR, which brings semantic ambiguity to BEV space. Other recent studies try to capture geometry clues from multi-view images in an implicit manner [8, 10], which losses the chance for direct spatial interactions. In this paper, we represent image features in an explicit voxel space to alleviate the semantic ambiguity and facilitate further feature interactions, as depicted in Figure 1d.
|
| 34 |
+
|
| 35 |
+
Cross-modality Interaction. With input data from various sensors, cross-modality interaction is conducted to benefit from different inputs, e.g., modality fusion and knowledge transfer. For modality fusion, the model takes data from different sensors and conducts fusion at point- and instance-level. Specifically, point-level fusion [30, 31, 32, 7] combines features from different modalities at the early stage of the network, which enables sufficient interaction. And instance-level fusion [33, 34, 35] is usually applied at the later stage to combine object-level features. Cross-modality knowledge transfer aims to distill specific knowledge [36] across modalities in the training phase. Compared with cross-modality fusion, knowledge transfer is seldom studied for 3D object detection. A prior work is LIGA-Stereo [37] that transfers geometry-aware representations from LiDAR to stereo images via distillation. Different from [37], UVTR represents each modality in a unified manner and supports cross-modality fusion and knowledge transfer simultaneously, which further enables distillation from multi-modality or consecutive frames to the single input.
|
| 36 |
+
|
| 37 |
+
# 3 UVTR Framework
|
| 38 |
+
|
| 39 |
+
The overall framework of UVTR is relatively simple: modality-specific space is constructed to unify the representation of inputs; cross-modality interaction is designed for feature learning across spaces; and transformer decoder is introduced for object-level interaction and final prediction.
|
| 40 |
+
|
| 41 |
+
# 3.1 Modality-specific Space
|
| 42 |
+
|
| 43 |
+
Given images $\mathbf { X } _ { I }$ captured from cameras and point cloud $\mathbf { X } _ { P }$ from LiDAR, different branches are utilized to respectively generate and enhance voxel space for each modality, as presented in Figure 2.
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
Figure 3: Details in the view transform.
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
Figure 4: Details in the knowledge transfer.
|
| 50 |
+
|
| 51 |
+
Image Voxel Space. For image voxel space, a shared backbone is adopted to extract features from multi-view or multi-frame images. In this process, FPN [38] is utilized for multi-scale context aggregation that is summed to formulate the feature $\mathbf { F } _ { I } \in \mathbb { R } ^ { \bar { H } \times W \times C }$ , where $H$ and $W$ vary with FPN stages. To construct the voxel feature for images, we then transform the image feature of each view to the predefined space with the designed view transform in Figure 3. Motivated by [39, 5], we first generate the depth distribution $\mathbf { D } _ { I } \in \mathbf { \mathbb { R } } ^ { D \times H \times W }$ of each image with a single convolution as
|
| 52 |
+
|
| 53 |
+
$$
|
| 54 |
+
\mathbf D _ { I } ( u , v ) = \mathrm { S o f t m a x } ( \mathrm { C o n v } ( \mathbf F _ { I } ) ( u , v ) ) .
|
| 55 |
+
$$
|
| 56 |
+
|
| 57 |
+
Here, $( u , v )$ indicates coordinate in the image plane, and $D$ is set to 64 to represent the perception limit $6 4 m$ . It is noted that $\mathbf { D } _ { I }$ is predicted without supervision. With the predicted $\mathbf { D } _ { I }$ in $D$ depth bins, we can easily get the depth distribution of each pixel in ${ \bf F } _ { I }$ . Let $( x , y , z )$ indicates a sampling point that is generated at the center of each bin from the voxel space $\mathbf { V } _ { I }$ . The point $( u , v , d )$ in the image plane is calculated from $( x , y , z )$ with the calibration matrix $\mathbf { P }$ , where $d$ denotes the reference depth along axis $D$ of $\mathbf { D } _ { I }$ . Thus, the corresponding feature in voxel space $\mathbf { V } _ { I }$ is easily captured by
|
| 58 |
+
|
| 59 |
+
$$
|
| 60 |
+
{ \bf V } _ { I } ( x , y , z ) = { \bf D } _ { I } ( u , v , d ) \times { \bf F } _ { I } ( u , v ) ,
|
| 61 |
+
$$
|
| 62 |
+
|
| 63 |
+
where $ { \mathbf Ḋ I Ḍ } ( u , v , d )$ represents the occupancy probability of feature ${ \mathbf { F } } _ { I } ( u , v )$ in voxel $( x , y , z )$ . For the multi-frame setting with $n$ sweeps, we use the shared network for all of them and formulate $n$ voxel spaces in total. In this process, each calibration matrix $\mathbf { P }$ is aligned to the ego vehicle in the initial frame. To gather temporal cues in each voxel space, relative time offsets from the initial frame are attached along the channel axis and merged using a single convolution. Then, $n$ voxel spaces are concatenated together, and the space-level fusion is conducted with a convolutional layer. In this way, features along the temporal dimension are integrated into a unified space $\mathbf { V } _ { I }$ , which is proved to bring significant gain in Table 3. Different from methods [5, 6] for BEV space, we preserve the 3D voxel space without collapsing in $Z$ axis to avoid the aforementioned semantic ambiguity and enable further interactions. The effectiveness of the 3D voxel space is empirically studied in Table 1.
|
| 64 |
+
|
| 65 |
+
Point Voxel Space. With the accurate position, we naturally split point cloud $\mathbf { X } _ { P }$ into several regular voxels. Then, the voxel backbone in Figure 2 is utilized to process input voxels with sparse convolution [17]. To enhance multi-scale features in the generated voxel space, parallel heads with various strides are designed to extract feature $\mathbf { F } _ { P }$ from the output. In particular, several 2D convolutions are applied in each head to aggregate the spatial cues at each height. Then, multi-scale features are upsampled to a same resolution and summed together to formulate the voxel space $\mathbf { V } _ { P } \in \mathbb { R } ^ { X \times Y \times Z \times C ^ { \bullet } }$ . For multi-frame setting with $n$ sweeps, we follow previous work [24] and attach all point clouds together with relative time offsets to formulate the input $\mathbf { X } _ { P }$ .
|
| 66 |
+
|
| 67 |
+
Due to the accurate position of point cloud, the semantic ambiguity in $Z$ axis is much reduced compared with that of images. But we still preserve the 3D space $\mathbf { V } _ { P }$ without height compression for convenient cross-modality interaction in Section 3.2 and fine-grained object interaction in Section 3.3. This is also proved to bring superior experimental results in Table 1.
|
| 68 |
+
|
| 69 |
+
Voxel Encoder. In the above-generated space $\mathbf { V } _ { I }$ , features of adjacent voxels projected from different views have no connection with each other. To solve this issue and facilitate local feature interaction, the voxel encoder is proposed in each voxel space, as presented in Figure 2. Specifically, we keep the simplicity of UVTR, and only three basic convolutional blocks are applied in each voxel encoder of Figure 4. In this process, features in each space $\mathbf { V } _ { I }$ or $\mathbf { V } _ { P }$ are aggregated in both coplanar and vertical dimensions. The spatial interaction in voxel space establishes connections among adjacent features, which is proved to be essential in Table 2, especially for $\mathbf { V } _ { I }$ .
|
| 70 |
+
|
| 71 |
+
# 3.2 Cross-modality Interaction
|
| 72 |
+
|
| 73 |
+
With the unified representation in space $\mathbf { V } _ { I }$ and $\mathbf { V } _ { P }$ , interactions across modalities can be easily conducted. Given the prior that LiDAR is advanced in localization and cameras provide context for classification, the cross-modality interaction is proposed from two separate aspects, i.e., transferring geometry-aware knowledge to images in a single-modality setting and fusing context-aware features with point clouds in a multi-modality setting. In particular, knowledge transfer aims to optimize the features of the student with guidance from the teacher in the single-modality setting. Meanwhile, modality fusion is designed to better utilize all modalities in both training and inference stages.
|
| 74 |
+
|
| 75 |
+
Knowledge Transfer. Considering single modality input in the inference stage, knowledge transfer is first designed to optimize features of the student with guidance from the teacher during training, which is important in an environment that lacks multi-modality data. Due to inherent properties, the geometry structure contained in images can be further exploited with the aid of point clouds, while the rich context in images can hardly be transferred to sparse point clouds. Therefore, we mainly focus on transferring knowledge from the geometry-rich modality to the poor one in this work. Benefiting from unified feature spaces, the cross-modality transfer can be easily supported, as illustrated in Figure 4. In particular, we take features before the last ReLU layer in the voxel encoder of $\mathbf { V } _ { P }$ as the geometry-rich teacher, marked as $\mathbf { T } _ { P }$ . Meanwhile, the feature in the same position of $\mathbf { V } _ { I }$ is taken as the geometry-poor student, denoted as $\mathbf { S } _ { I }$ . If we take one object query position $( x , y , z )$ from Section 3.3, the feature distance for knowledge transfer is formulated as
|
| 76 |
+
|
| 77 |
+
$$
|
| 78 |
+
d _ { K T } = P L _ { 2 } ( { \bf T } _ { P } ( x , y , z ) , { \bf S } _ { I } ( x , y , z ) ) ,
|
| 79 |
+
$$
|
| 80 |
+
|
| 81 |
+
where $P L _ { 2 }$ represents the partial $L _ { 2 }$ distance [40]. Without bells-and-whistles, the optimization objective for knowledge transfer is averaged from $N$ object queries of transformer decoder in Section 3.3, namely $\begin{array} { r } { \mathcal { L } _ { K T } = \frac { 1 } { N } \sum _ { i } ( d _ { K T } ) } \end{array}$ . It should be noted that the whole network is optimized in an end-to-end manner, with no need for extra procedures. Given the object position in each query, we can directly minimize the object-level distance with no need to exclude background features like [37]. In a similar pipeline, the knowledge transfer is further extended to support more input streams, like multi-frame images. The proposed cross-modality knowledge transfer is flexible with input modalities and brings consistent gains over various baselines in Tables 5 and 7.
|
| 82 |
+
|
| 83 |
+
Modality Fusion. Different from the knowledge transfer, modality fusion aims to better utilize all modalities in both training and inference stages, which utilizes the complementary knowledge of point cloud and images to improve the performance and robustness. Thanks to the unified representation of each modality, feature fusion can be naturally applied. To be specific, given the processed feature space $\mathbf { V } _ { I } ^ { \prime }$ and $\mathbf { V } _ { P } ^ { \prime }$ , we first select candidate modality for final prediction via modality switch, as depicted in Figure 2. That means we support single- or multi-modality input for prediction according to different settings. If both modalities are taken, $\mathbf { V } _ { I } ^ { \prime }$ and $\mathbf { V } _ { P } ^ { \prime }$ are added together to formulate the unified voxel space $\mathbf { V } _ { U } \in \mathbb { R } ^ { X \times Y \times Z \times C }$ . In this way, both modalities are well expressed in a unified manner, which can be further fused with a single convolution. The space $\mathbf { V } _ { U }$ unifies modalities with the explicit representation, which provides an expressive space for object interactions in Section 3.3.
|
| 84 |
+
|
| 85 |
+
# 3.3 Transformer Decoder
|
| 86 |
+
|
| 87 |
+
To obtain accurate and robust predictions, the transformer decoder is utilized for further object-level interaction in the unified voxel space $\mathbf { V } _ { U }$ . We draw inspirations from deformable DETR [11] and apply reference positions to efficiently sample representative features, regardless of the spatial size of 3D voxel spaces. In particular, we first initialize $N$ object queries $\mathbf { \bar { Q } } \in \mathbb { R } ^ { N \times C }$ and generate $N$ reference points from object query embedding. Then, object queries are interacted with each other in the self-attention module and summed via skip-connection, as shown in Figure 2. Let $q$ represents a specific query in $\mathbf { Q }$ with corresponding reference point $p = ( x , y , z )$ . The process of the cross-attention module in Figure 2 is modeled as
|
| 88 |
+
|
| 89 |
+
$$
|
| 90 |
+
\mathrm { C r o s s A t t n } ( q , { \mathbf { V } } _ { U } ( p ) ) = \mathrm { D e f o r m A t t n } ( q , p , { \mathbf { V } } _ { U } ) ,
|
| 91 |
+
$$
|
| 92 |
+
|
| 93 |
+
where ${ \bf V } _ { U } ( p )$ denotes the sampled feature at $( x , y , z )$ of $\mathbf { V } _ { U }$ , and DeformAttn indicates the deformable attention in [11]. With the feed-forward network and normalization, each object query can easily interact with unified features from $\mathbf { V } _ { U }$ inside each block. There are total $M$ blocks in the transformer decoder, where $M$ is respectively set to 3 and 6 for LiDAR-based and camera-based settings. Finally, a shared MLP head is utilized for prediction according to the output of each block. And iterative box refinement [11, 8] is applied to refine 3D bounding boxes based on the predictions.
|
| 94 |
+
|
| 95 |
+
# 3.4 Optimization Objectives
|
| 96 |
+
|
| 97 |
+
Following a general paradigm in recent transformers [41, 11], Hungarian algorithm [42] is adopted for one-to-one target assignment in the training phase. Thus, a set-to-set loss $\mathcal { L } _ { D e t }$ is computed to optimize detection results, including box regression loss and classification loss. If knowledge transfer is applied, the loss $\mathcal { L } _ { K T }$ is contained to reduce cross-modality feature distance with a weight 0.01.
|
| 98 |
+
|
| 99 |
+
# 4 Experiments
|
| 100 |
+
|
| 101 |
+
In this section, we first introduce our detailed experimental setup. Then, analyses of each component are conducted on different modalities. Comparisons with several leading benchmarks on the nuScenes [43] dataset are presented in the end. More results are attached in supplementary material.
|
| 102 |
+
|
| 103 |
+
# 4.1 Experimental Setup
|
| 104 |
+
|
| 105 |
+
Dataset. nuScenes [43] dataset is a large-scale benchmark for autonomous driving, which is widely adopted for single- or multi-modality 3D object detection. It contains 700, 150, 150 scenes in the train, val, and test set, respectively. We use the synced data with 10 object categories that are captured from a 32-beam LiDAR at $2 0 \mathrm { H z }$ and six cameras in a 360-degree field of view at $1 2 \mathrm { H z }$ . Only annotations of keyframes are given at $2 \mathrm { H z }$ . Here, ablation studies are optimized on a mini 1/4 train split by default, and final models are optimized on the whole train set.
|
| 106 |
+
|
| 107 |
+
Implementation Details. In this work, we conduct experiments on different modalities with 900 object queries $\mathbf { Q }$ . Constructed voxel spaces $\mathbf { V } _ { I }$ , $\mathbf { V } _ { P }$ , and $\mathbf { V } _ { U }$ share the same shape $1 2 8 \times 1 2 8 \times Z$ , where $Z$ indicates the height of voxel space and is further investigated in Table 1. The channel number $C$ in voxel spaces and transformer decoder is set to 256. And the amount of block $M$ in the decoder is set to 3, 6, and 6 for LiDAR-based, camera-based, and fusion settings, respectively. In particular, for the LiDAR-based setting, only the branch with voxel space $\mathbf { V } _ { P }$ is kept. With grid size $0 . 1 m$ , the input point clouds are filtered in range $[ - 5 1 . 2 m , 5 1 . 2 m ]$ for $X$ and $Y$ axis with $[ - 5 . 0 m , 3 . 0 m ]$ for $Z$ axis. While for grid size $0 . 0 7 5 m$ , the range in $X$ and $Y$ axis is modified to $[ - 5 4 . 0 m , 5 4 . 0 m ]$ . The framework is trained with AdamW optimizer with an initial learning rate $2 e ^ { - 5 }$ for 20 epochs. For a camera-based setting, the network is optimized with an initial learning rate $1 e ^ { - 4 }$ for 24 epochs. As for fusion, we initialize two modality-specific branches with corresponding pretrained models and optimize the model with an initial learning rate $4 e ^ { - 5 }$ for 20 epochs. The whole framework is trained in an end-to-end manner with different modalities. More details are given in supplementary material.
|
| 108 |
+
|
| 109 |
+
# 4.2 Component-wise Analysis
|
| 110 |
+
|
| 111 |
+
In this subsection, we use a randomly sampled 1/4 split of nuScenes train set for efficient validation.
|
| 112 |
+
|
| 113 |
+
Effect of Height in Voxel Space. As elaborated in Section 3.1, the height axis $Z$ plays a vital role in voxel space, especially for camera-based $\mathbf { V } _ { I }$ . In Table 1, we validate this with different heights on both modalities. Compared with the BEV space with height 1, the increase in height contributes significantly for camera-based $\mathbf { V } _ { I }$ , which respectively improves $3 . 1 \%$ and $4 . 2 \%$ NDS with height 5 and 11. Consistent with our analysis, the gain brought by increase of height contributes less to that of LiDAR because of the accurate position, which is up to $1 \%$ NDS and $1 . 9 \%$ mAP.
|
| 114 |
+
|
| 115 |
+
Operations in Voxel Encoder. The voxel encoder in Section 3.1 aims to facilitate spatial feature interactions that are essential, especially in the camera-based manner. As presented in Table 2, camera-based network cannot converge if given no spatial interaction. For LiDAR-based network, it performs satisfactorily without the voxel encoder, which can be attributed to the established relations at the early stage. Overall, for both of them, 3D spatial interaction still plays a vital role and improves $2 . 6 \%$ and $0 . 6 \%$ NDS over 2D convolution for the camera- and LiDAR-based methods, respectively.
|
| 116 |
+
|
| 117 |
+
Effect of Multi-frame Input. To further release the potential of the designed paradigm, we input multi-frame sweeps and represent them in each voxel space, as shown in Table 3. With more input frames, networks for both modalities achieve consistent gains. The performance gap reaches $5 \%$ and $1 8 . 1 \%$ NDS for the camera- and LiDAR-based manner with 5 and 10 sweeps, respectively.
|
| 118 |
+
|
| 119 |
+
Networks for Voxel Space. In Table 4, we exploit the network for voxel space generation. It is clear that the deeper network with larger voxel space contributes more to the final result. With ResNet-101 and height 11 for space $\mathbf { V } _ { I }$ , the camera-based method attains nearly $5 \%$ gains over the baseline in both NDS and mAP. For LiDAR-based manner, the model performs slightly better with finer grid size that brings higher resolution to voxel space. It means the image-based voxel space requires strong backbones to extract expressive features, while the LiDAR-based one is less dependent on that.
|
| 120 |
+
|
| 121 |
+
Table 1: Effect of different heights $Z$ in voxel space on nuScenes val set.
|
| 122 |
+
|
| 123 |
+
<table><tr><td>modality</td><td>height</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan="4">Camera</td><td>1</td><td>31.4</td><td>24.9</td></tr><tr><td>5</td><td>34.5</td><td>27.0</td></tr><tr><td>11</td><td>35.6</td><td>28.7</td></tr><tr><td>1</td><td>62.8</td><td>54.4</td></tr><tr><td rowspan="2">LiDAR</td><td>5</td><td>63.8</td><td>55.5</td></tr><tr><td>11</td><td>63.8</td><td>56.3</td></tr></table>
|
| 124 |
+
|
| 125 |
+
Table 2: Effect of different operations in voxel encoder on nuScenes val set.
|
| 126 |
+
|
| 127 |
+
<table><tr><td>modality</td><td>type</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan="3">Camera</td><td>None</td><td>12.0</td><td>2.5</td></tr><tr><td>Conv2D</td><td>31.9</td><td>24.8</td></tr><tr><td>Conv3D</td><td>34.5</td><td>27.0</td></tr><tr><td rowspan="3">LiDAR</td><td>None</td><td>63.1</td><td>54.3</td></tr><tr><td>Conv2D</td><td>63.2</td><td>54.6</td></tr><tr><td>Conv3D</td><td>63.8</td><td>55.5</td></tr></table>
|
| 128 |
+
|
| 129 |
+
Table 3: Effect of different number of frames on nuScenes val set. sweep denotes the sweep number of multi-frame input.
|
| 130 |
+
|
| 131 |
+
<table><tr><td>modality</td><td>sweep</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan="3">Camera</td><td>1</td><td>34.5</td><td>27.0</td></tr><tr><td>3</td><td>38.0</td><td>28.7</td></tr><tr><td>5</td><td>39.5</td><td>29.4</td></tr><tr><td rowspan="2">LiDAR</td><td>1</td><td>45.7</td><td>42.8</td></tr><tr><td>10</td><td>63.8</td><td>55.5</td></tr></table>
|
| 132 |
+
|
| 133 |
+
Table 4: Effect of different models for voxel space construction on nuScenes val set. H and V denote space height and grid size.
|
| 134 |
+
|
| 135 |
+
<table><tr><td>modality</td><td>voxel net</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan="3">Camera</td><td>R50-H5</td><td>34.5</td><td>27.0</td></tr><tr><td>R50-H11</td><td>35.6</td><td>28.7</td></tr><tr><td>R101-H11</td><td>39.4</td><td>32.0</td></tr><tr><td rowspan="2">LiDAR</td><td>v0.1</td><td>63.8</td><td>55.5</td></tr><tr><td>V0.075</td><td>64.3</td><td>56.3</td></tr></table>
|
| 136 |
+
|
| 137 |
+
Table 5: Effect of different knowledge transfer settings on nuScenes val set. CS denotes multiframe camera sweeps.
|
| 138 |
+
|
| 139 |
+
<table><tr><td>student</td><td>teacher</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan="5">Camera</td><td>1</td><td>34.5</td><td>27.0</td></tr><tr><td>CS</td><td>36.3</td><td>28.1</td></tr><tr><td>LiDAR</td><td>36.4</td><td>28.2</td></tr><tr><td>Multi-mod</td><td>37.1</td><td>28.8</td></tr><tr><td></td><td>63.8</td><td>55.5</td></tr><tr><td rowspan="2">LiDAR</td><td></td><td></td><td></td></tr><tr><td>Multi-mod</td><td>64.4</td><td>56.1</td></tr></table>
|
| 140 |
+
|
| 141 |
+
Table 6: Effect of different network settings for cross-modality fusion on nuScenes val set. V denotes the split voxel grid size.
|
| 142 |
+
|
| 143 |
+
<table><tr><td>camera</td><td>lidar</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td>R50</td><td>一</td><td>34.5</td><td>27.0</td></tr><tr><td>1</td><td>V0.1</td><td>63.8</td><td>55.5</td></tr><tr><td rowspan="2">R50</td><td>V0.1</td><td>65.1</td><td>59.0</td></tr><tr><td>V0.075</td><td>65.6</td><td>60.1</td></tr><tr><td rowspan="2">R101</td><td>V0.1</td><td>65.4</td><td>59.4</td></tr><tr><td>V0.075</td><td>66.3</td><td>61.0</td></tr></table>
|
| 144 |
+
|
| 145 |
+
Knowledge Transfer. Benefiting from the unified representation, knowledge can be easily transferred across modalities. In Table 5, we compare combinations with different students and teachers. In particular, the camera-based student captures geometry-aware cues from camera sweeps or the LiDAR-based teacher, which brings up to $1 . 9 \%$ NDS gain. If coupled with context features in the multi-modality setting, the gap is enlarged to $2 . 6 \%$ NDS and $1 . 8 \%$ mAP. For the LiDAR-based student, the increase brought by knowledge transfer saturated with $0 . 6 \%$ NDS. This can be attributed to that rich context in images can not be well expressed only with sparse points during inference. Therefore, images are required as input if rich context is supplemented, like the following fusion part.
|
| 146 |
+
|
| 147 |
+
Cross-modality Fusion. In Table 6, we validate capability of UVTR with cross-modality fusion. As presented in the table, feature fusion brings significant gains over a single modality. And the LiDARbased representation dominates the final results, while the camera-based one provides supplementary context. It is reasonable because point clouds are more accurate in position and more representative in geometry expression. Cameras still provide sufficient context for better classification, which yields up to $1 . 6 \%$ NDS and $3 . 9 \%$ mAP gain compared with the LiDAR-based baseline. With finer voxel grids, the performance gap is enlarged to $2 . 5 \%$ NDS and $5 . 5 \%$ mAP.
|
| 148 |
+
|
| 149 |
+
Table 7: Comparisons of different methods with a single model on the nuScenes val set. We compare with classic methods on different modalities without test-time augmentation. † denotes the implementation from MMDetection3D [44]. L, C, CS, and M indicate the LiDAR, Camera, Camera Sweep, and Multi-modality input, respectively. L2 represents knowledge transfer from LiDAR.
|
| 150 |
+
|
| 151 |
+
<table><tr><td>Method</td><td>Backbone</td><td>NDS(%)</td><td>mAP(%)</td><td>mATE↓</td><td>mASE↓</td><td>mAOE↓</td><td>mAVE↓</td><td>mAAE↓</td></tr><tr><td colspan="9">LiDAR-based</td></tr><tr><td>CenterPoint [24] V0.1</td><td></td><td>64.9</td><td>56.6</td><td>0.291</td><td>0.252</td><td>0.324</td><td>0.284</td><td>0.189</td></tr><tr><td>UVTR-L</td><td>V0.1</td><td>66.4</td><td>59.3</td><td>0.345</td><td>0.259</td><td>0.313</td><td>0.218</td><td>0.185</td></tr><tr><td> UVTR-L</td><td>V0.075</td><td>67.7</td><td>60.9</td><td>0.334</td><td>0.257</td><td>0.300</td><td>0.204</td><td>0.182</td></tr><tr><td colspan="9">Camera-based</td></tr><tr><td>DETR3D [8]</td><td>R101</td><td>42.5</td><td>34.6</td><td>0.773</td><td>0.268</td><td>0.383</td><td>0.842</td><td>0.216</td></tr><tr><td>UVTR-C</td><td>R50</td><td>41.9</td><td>33.3</td><td>0.793</td><td>0.276</td><td>0.454</td><td>0.760</td><td>0.196</td></tr><tr><td>UVTR-C</td><td>R101</td><td>44.1</td><td>36.2</td><td>0.758</td><td>0.272</td><td>0.410</td><td>0.758</td><td>0.203</td></tr><tr><td>UVTR-CS</td><td>R50</td><td>47.2</td><td>36.2</td><td>0.756</td><td>0.276</td><td>0.399</td><td>0.467</td><td>0.189</td></tr><tr><td> UVTR-CS</td><td>R101</td><td>48.3</td><td>37.9</td><td>0.731</td><td>0.267</td><td>0.350</td><td>0.510</td><td>0.200</td></tr><tr><td>UVTR-L2C</td><td>R101</td><td>45.0</td><td>37.2</td><td>0.735</td><td>0.269</td><td>0.397</td><td>0.761</td><td>0.193</td></tr><tr><td>UVTR-L2CS</td><td>R101</td><td>48.8</td><td>39.2</td><td>0.720</td><td>0.268</td><td>0.354</td><td>0.534</td><td>0.206</td></tr><tr><td colspan="9">LiDAR+Camera</td></tr><tr><td>FUTR3D [9]</td><td>V0.075-R101</td><td>68.3</td><td>64.5</td><td>1</td><td>-</td><td>-</td><td>-</td><td>1</td></tr><tr><td>UVTR-M</td><td>V0.075-R101</td><td>70.2</td><td>65.4</td><td>0.332</td><td>0.258</td><td>0.268</td><td>0.212</td><td>0.177</td></tr></table>
|
| 152 |
+
|
| 153 |
+
Table 8: Comparisons of different distances, weather, and lighting conditions on nuScenes val set.
|
| 154 |
+
|
| 155 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Modality</td><td colspan="3">Distance: NDS(%)</td><td colspan="2">Weather: NDS(%)</td><td colspan="2">Lighting: NDS(%)</td></tr><tr><td><20m</td><td>20-30m</td><td>>30m</td><td>Sunny</td><td>Rainy</td><td>Day</td><td>Night</td></tr><tr><td>CenterPoint[ t[24]</td><td>LiDAR</td><td>74.1</td><td>62.1</td><td>34.6</td><td>64.6</td><td>64.4</td><td>65.1</td><td>40.1</td></tr><tr><td>UVTR-L</td><td>LiDAR</td><td>75.9</td><td>64.9</td><td>37.3</td><td>67.4</td><td>67.9</td><td>67.8</td><td>41.4</td></tr><tr><td>UVTR-C</td><td>Camera</td><td>52.8</td><td>39.7</td><td>20.4</td><td>43.1</td><td>48.3</td><td>44.5</td><td>23.5</td></tr><tr><td>UVTR-M</td><td>Multi-mod</td><td>77.2</td><td>68.2</td><td>38.9</td><td>69.7</td><td>72.0</td><td>70.3</td><td>42.6</td></tr></table>
|
| 156 |
+
|
| 157 |
+
# 4.3 Main Results
|
| 158 |
+
|
| 159 |
+
In this section, we first report results with various modalities that are optimized on the whole train set. Then, we give analyses of the framework robustness, including camera view drop and translational noise. Comparisons with leading methods on the nuScenes test set are presented in the end.
|
| 160 |
+
|
| 161 |
+
Results with Different Modalities. In Table 7, we carry out experiments with different modalities on the nuScenes val set. Compared with classic methods, UVTR achieves significant improvement. In particular, for LiDAR-based method, it attains $1 . 5 \%$ NDS and $2 . 7 \%$ mAP gain over CenterPoint [24]. And a finer resolution contributes better results with $6 7 . 7 \%$ NDS. For camera-based manner, UVTR-C performs better in single frame setting with $1 . 6 \%$ NDS gain over DETR3D [8]. If applied knowledge transfer in UVTR-L2C, the gap is enlarged to $2 . 5 \%$ NDS. The performance improves with more frames and attains up to $4 8 . 8 \%$ NDS. For multi-modality input, UVTR-M achieves $4 . 5 \%$ mAP gain over UVTR-L and outperforms the contemporary FUTR3D [9] with $1 . 9 \%$ NDS in a same setting.
|
| 162 |
+
|
| 163 |
+
Results in Different Conditions. In Table 8, we report the performance with different distances, weather conditions, and light situations. (1) Distance: For LiDAR-based approaches, the proposed UVTR-L achieves better performance in all situations compared with CenterPoint [24]. Equipped with both LiDAR and camera inputs in UVTR-M, the framework attains significant gains, especially in a relatively far distance $3 . 3 \%$ NDS gain in $2 0 { - } 3 0 m$ ). If the object is too far $( > 3 0 m )$ , the performance gain decreases to $1 . 6 \%$ NDS, but still much better than CenterPoint and UVTR-L. (2) Weather condition: It is clear that the proposed UVTR-L achieves significant gain compared with CenterPoint in both conditions. And additional camera input brings much better results, especially in rainy weather $4 . 1 \%$ NDS gain). (3) Light situation: Compared with that in the daylight situation, both LiDAR-based and camera-based approaches perform inferior in the dark night. Compared with
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
Figure 5: We validate the robustness of UVTR by adding two typical errors during inference. For dropped view in 5a, we randomly drop a fixed number of camera views to simulate the camera failure. For sensor noise in 5b, we randomly add translational noises in LiDAR to camera calibration matrix.
|
| 167 |
+
|
| 168 |
+
Table 9: Comparisons of leading methods with a single model on the nuScenes test set. L, C, CS, and M indicate the LiDAR, Camera, Camera Sweep, and Multi-modality input, respectively. L2 represents knowledge transfer from LiDAR. Flipping augmentation is adopted for LiDAR. It should be noted that the performance of UVTR-L2CS3 can be further improved with more than 3 sweeps.
|
| 169 |
+
|
| 170 |
+
<table><tr><td>Method</td><td>Backbone</td><td>NDS(%) 1</td><td>mAP(%)</td><td>mATE↓</td><td>mASE↓</td><td>mAOE↓</td><td></td><td>mAVE↓mAAE↓</td></tr><tr><td colspan="9">LiDAR-based</td></tr><tr><td>3DSSD [45]</td><td>Point-based</td><td>56.4</td><td>42.6</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CenterPoint [24]</td><td>V0.075</td><td>65.5</td><td>58.0</td><td>1</td><td>=</td><td>-</td><td></td><td>1</td></tr><tr><td>HotSpotNet [46]</td><td>V0.1</td><td>66.0</td><td>59.3</td><td>0.274</td><td>0.239</td><td>0.384</td><td>0.333</td><td>0.133</td></tr><tr><td>AFDetV2 [47]</td><td>V0.075</td><td>68.5</td><td>62.4</td><td>0.257</td><td>0.234</td><td>0.341</td><td>0.299</td><td>0.137</td></tr><tr><td>UVTR-L</td><td>V0.075</td><td>69.7</td><td>63.9</td><td>0.302</td><td>0.246</td><td>0.350</td><td>0.207</td><td>0.123</td></tr><tr><td colspan="9">Camera-based</td></tr><tr><td>FCOS3D [27]</td><td>R101</td><td>42.8</td><td>35.8</td><td>0.690</td><td>0.249</td><td>0.452</td><td>1.434</td><td>0.124</td></tr><tr><td>DD3D [48]</td><td>V2-99</td><td>47.7</td><td>41.8</td><td>0.572</td><td>0.249</td><td>0.368</td><td>1.014</td><td>0.124</td></tr><tr><td>DETR3D [8]</td><td>V2-99</td><td>47.9</td><td>41.2</td><td>0.641</td><td>0.255</td><td>0.394</td><td>0.845</td><td>0.133</td></tr><tr><td>BEVDet [6]</td><td>V2-99</td><td>48.8</td><td>42.4</td><td>0.524</td><td>0.242</td><td>0.373</td><td>0.950</td><td>0.148</td></tr><tr><td>PETR[10]</td><td>V2-99</td><td>50.4</td><td>44.1</td><td>0.593</td><td>0.249</td><td>0.383</td><td>0.808</td><td>0.132</td></tr><tr><td>UVTR-L2C</td><td>V2-99</td><td>52.2</td><td>45.2</td><td>0.612</td><td>0.256</td><td>0.385</td><td>0.664</td><td>0.125</td></tr><tr><td>UVTR-L2CS3</td><td>V2-99</td><td>55.1</td><td>47.2</td><td>0.577</td><td>0.253</td><td>0.391</td><td>0.508</td><td>0.123</td></tr><tr><td colspan="9">LiDAR+Camera</td></tr><tr><td>FusionPainting [49]</td><td>V0.075-R50</td><td>70.4</td><td>66.3</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>MVP [32]</td><td>V0.075-DLA34</td><td>70.5</td><td>66.4</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>PointAugmenting [50]</td><td>V0.075-DLA34</td><td>71.0</td><td>66.8</td><td>1</td><td>=</td><td></td><td></td><td>=</td></tr><tr><td>UVTR-M</td><td>V0.075-R101</td><td>71.1</td><td>67.1</td><td>0.306</td><td>0.245</td><td>0.351</td><td>0.225</td><td>0.124</td></tr></table>
|
| 171 |
+
|
| 172 |
+
CenterPoint, the proposed UVTR-L still performs better. And the camera inputs still bring significant gains in both situations, especially in a daylight environment ( $2 . 5 \%$ NDS gain).
|
| 173 |
+
|
| 174 |
+
Robustness of the Framework. To validate the robustness of UVTR, we simulate two typical sensor errors during inference in Figure 5. For loss of view, the multi-modality manner achieves well robustness in Figure 5a. Because LiDAR can still capture surrounding scenes if cameras are offline. As for the camera-based manner, the model losses inputs from dropped scenes, but the network still works well and outputs predictions within the field of view. For sensor jitter in Figure 5b, the model performs stable especially in the multi-modality setting because of the accurate perception from LiDAR. Meanwhile, knowledge transfer consistently improves performance in both settings.
|
| 175 |
+
|
| 176 |
+
Comparison with Leading Methods. In Table 9, we present comparisons with leading methods on the nuScenes test set. For LiDAR-based method, UVTR-L surpasses AFDetV2 [47] with $1 . 2 \%$ NDS and attains $6 9 . 7 \%$ NDS. For camera-based manner, with single frame, UVTR-L2C outperforms
|
| 177 |
+
|
| 178 |
+
Table 10: Comparisons of different leading tracking methods on nuScenes test set. \* indicates the method at the leaderboard with no released publication.
|
| 179 |
+
|
| 180 |
+
<table><tr><td>Method</td><td>Tracker</td><td>AMOTA(%)</td><td>AMOTP</td><td>Recall</td></tr><tr><td colspan="5">LiDAR-based</td></tr><tr><td>CenterPoint [24]</td><td>Greedy</td><td>63.8</td><td>0.555</td><td>0.675</td></tr><tr><td>UVTR-L</td><td>Greedy</td><td>67.0</td><td>0.656</td><td>0.703</td></tr><tr><td colspan="5">Camera-based</td></tr><tr><td>PolarDETR [51]</td><td>Transformer</td><td>27.3</td><td>1.185</td><td>0.404</td></tr><tr><td>BEVTrack*</td><td>Private</td><td>34.1</td><td>1.107</td><td>0.463</td></tr><tr><td>UVTR-L2CS3</td><td>Greedy</td><td>51.9</td><td>1.125</td><td>0.599</td></tr><tr><td colspan="5">LiDAR+Camera</td></tr><tr><td>EagerMOT [52]</td><td>Two-stage</td><td>67.7</td><td>0.550</td><td>0.727</td></tr><tr><td>AlphaTrack [53]</td><td>Position+Appearance</td><td>69.3</td><td>0.585</td><td>0.723</td></tr><tr><td>UVTR-M</td><td>Greedy</td><td>70.1</td><td>0.686</td><td>0.750</td></tr></table>
|
| 181 |
+
|
| 182 |
+
PETR [10] with $1 . 8 \%$ NDS and $1 . 1 \%$ mAP. With 3 camera sweeps, UVTR-L2CS3 obtains significant gain and attains $5 5 . 1 \%$ NDS, which can be further improved with more frames. For the multi-modality setting, we directly adopt pretrained models from LiDAR- and camera-based manner with simple fine-tuning without bells-and-whistles. Compared with similar approaches without special module, UVTR-M achieves $7 1 . 1 \%$ NDS and $6 7 . 1 \%$ mAP, which is on par with leading approaches.
|
| 183 |
+
|
| 184 |
+
Tracking Extension. To better illustrate the capability and generality of the proposed UVTR, we further conduct experiments on the downstream tracking task. In particular, we follow the classic tracking-by-detection paradigm and apply the simple greedy tracker in CenterPoint. The only difference lies in that we adopt threshold 0.2 and NMS to remove low quality results. As presented in Table 10, the proposed UVTR achieves leading tracking performance with the greedy tracker in different settings. Specifically, in a camera-based setting, the proposed UVTR-L2CS3 surpasses previous SOTA at the leaderboard (BEVTrack) with $1 7 . 8 \%$ AMOTA. It further proves the effectiveness and generality of the proposed cross-modality interaction in UVTR.
|
| 185 |
+
|
| 186 |
+
# 5 Discussion and Conclusion
|
| 187 |
+
|
| 188 |
+
We presented the UVTR, a conceptually simple yet effective framework for multi-modality 3D object detection. The key innovation lies in that it unifies the voxel-based representation for different modalities and facilitates multi-level interactions. In particular, it uniformly encodes inputs from different sensors in the modality-specific space to reduce the semantic ambiguity and enable spatial interaction. With the unified representation, the cross-modality interaction can be easily conducted for knowledge transfer and modality fusion. Moreover, object-level interactions in the unified space are further supported by the transformer decoder for accurate and robust detection. Experiments on the nuScenes dataset prove the effectiveness of UVTR, which attains consistent improvements over various benchmarks and surpasses previous methods with leading performance.
|
| 189 |
+
|
| 190 |
+
There still exist certain limitations in the current method. First, to construct voxel space for multi-view images, we need to process all of them in the shared image backbone, which brings computational cost, especially for multi-frame setting. In the future, we plan to explore a new manner for voxel space construction at the early stage of the network, like that of point clouds. Moreover, we construct the voxel space with a spatial resolution $1 2 8 \times 1 2 8$ . We believe a higher resolution and more image frames could bring stronger voxel space with better results, which remains to be explored.
|
| 191 |
+
|
| 192 |
+
Societal Impacts. The proposed method focuses on 3D object detection that can be used in autonomous driving. Theoretically, a better 3D detector leads to safer autonomous vehicles. However, in the short term, the current technique could not solve all the corner cases and extreme situations. It may bring potential risks to the decision process in real-world autonomous systems.
|
| 193 |
+
|
| 194 |
+
Acknowledgement. This work is supported by Shenzhen Science and Technology Program KQTD20210811090149095.
|
| 195 |
+
|
| 196 |
+
References
|
| 197 |
+
[1] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2019.
|
| 198 |
+
[2] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar $^ { + + }$ : Accurate depth for 3d object detection in autonomous driving. In International Conference on Learning Representations, 2020.
|
| 199 |
+
[3] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. arXiv:1608.07916, 2016.
|
| 200 |
+
[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2021.
|
| 201 |
+
[5] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2021.
|
| 202 |
+
[6] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv:2112.11790, 2021.
|
| 203 |
+
[7] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2022.
|
| 204 |
+
[8] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, 2022.
|
| 205 |
+
[9] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv:2203.10642, 2022.
|
| 206 |
+
[10] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv:2203.05625, 2022.
|
| 207 |
+
[11] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations, 2021.
|
| 208 |
+
[12] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv:2012.12741, 2020.
|
| 209 |
+
[13] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017.
|
| 210 |
+
[14] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 211 |
+
[15] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019.
|
| 212 |
+
[16] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 213 |
+
[17] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017.
|
| 214 |
+
[18] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2022.
|
| 215 |
+
[19] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Real-time 3d object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2018.
|
| 216 |
+
[20] Bin Yang, Ming Liang, and Raquel Urtasun. Hdnet: Exploiting hd maps for 3d object detection. In Conference on Robot Learning, 2018.
|
| 217 |
+
[21] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019.
|
| 218 |
+
[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018.
|
| 219 |
+
[23] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI Conference on Artificial Intelligence, 2021.
|
| 220 |
+
[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In IEEE Conference on Computer Vision and Pattern Recognition, 2021.
|
| 221 |
+
[25] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 222 |
+
[26] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 223 |
+
[27] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2021.
|
| 224 |
+
[28] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 225 |
+
[29] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2020.
|
| 226 |
+
[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In European Conference on Computer Vision, 2018.
|
| 227 |
+
[31] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, 2020.
|
| 228 |
+
[32] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbühl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 2021.
|
| 229 |
+
[33] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2017.
|
| 230 |
+
[34] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018.
|
| 231 |
+
[35] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, 2020.
|
| 232 |
+
[36] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015.
|
| 233 |
+
[37] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In IEEE/CVF International Conference on Computer Vision, 2021.
|
| 234 |
+
[38] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017.
|
| 235 |
+
[39] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, 2020.
|
| 236 |
+
[40] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In IEEE/CVF International Conference on Computer Vision, 2019.
|
| 237 |
+
[41] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020.
|
| 238 |
+
[42] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955.
|
| 239 |
+
[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2020.
|
| 240 |
+
[44] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.
|
| 241 |
+
[45] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In IEEE Conference on Computer Vision and Pattern Recognition, 2020.
|
| 242 |
+
[46] Qi Chen, Lin Sun, Zhixin Wang, Kui Jia, and Alan Yuille. Object as hotspots: An anchor-free 3d object detection approach via firing of hotspots. In European Conference on Computer Vision, 2020.
|
| 243 |
+
[47] Yihan Hu, Zhuangzhuang Ding, Runzhou Ge, Wenxin Shao, Li Huang, Kun Li, and Qiang Liu. Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. AAAI Conference on Artificial Intelligence, 2022.
|
| 244 |
+
[48] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In IEEE/CVF International Conference on Computer Vision, 2021.
|
| 245 |
+
[49] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In IEEE International Conference on Intelligent Transportation Systems, 2021.
|
| 246 |
+
[50] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2021.
|
| 247 |
+
[51] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv:2206.10965, 2022.
|
| 248 |
+
[52] Aleksandr Kim, Aljoša Ošep, and Laura Leal-Taixé. Eagermot: 3d multi-object tracking via sensor fusion. In IEEE International Conference on Robotics and Automation, 2021.
|
| 249 |
+
[53] Yihan Zeng, Chao Ma, Ming Zhu, Zhiming Fan, and Xiaokang Yang. Cross-modal 3d object detection and tracking for auto-driving. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2021.
|
| 250 |
+
|
| 251 |
+
# Checklist
|
| 252 |
+
|
| 253 |
+
1. For all authors...
|
| 254 |
+
|
| 255 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes]
|
| 256 |
+
(b) Did you describe the limitations of your work? [Yes] See Section 5.
|
| 257 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] See Section 5.
|
| 258 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 259 |
+
|
| 260 |
+
2. If you are including theoretical results...
|
| 261 |
+
|
| 262 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 263 |
+
|
| 264 |
+
3. If you ran experiments...
|
| 265 |
+
|
| 266 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We give detailed experimental setup in Section 4.1 and publicly release the code.
|
| 267 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 4.1 and supplementary material.
|
| 268 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No]
|
| 269 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We report them in the supplementary material. Our experiments in this paper require about 8 NVIDIA Tesla V100 GPUs.
|
| 270 |
+
|
| 271 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 272 |
+
|
| 273 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes] See Section 4.1.
|
| 274 |
+
(b) Did you mention the license of the assets? [Yes] We conduct experiments on nuScenes [43] dataset with custom non-commercial license.
|
| 275 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [No]
|
| 276 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 277 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [No] We conduct experiments only on public datasets.
|
| 278 |
+
|
| 279 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 280 |
+
|
| 281 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 282 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 283 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_content_list.json
ADDED
|
@@ -0,0 +1,1162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Unifying Voxel-based Representation with Transformer for 3D Object Detection ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
245,
|
| 8 |
+
122,
|
| 9 |
+
753,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Yanwei Li1 Yilun Chen1 Xiaojuan Qi2 Zeming Li3 Jian Sun3 Jiaya Jia1,4 ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
214,
|
| 19 |
+
224,
|
| 20 |
+
784,
|
| 21 |
+
241
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "The Chinese University of Hong Kong1 The University of Hong Kong2 MEGVII Technology3 SmartMore4 ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
259,
|
| 30 |
+
246,
|
| 31 |
+
738,
|
| 32 |
+
276
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
462,
|
| 42 |
+
311,
|
| 43 |
+
535,
|
| 44 |
+
329
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "In this work, we present a unified framework for multi-modality 3D object detection, named UVTR. The proposed method aims to unify multi-modality representations in the voxel space for accurate and robust single- or cross-modality 3D detection. To this end, the modality-specific space is first designed to represent different inputs in the voxel feature space. Different from previous work, our approach preserves the voxel space without height compression to alleviate semantic ambiguity and enable spatial connections. To make full use of the inputs from different sensors, the cross-modality interaction is then proposed, including knowledge transfer and modality fusion. In this way, geometry-aware expressions in point clouds and context-rich features in images are well utilized for better performance and robustness. The transformer decoder is applied to efficiently sample features from the unified space with learnable positions, which facilitates object-level interactions. In general, UVTR presents an early attempt to represent different modalities in a unified framework. It surpasses previous work in single- or multi-modality entries. The proposed method achieves leading performance in the nuScenes test set for both object detection and the following object tracking task. Code is made publicly available at https://github.com/dvlab-research/UVTR. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
344,
|
| 54 |
+
766,
|
| 55 |
+
579
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 Introduction ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
609,
|
| 66 |
+
310,
|
| 67 |
+
627
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Detecting 3D objects with multi-modality sensors (i.e., LiDAR and camera) is regarded as a fundamental task in real-world scenes. For accurate object detection, data from different modalities are utilized to provide complementary knowledge, like accurate positions from point clouds and rich context from images. Toward this purpose, a unified representation is essential to facilitate knowledge transfer and feature fusion across modalities. However, due to the lack of accurate depth from cameras, images can not be naturally represented in voxel space like that of point clouds. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
643,
|
| 77 |
+
825,
|
| 78 |
+
726
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "In the unified progress, several representations have been studied that can be roughly separated into input- and feature-level streams. For the first one, multi-modality data is aligned at the beginning of network. In particular, pseudo point clouds in Figure 1a are transformed from image aided by predicted depth [1, 2], while the range-view image in Figure 1b is projected from point clouds [3, 4]. Because of inaccurate depth in pseudo point clouds and collapsed 3D geometry in range-view images, the spatial structure of data is damaged, which brings inferior results. For feature-level method, a typical approach is to transform image features as frustum and then compress to BEV space [5, 6], like that in Figure 1c. However, due to ray-like trajectories in the frustum [7], height compression at each position aggregates features from various objects and thus introduces semantic ambiguity. Meanwhile, other implicit manners in contemporary work [8, 9, 10] can hardly support explicit feature interactions in 3D space and restrict further knowledge transfer. Therefore, a more unified representation is desired to bridge modality gap and facilitate interactions from multiple aspects. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
733,
|
| 88 |
+
825,
|
| 89 |
+
898
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "image",
|
| 95 |
+
"img_path": "images/ad3794a20dc5c227e5e92009fc24c2fdeb87714e65228d14ff29161f78640b60.jpg",
|
| 96 |
+
"image_caption": [
|
| 97 |
+
"Figure 1: Toy example of methods for unified representation. Compared with others, the proposed manner in 1d constructs the voxel space by sampling features from the image plane and represents multi-modalities uniformly without height-level compression in 1c that brings semantic ambiguity. "
|
| 98 |
+
],
|
| 99 |
+
"image_footnote": [],
|
| 100 |
+
"bbox": [
|
| 101 |
+
183,
|
| 102 |
+
85,
|
| 103 |
+
813,
|
| 104 |
+
321
|
| 105 |
+
],
|
| 106 |
+
"page_idx": 1
|
| 107 |
+
},
|
| 108 |
+
{
|
| 109 |
+
"type": "text",
|
| 110 |
+
"text": "In this paper, we present a simple yet effective framework to unify the voxel-based representation with transformer, called UVTR. In particular, features from images and point clouds are represented and interacted in the explicit voxel-based space. For images, we construct the voxel space by sampling features from the image plane according to predicted depth scores and geometric constraints, as briefly depicted in Figure 1d. For point clouds, the accurate position naturally allows us to associate features with voxels. Then, voxel encoder is introduced for spatial interaction that establishes the relationship among adjacent features. In this way, cross-modality interaction is naturally conducted with features in each voxel space. For object-level interaction, deformable transformer [11] is adopted as the decoder that samples specific feature for each object query with position $( x , y , z )$ in the unified voxel space, as illustrated in Figure 1d. Meanwhile, the introduction of 3D query position efficiently alleviates the semantic ambiguity brought by height compression in BEV space as analysed before. ",
|
| 111 |
+
"bbox": [
|
| 112 |
+
174,
|
| 113 |
+
400,
|
| 114 |
+
825,
|
| 115 |
+
553
|
| 116 |
+
],
|
| 117 |
+
"page_idx": 1
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"type": "text",
|
| 121 |
+
"text": "Compared with previous and even concurrent studies [8, 9], more key advances can be achieved with the proposed framework. First, the explicit voxel-based representation supports spatial interaction in 3D space and multi-frame scenes that bring significant improvements. Second, the proposed unified manner facilitates cross-modality learning and can be naturally applied for knowledge transfer and feature fusion, which further boosts the performance. Finally, data augmentation for both modalities can be directly synchronized in the voxel space without the complex aligning process [12, 7]. ",
|
| 122 |
+
"bbox": [
|
| 123 |
+
174,
|
| 124 |
+
559,
|
| 125 |
+
825,
|
| 126 |
+
642
|
| 127 |
+
],
|
| 128 |
+
"page_idx": 1
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"type": "text",
|
| 132 |
+
"text": "The overall framework, called UVTR, can be easily instantiated and improved with various imageor voxel-based backbones for single- and multi-modality 3D object detection. Extensive empirical studies are conducted in Section 4 to reveal the effect of each component. The proposed UVTR attains leading performance in various settings. For detection, it achieves $6 9 . 7 \\%$ , $5 5 . 1 \\%$ , and $7 1 . 1 \\%$ NDS on nuScenes test set with point clouds, images, and multi-modality inputs, respectively. Given naive association strategy, UVTR also achieves strong tracking results with $6 7 . 0 \\%$ , ${ \\mathfrak { s } } 1 . 9 \\%$ , and $7 0 . 1 \\%$ AMOTA on LiDAR-based, camera-based, and multi-modality setting, respectively. ",
|
| 133 |
+
"bbox": [
|
| 134 |
+
174,
|
| 135 |
+
648,
|
| 136 |
+
825,
|
| 137 |
+
746
|
| 138 |
+
],
|
| 139 |
+
"page_idx": 1
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"type": "text",
|
| 143 |
+
"text": "2 Related Work ",
|
| 144 |
+
"text_level": 1,
|
| 145 |
+
"bbox": [
|
| 146 |
+
174,
|
| 147 |
+
767,
|
| 148 |
+
321,
|
| 149 |
+
784
|
| 150 |
+
],
|
| 151 |
+
"page_idx": 1
|
| 152 |
+
},
|
| 153 |
+
{
|
| 154 |
+
"type": "text",
|
| 155 |
+
"text": "LiDAR-based 3D Detection. With point clouds captured from LiDAR, traditional methods process the irregular input and generate 3D boxes with different representations, e.g., point, voxel, and range view. Point-based detectors usually aggregate features from raw point clouds with set abstraction [13] and then predict box proposals [14, 15, 16]. For voxel-based methods, point clouds are transformed into regular grids and processed with 3D sparse convolutions [17, 18] or 2D convolutions [19, 20, 21] directly. Final predictions are usually generated on top of the bird-eye view (BEV) space with the flatted height axis [22, 23, 24]. There are also studies [3, 4] that project point clouds to range view and process them like images. However, due to the collapsed 3D geometry in range-view images, the relationship in point clouds cannot be fully explored. In this work, we follow the voxel-based pipeline but keep the fine-grained voxel space without height compression, as shown in Figure 2. ",
|
| 156 |
+
"bbox": [
|
| 157 |
+
174,
|
| 158 |
+
800,
|
| 159 |
+
825,
|
| 160 |
+
911
|
| 161 |
+
],
|
| 162 |
+
"page_idx": 1
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"type": "image",
|
| 166 |
+
"img_path": "images/313f2b03040d107f882f32c1c5a338a80da06d5aa2b71fee22c7219bd2237303.jpg",
|
| 167 |
+
"image_caption": [
|
| 168 |
+
"Figure 2: The framework of UVTR with multi-modality input. Given single- or multi-frame images and point clouds, we first process them in individual backbone and convert to modality-specific space $\\mathbf { V } _ { I }$ and $\\mathbf { V } _ { P }$ , where view transform is utilized for that of image. In voxel encoder, features are spatially interacted, and knowledge transfer is easily supported during training. Single- or multi-modality features are selected via modality switch according to different settings. Finally, transformer decoder is utilized for prediction by sampling features from the unified space $\\mathbf { V } _ { U }$ with learnable positions. "
|
| 169 |
+
],
|
| 170 |
+
"image_footnote": [],
|
| 171 |
+
"bbox": [
|
| 172 |
+
173,
|
| 173 |
+
89,
|
| 174 |
+
825,
|
| 175 |
+
262
|
| 176 |
+
],
|
| 177 |
+
"page_idx": 2
|
| 178 |
+
},
|
| 179 |
+
{
|
| 180 |
+
"type": "text",
|
| 181 |
+
"text": "",
|
| 182 |
+
"bbox": [
|
| 183 |
+
174,
|
| 184 |
+
382,
|
| 185 |
+
821,
|
| 186 |
+
411
|
| 187 |
+
],
|
| 188 |
+
"page_idx": 2
|
| 189 |
+
},
|
| 190 |
+
{
|
| 191 |
+
"type": "text",
|
| 192 |
+
"text": "Camera-based 3D Detection. Camera-based methods perform 3D detection on single- or multi-view images. With monocular image, previous approaches try to predict 3D boxes based on image features directly [25, 26, 27] or utilize the middle representation [1, 2, 5]. For multi-view input, image features are usually optimized in the constructed 3D geometry volume [28, 29]. Most recently, multi-view features are projected and merged in the frustum feature space with the aid of predicted depth [6]. Following the LiDAR-based paradigm, the frustum feature is collapsed to the BEV space, as briefly introduced in Figure 1c. However, the accuracy of the predicted depth map is much inferior to that of LiDAR, which brings semantic ambiguity to BEV space. Other recent studies try to capture geometry clues from multi-view images in an implicit manner [8, 10], which losses the chance for direct spatial interactions. In this paper, we represent image features in an explicit voxel space to alleviate the semantic ambiguity and facilitate further feature interactions, as depicted in Figure 1d. ",
|
| 193 |
+
"bbox": [
|
| 194 |
+
174,
|
| 195 |
+
416,
|
| 196 |
+
825,
|
| 197 |
+
569
|
| 198 |
+
],
|
| 199 |
+
"page_idx": 2
|
| 200 |
+
},
|
| 201 |
+
{
|
| 202 |
+
"type": "text",
|
| 203 |
+
"text": "Cross-modality Interaction. With input data from various sensors, cross-modality interaction is conducted to benefit from different inputs, e.g., modality fusion and knowledge transfer. For modality fusion, the model takes data from different sensors and conducts fusion at point- and instance-level. Specifically, point-level fusion [30, 31, 32, 7] combines features from different modalities at the early stage of the network, which enables sufficient interaction. And instance-level fusion [33, 34, 35] is usually applied at the later stage to combine object-level features. Cross-modality knowledge transfer aims to distill specific knowledge [36] across modalities in the training phase. Compared with cross-modality fusion, knowledge transfer is seldom studied for 3D object detection. A prior work is LIGA-Stereo [37] that transfers geometry-aware representations from LiDAR to stereo images via distillation. Different from [37], UVTR represents each modality in a unified manner and supports cross-modality fusion and knowledge transfer simultaneously, which further enables distillation from multi-modality or consecutive frames to the single input. ",
|
| 204 |
+
"bbox": [
|
| 205 |
+
173,
|
| 206 |
+
575,
|
| 207 |
+
825,
|
| 208 |
+
741
|
| 209 |
+
],
|
| 210 |
+
"page_idx": 2
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"type": "text",
|
| 214 |
+
"text": "3 UVTR Framework ",
|
| 215 |
+
"text_level": 1,
|
| 216 |
+
"bbox": [
|
| 217 |
+
176,
|
| 218 |
+
762,
|
| 219 |
+
362,
|
| 220 |
+
780
|
| 221 |
+
],
|
| 222 |
+
"page_idx": 2
|
| 223 |
+
},
|
| 224 |
+
{
|
| 225 |
+
"type": "text",
|
| 226 |
+
"text": "The overall framework of UVTR is relatively simple: modality-specific space is constructed to unify the representation of inputs; cross-modality interaction is designed for feature learning across spaces; and transformer decoder is introduced for object-level interaction and final prediction. ",
|
| 227 |
+
"bbox": [
|
| 228 |
+
176,
|
| 229 |
+
796,
|
| 230 |
+
825,
|
| 231 |
+
838
|
| 232 |
+
],
|
| 233 |
+
"page_idx": 2
|
| 234 |
+
},
|
| 235 |
+
{
|
| 236 |
+
"type": "text",
|
| 237 |
+
"text": "3.1 Modality-specific Space ",
|
| 238 |
+
"text_level": 1,
|
| 239 |
+
"bbox": [
|
| 240 |
+
174,
|
| 241 |
+
856,
|
| 242 |
+
377,
|
| 243 |
+
872
|
| 244 |
+
],
|
| 245 |
+
"page_idx": 2
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"type": "text",
|
| 249 |
+
"text": "Given images $\\mathbf { X } _ { I }$ captured from cameras and point cloud $\\mathbf { X } _ { P }$ from LiDAR, different branches are utilized to respectively generate and enhance voxel space for each modality, as presented in Figure 2. ",
|
| 250 |
+
"bbox": [
|
| 251 |
+
174,
|
| 252 |
+
883,
|
| 253 |
+
823,
|
| 254 |
+
911
|
| 255 |
+
],
|
| 256 |
+
"page_idx": 2
|
| 257 |
+
},
|
| 258 |
+
{
|
| 259 |
+
"type": "image",
|
| 260 |
+
"img_path": "images/8027bc0823ef309ec100839a26625ffd43a780e8a85d2d7022c4dbf3db29da0f.jpg",
|
| 261 |
+
"image_caption": [
|
| 262 |
+
"Figure 3: Details in the view transform. "
|
| 263 |
+
],
|
| 264 |
+
"image_footnote": [],
|
| 265 |
+
"bbox": [
|
| 266 |
+
179,
|
| 267 |
+
90,
|
| 268 |
+
496,
|
| 269 |
+
222
|
| 270 |
+
],
|
| 271 |
+
"page_idx": 3
|
| 272 |
+
},
|
| 273 |
+
{
|
| 274 |
+
"type": "image",
|
| 275 |
+
"img_path": "images/8000ddf8c70ebe59c5a2ace8e1b530509752fbd1cdf403dc1d3c22670503472d.jpg",
|
| 276 |
+
"image_caption": [
|
| 277 |
+
"Figure 4: Details in the knowledge transfer. "
|
| 278 |
+
],
|
| 279 |
+
"image_footnote": [],
|
| 280 |
+
"bbox": [
|
| 281 |
+
555,
|
| 282 |
+
92,
|
| 283 |
+
764,
|
| 284 |
+
218
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 3
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "Image Voxel Space. For image voxel space, a shared backbone is adopted to extract features from multi-view or multi-frame images. In this process, FPN [38] is utilized for multi-scale context aggregation that is summed to formulate the feature $\\mathbf { F } _ { I } \\in \\mathbb { R } ^ { \\bar { H } \\times W \\times C }$ , where $H$ and $W$ vary with FPN stages. To construct the voxel feature for images, we then transform the image feature of each view to the predefined space with the designed view transform in Figure 3. Motivated by [39, 5], we first generate the depth distribution $\\mathbf { D } _ { I } \\in \\mathbf { \\mathbb { R } } ^ { D \\times H \\times W }$ of each image with a single convolution as ",
|
| 291 |
+
"bbox": [
|
| 292 |
+
173,
|
| 293 |
+
267,
|
| 294 |
+
825,
|
| 295 |
+
352
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 3
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "equation",
|
| 301 |
+
"img_path": "images/f34b36a7ae4e9a541ec6d0926e8ef5cfd515c3a112dc8fb500f0fdce6ac09d72.jpg",
|
| 302 |
+
"text": "$$\n\\mathbf D _ { I } ( u , v ) = \\mathrm { S o f t m a x } ( \\mathrm { C o n v } ( \\mathbf F _ { I } ) ( u , v ) ) .\n$$",
|
| 303 |
+
"text_format": "latex",
|
| 304 |
+
"bbox": [
|
| 305 |
+
364,
|
| 306 |
+
354,
|
| 307 |
+
632,
|
| 308 |
+
372
|
| 309 |
+
],
|
| 310 |
+
"page_idx": 3
|
| 311 |
+
},
|
| 312 |
+
{
|
| 313 |
+
"type": "text",
|
| 314 |
+
"text": "Here, $( u , v )$ indicates coordinate in the image plane, and $D$ is set to 64 to represent the perception limit $6 4 m$ . It is noted that $\\mathbf { D } _ { I }$ is predicted without supervision. With the predicted $\\mathbf { D } _ { I }$ in $D$ depth bins, we can easily get the depth distribution of each pixel in ${ \\bf F } _ { I }$ . Let $( x , y , z )$ indicates a sampling point that is generated at the center of each bin from the voxel space $\\mathbf { V } _ { I }$ . The point $( u , v , d )$ in the image plane is calculated from $( x , y , z )$ with the calibration matrix $\\mathbf { P }$ , where $d$ denotes the reference depth along axis $D$ of $\\mathbf { D } _ { I }$ . Thus, the corresponding feature in voxel space $\\mathbf { V } _ { I }$ is easily captured by ",
|
| 315 |
+
"bbox": [
|
| 316 |
+
174,
|
| 317 |
+
382,
|
| 318 |
+
823,
|
| 319 |
+
467
|
| 320 |
+
],
|
| 321 |
+
"page_idx": 3
|
| 322 |
+
},
|
| 323 |
+
{
|
| 324 |
+
"type": "equation",
|
| 325 |
+
"img_path": "images/ef27f8fe7cb55dbf7ace41e65cbc4e829826875aa0b1c7b3a049c003dfa7460d.jpg",
|
| 326 |
+
"text": "$$\n{ \\bf V } _ { I } ( x , y , z ) = { \\bf D } _ { I } ( u , v , d ) \\times { \\bf F } _ { I } ( u , v ) ,\n$$",
|
| 327 |
+
"text_format": "latex",
|
| 328 |
+
"bbox": [
|
| 329 |
+
369,
|
| 330 |
+
469,
|
| 331 |
+
625,
|
| 332 |
+
487
|
| 333 |
+
],
|
| 334 |
+
"page_idx": 3
|
| 335 |
+
},
|
| 336 |
+
{
|
| 337 |
+
"type": "text",
|
| 338 |
+
"text": "where $ { \\mathbf Ḋ I Ḍ } ( u , v , d )$ represents the occupancy probability of feature ${ \\mathbf { F } } _ { I } ( u , v )$ in voxel $( x , y , z )$ . For the multi-frame setting with $n$ sweeps, we use the shared network for all of them and formulate $n$ voxel spaces in total. In this process, each calibration matrix $\\mathbf { P }$ is aligned to the ego vehicle in the initial frame. To gather temporal cues in each voxel space, relative time offsets from the initial frame are attached along the channel axis and merged using a single convolution. Then, $n$ voxel spaces are concatenated together, and the space-level fusion is conducted with a convolutional layer. In this way, features along the temporal dimension are integrated into a unified space $\\mathbf { V } _ { I }$ , which is proved to bring significant gain in Table 3. Different from methods [5, 6] for BEV space, we preserve the 3D voxel space without collapsing in $Z$ axis to avoid the aforementioned semantic ambiguity and enable further interactions. The effectiveness of the 3D voxel space is empirically studied in Table 1. ",
|
| 339 |
+
"bbox": [
|
| 340 |
+
173,
|
| 341 |
+
489,
|
| 342 |
+
825,
|
| 343 |
+
628
|
| 344 |
+
],
|
| 345 |
+
"page_idx": 3
|
| 346 |
+
},
|
| 347 |
+
{
|
| 348 |
+
"type": "text",
|
| 349 |
+
"text": "Point Voxel Space. With the accurate position, we naturally split point cloud $\\mathbf { X } _ { P }$ into several regular voxels. Then, the voxel backbone in Figure 2 is utilized to process input voxels with sparse convolution [17]. To enhance multi-scale features in the generated voxel space, parallel heads with various strides are designed to extract feature $\\mathbf { F } _ { P }$ from the output. In particular, several 2D convolutions are applied in each head to aggregate the spatial cues at each height. Then, multi-scale features are upsampled to a same resolution and summed together to formulate the voxel space $\\mathbf { V } _ { P } \\in \\mathbb { R } ^ { X \\times Y \\times Z \\times C ^ { \\bullet } }$ . For multi-frame setting with $n$ sweeps, we follow previous work [24] and attach all point clouds together with relative time offsets to formulate the input $\\mathbf { X } _ { P }$ . ",
|
| 350 |
+
"bbox": [
|
| 351 |
+
173,
|
| 352 |
+
635,
|
| 353 |
+
825,
|
| 354 |
+
746
|
| 355 |
+
],
|
| 356 |
+
"page_idx": 3
|
| 357 |
+
},
|
| 358 |
+
{
|
| 359 |
+
"type": "text",
|
| 360 |
+
"text": "Due to the accurate position of point cloud, the semantic ambiguity in $Z$ axis is much reduced compared with that of images. But we still preserve the 3D space $\\mathbf { V } _ { P }$ without height compression for convenient cross-modality interaction in Section 3.2 and fine-grained object interaction in Section 3.3. This is also proved to bring superior experimental results in Table 1. ",
|
| 361 |
+
"bbox": [
|
| 362 |
+
174,
|
| 363 |
+
752,
|
| 364 |
+
825,
|
| 365 |
+
808
|
| 366 |
+
],
|
| 367 |
+
"page_idx": 3
|
| 368 |
+
},
|
| 369 |
+
{
|
| 370 |
+
"type": "text",
|
| 371 |
+
"text": "Voxel Encoder. In the above-generated space $\\mathbf { V } _ { I }$ , features of adjacent voxels projected from different views have no connection with each other. To solve this issue and facilitate local feature interaction, the voxel encoder is proposed in each voxel space, as presented in Figure 2. Specifically, we keep the simplicity of UVTR, and only three basic convolutional blocks are applied in each voxel encoder of Figure 4. In this process, features in each space $\\mathbf { V } _ { I }$ or $\\mathbf { V } _ { P }$ are aggregated in both coplanar and vertical dimensions. The spatial interaction in voxel space establishes connections among adjacent features, which is proved to be essential in Table 2, especially for $\\mathbf { V } _ { I }$ . ",
|
| 372 |
+
"bbox": [
|
| 373 |
+
174,
|
| 374 |
+
814,
|
| 375 |
+
825,
|
| 376 |
+
911
|
| 377 |
+
],
|
| 378 |
+
"page_idx": 3
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"type": "text",
|
| 382 |
+
"text": "3.2 Cross-modality Interaction ",
|
| 383 |
+
"text_level": 1,
|
| 384 |
+
"bbox": [
|
| 385 |
+
176,
|
| 386 |
+
90,
|
| 387 |
+
401,
|
| 388 |
+
106
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 4
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "text",
|
| 394 |
+
"text": "With the unified representation in space $\\mathbf { V } _ { I }$ and $\\mathbf { V } _ { P }$ , interactions across modalities can be easily conducted. Given the prior that LiDAR is advanced in localization and cameras provide context for classification, the cross-modality interaction is proposed from two separate aspects, i.e., transferring geometry-aware knowledge to images in a single-modality setting and fusing context-aware features with point clouds in a multi-modality setting. In particular, knowledge transfer aims to optimize the features of the student with guidance from the teacher in the single-modality setting. Meanwhile, modality fusion is designed to better utilize all modalities in both training and inference stages. ",
|
| 395 |
+
"bbox": [
|
| 396 |
+
174,
|
| 397 |
+
116,
|
| 398 |
+
825,
|
| 399 |
+
213
|
| 400 |
+
],
|
| 401 |
+
"page_idx": 4
|
| 402 |
+
},
|
| 403 |
+
{
|
| 404 |
+
"type": "text",
|
| 405 |
+
"text": "Knowledge Transfer. Considering single modality input in the inference stage, knowledge transfer is first designed to optimize features of the student with guidance from the teacher during training, which is important in an environment that lacks multi-modality data. Due to inherent properties, the geometry structure contained in images can be further exploited with the aid of point clouds, while the rich context in images can hardly be transferred to sparse point clouds. Therefore, we mainly focus on transferring knowledge from the geometry-rich modality to the poor one in this work. Benefiting from unified feature spaces, the cross-modality transfer can be easily supported, as illustrated in Figure 4. In particular, we take features before the last ReLU layer in the voxel encoder of $\\mathbf { V } _ { P }$ as the geometry-rich teacher, marked as $\\mathbf { T } _ { P }$ . Meanwhile, the feature in the same position of $\\mathbf { V } _ { I }$ is taken as the geometry-poor student, denoted as $\\mathbf { S } _ { I }$ . If we take one object query position $( x , y , z )$ from Section 3.3, the feature distance for knowledge transfer is formulated as ",
|
| 406 |
+
"bbox": [
|
| 407 |
+
173,
|
| 408 |
+
219,
|
| 409 |
+
826,
|
| 410 |
+
372
|
| 411 |
+
],
|
| 412 |
+
"page_idx": 4
|
| 413 |
+
},
|
| 414 |
+
{
|
| 415 |
+
"type": "equation",
|
| 416 |
+
"img_path": "images/afd8bba3db4eabd0ddcd5558de1da13cb42d6698f7c1588116fee576eabf4c01.jpg",
|
| 417 |
+
"text": "$$\nd _ { K T } = P L _ { 2 } ( { \\bf T } _ { P } ( x , y , z ) , { \\bf S } _ { I } ( x , y , z ) ) ,\n$$",
|
| 418 |
+
"text_format": "latex",
|
| 419 |
+
"bbox": [
|
| 420 |
+
367,
|
| 421 |
+
375,
|
| 422 |
+
625,
|
| 423 |
+
392
|
| 424 |
+
],
|
| 425 |
+
"page_idx": 4
|
| 426 |
+
},
|
| 427 |
+
{
|
| 428 |
+
"type": "text",
|
| 429 |
+
"text": "where $P L _ { 2 }$ represents the partial $L _ { 2 }$ distance [40]. Without bells-and-whistles, the optimization objective for knowledge transfer is averaged from $N$ object queries of transformer decoder in Section 3.3, namely $\\begin{array} { r } { \\mathcal { L } _ { K T } = \\frac { 1 } { N } \\sum _ { i } ( d _ { K T } ) } \\end{array}$ . It should be noted that the whole network is optimized in an end-to-end manner, with no need for extra procedures. Given the object position in each query, we can directly minimize the object-level distance with no need to exclude background features like [37]. In a similar pipeline, the knowledge transfer is further extended to support more input streams, like multi-frame images. The proposed cross-modality knowledge transfer is flexible with input modalities and brings consistent gains over various baselines in Tables 5 and 7. ",
|
| 430 |
+
"bbox": [
|
| 431 |
+
173,
|
| 432 |
+
395,
|
| 433 |
+
825,
|
| 434 |
+
507
|
| 435 |
+
],
|
| 436 |
+
"page_idx": 4
|
| 437 |
+
},
|
| 438 |
+
{
|
| 439 |
+
"type": "text",
|
| 440 |
+
"text": "Modality Fusion. Different from the knowledge transfer, modality fusion aims to better utilize all modalities in both training and inference stages, which utilizes the complementary knowledge of point cloud and images to improve the performance and robustness. Thanks to the unified representation of each modality, feature fusion can be naturally applied. To be specific, given the processed feature space $\\mathbf { V } _ { I } ^ { \\prime }$ and $\\mathbf { V } _ { P } ^ { \\prime }$ , we first select candidate modality for final prediction via modality switch, as depicted in Figure 2. That means we support single- or multi-modality input for prediction according to different settings. If both modalities are taken, $\\mathbf { V } _ { I } ^ { \\prime }$ and $\\mathbf { V } _ { P } ^ { \\prime }$ are added together to formulate the unified voxel space $\\mathbf { V } _ { U } \\in \\mathbb { R } ^ { X \\times Y \\times Z \\times C }$ . In this way, both modalities are well expressed in a unified manner, which can be further fused with a single convolution. The space $\\mathbf { V } _ { U }$ unifies modalities with the explicit representation, which provides an expressive space for object interactions in Section 3.3. ",
|
| 441 |
+
"bbox": [
|
| 442 |
+
173,
|
| 443 |
+
512,
|
| 444 |
+
825,
|
| 445 |
+
654
|
| 446 |
+
],
|
| 447 |
+
"page_idx": 4
|
| 448 |
+
},
|
| 449 |
+
{
|
| 450 |
+
"type": "text",
|
| 451 |
+
"text": "3.3 Transformer Decoder ",
|
| 452 |
+
"text_level": 1,
|
| 453 |
+
"bbox": [
|
| 454 |
+
174,
|
| 455 |
+
667,
|
| 456 |
+
362,
|
| 457 |
+
683
|
| 458 |
+
],
|
| 459 |
+
"page_idx": 4
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"type": "text",
|
| 463 |
+
"text": "To obtain accurate and robust predictions, the transformer decoder is utilized for further object-level interaction in the unified voxel space $\\mathbf { V } _ { U }$ . We draw inspirations from deformable DETR [11] and apply reference positions to efficiently sample representative features, regardless of the spatial size of 3D voxel spaces. In particular, we first initialize $N$ object queries $\\mathbf { \\bar { Q } } \\in \\mathbb { R } ^ { N \\times C }$ and generate $N$ reference points from object query embedding. Then, object queries are interacted with each other in the self-attention module and summed via skip-connection, as shown in Figure 2. Let $q$ represents a specific query in $\\mathbf { Q }$ with corresponding reference point $p = ( x , y , z )$ . The process of the cross-attention module in Figure 2 is modeled as ",
|
| 464 |
+
"bbox": [
|
| 465 |
+
173,
|
| 466 |
+
693,
|
| 467 |
+
825,
|
| 468 |
+
804
|
| 469 |
+
],
|
| 470 |
+
"page_idx": 4
|
| 471 |
+
},
|
| 472 |
+
{
|
| 473 |
+
"type": "equation",
|
| 474 |
+
"img_path": "images/85b06ad5987e132013b6f7ef5e11c7f96ed8dafdcdbde921e8590c2cad8def25.jpg",
|
| 475 |
+
"text": "$$\n\\mathrm { C r o s s A t t n } ( q , { \\mathbf { V } } _ { U } ( p ) ) = \\mathrm { D e f o r m A t t n } ( q , p , { \\mathbf { V } } _ { U } ) ,\n$$",
|
| 476 |
+
"text_format": "latex",
|
| 477 |
+
"bbox": [
|
| 478 |
+
331,
|
| 479 |
+
808,
|
| 480 |
+
661,
|
| 481 |
+
825
|
| 482 |
+
],
|
| 483 |
+
"page_idx": 4
|
| 484 |
+
},
|
| 485 |
+
{
|
| 486 |
+
"type": "text",
|
| 487 |
+
"text": "where ${ \\bf V } _ { U } ( p )$ denotes the sampled feature at $( x , y , z )$ of $\\mathbf { V } _ { U }$ , and DeformAttn indicates the deformable attention in [11]. With the feed-forward network and normalization, each object query can easily interact with unified features from $\\mathbf { V } _ { U }$ inside each block. There are total $M$ blocks in the transformer decoder, where $M$ is respectively set to 3 and 6 for LiDAR-based and camera-based settings. Finally, a shared MLP head is utilized for prediction according to the output of each block. And iterative box refinement [11, 8] is applied to refine 3D bounding boxes based on the predictions. ",
|
| 488 |
+
"bbox": [
|
| 489 |
+
174,
|
| 490 |
+
827,
|
| 491 |
+
825,
|
| 492 |
+
912
|
| 493 |
+
],
|
| 494 |
+
"page_idx": 4
|
| 495 |
+
},
|
| 496 |
+
{
|
| 497 |
+
"type": "text",
|
| 498 |
+
"text": "3.4 Optimization Objectives ",
|
| 499 |
+
"text_level": 1,
|
| 500 |
+
"bbox": [
|
| 501 |
+
174,
|
| 502 |
+
90,
|
| 503 |
+
382,
|
| 504 |
+
106
|
| 505 |
+
],
|
| 506 |
+
"page_idx": 5
|
| 507 |
+
},
|
| 508 |
+
{
|
| 509 |
+
"type": "text",
|
| 510 |
+
"text": "Following a general paradigm in recent transformers [41, 11], Hungarian algorithm [42] is adopted for one-to-one target assignment in the training phase. Thus, a set-to-set loss $\\mathcal { L } _ { D e t }$ is computed to optimize detection results, including box regression loss and classification loss. If knowledge transfer is applied, the loss $\\mathcal { L } _ { K T }$ is contained to reduce cross-modality feature distance with a weight 0.01. ",
|
| 511 |
+
"bbox": [
|
| 512 |
+
174,
|
| 513 |
+
117,
|
| 514 |
+
825,
|
| 515 |
+
172
|
| 516 |
+
],
|
| 517 |
+
"page_idx": 5
|
| 518 |
+
},
|
| 519 |
+
{
|
| 520 |
+
"type": "text",
|
| 521 |
+
"text": "4 Experiments ",
|
| 522 |
+
"text_level": 1,
|
| 523 |
+
"bbox": [
|
| 524 |
+
174,
|
| 525 |
+
193,
|
| 526 |
+
312,
|
| 527 |
+
209
|
| 528 |
+
],
|
| 529 |
+
"page_idx": 5
|
| 530 |
+
},
|
| 531 |
+
{
|
| 532 |
+
"type": "text",
|
| 533 |
+
"text": "In this section, we first introduce our detailed experimental setup. Then, analyses of each component are conducted on different modalities. Comparisons with several leading benchmarks on the nuScenes [43] dataset are presented in the end. More results are attached in supplementary material. ",
|
| 534 |
+
"bbox": [
|
| 535 |
+
176,
|
| 536 |
+
224,
|
| 537 |
+
825,
|
| 538 |
+
266
|
| 539 |
+
],
|
| 540 |
+
"page_idx": 5
|
| 541 |
+
},
|
| 542 |
+
{
|
| 543 |
+
"type": "text",
|
| 544 |
+
"text": "4.1 Experimental Setup ",
|
| 545 |
+
"text_level": 1,
|
| 546 |
+
"bbox": [
|
| 547 |
+
176,
|
| 548 |
+
284,
|
| 549 |
+
351,
|
| 550 |
+
297
|
| 551 |
+
],
|
| 552 |
+
"page_idx": 5
|
| 553 |
+
},
|
| 554 |
+
{
|
| 555 |
+
"type": "text",
|
| 556 |
+
"text": "Dataset. nuScenes [43] dataset is a large-scale benchmark for autonomous driving, which is widely adopted for single- or multi-modality 3D object detection. It contains 700, 150, 150 scenes in the train, val, and test set, respectively. We use the synced data with 10 object categories that are captured from a 32-beam LiDAR at $2 0 \\mathrm { H z }$ and six cameras in a 360-degree field of view at $1 2 \\mathrm { H z }$ . Only annotations of keyframes are given at $2 \\mathrm { H z }$ . Here, ablation studies are optimized on a mini 1/4 train split by default, and final models are optimized on the whole train set. ",
|
| 557 |
+
"bbox": [
|
| 558 |
+
174,
|
| 559 |
+
309,
|
| 560 |
+
825,
|
| 561 |
+
392
|
| 562 |
+
],
|
| 563 |
+
"page_idx": 5
|
| 564 |
+
},
|
| 565 |
+
{
|
| 566 |
+
"type": "text",
|
| 567 |
+
"text": "Implementation Details. In this work, we conduct experiments on different modalities with 900 object queries $\\mathbf { Q }$ . Constructed voxel spaces $\\mathbf { V } _ { I }$ , $\\mathbf { V } _ { P }$ , and $\\mathbf { V } _ { U }$ share the same shape $1 2 8 \\times 1 2 8 \\times Z$ , where $Z$ indicates the height of voxel space and is further investigated in Table 1. The channel number $C$ in voxel spaces and transformer decoder is set to 256. And the amount of block $M$ in the decoder is set to 3, 6, and 6 for LiDAR-based, camera-based, and fusion settings, respectively. In particular, for the LiDAR-based setting, only the branch with voxel space $\\mathbf { V } _ { P }$ is kept. With grid size $0 . 1 m$ , the input point clouds are filtered in range $[ - 5 1 . 2 m , 5 1 . 2 m ]$ for $X$ and $Y$ axis with $[ - 5 . 0 m , 3 . 0 m ]$ for $Z$ axis. While for grid size $0 . 0 7 5 m$ , the range in $X$ and $Y$ axis is modified to $[ - 5 4 . 0 m , 5 4 . 0 m ]$ . The framework is trained with AdamW optimizer with an initial learning rate $2 e ^ { - 5 }$ for 20 epochs. For a camera-based setting, the network is optimized with an initial learning rate $1 e ^ { - 4 }$ for 24 epochs. As for fusion, we initialize two modality-specific branches with corresponding pretrained models and optimize the model with an initial learning rate $4 e ^ { - 5 }$ for 20 epochs. The whole framework is trained in an end-to-end manner with different modalities. More details are given in supplementary material. ",
|
| 568 |
+
"bbox": [
|
| 569 |
+
174,
|
| 570 |
+
398,
|
| 571 |
+
826,
|
| 572 |
+
578
|
| 573 |
+
],
|
| 574 |
+
"page_idx": 5
|
| 575 |
+
},
|
| 576 |
+
{
|
| 577 |
+
"type": "text",
|
| 578 |
+
"text": "4.2 Component-wise Analysis ",
|
| 579 |
+
"text_level": 1,
|
| 580 |
+
"bbox": [
|
| 581 |
+
174,
|
| 582 |
+
594,
|
| 583 |
+
393,
|
| 584 |
+
609
|
| 585 |
+
],
|
| 586 |
+
"page_idx": 5
|
| 587 |
+
},
|
| 588 |
+
{
|
| 589 |
+
"type": "text",
|
| 590 |
+
"text": "In this subsection, we use a randomly sampled 1/4 split of nuScenes train set for efficient validation. ",
|
| 591 |
+
"bbox": [
|
| 592 |
+
174,
|
| 593 |
+
621,
|
| 594 |
+
823,
|
| 595 |
+
636
|
| 596 |
+
],
|
| 597 |
+
"page_idx": 5
|
| 598 |
+
},
|
| 599 |
+
{
|
| 600 |
+
"type": "text",
|
| 601 |
+
"text": "Effect of Height in Voxel Space. As elaborated in Section 3.1, the height axis $Z$ plays a vital role in voxel space, especially for camera-based $\\mathbf { V } _ { I }$ . In Table 1, we validate this with different heights on both modalities. Compared with the BEV space with height 1, the increase in height contributes significantly for camera-based $\\mathbf { V } _ { I }$ , which respectively improves $3 . 1 \\%$ and $4 . 2 \\%$ NDS with height 5 and 11. Consistent with our analysis, the gain brought by increase of height contributes less to that of LiDAR because of the accurate position, which is up to $1 \\%$ NDS and $1 . 9 \\%$ mAP. ",
|
| 602 |
+
"bbox": [
|
| 603 |
+
174,
|
| 604 |
+
641,
|
| 605 |
+
825,
|
| 606 |
+
724
|
| 607 |
+
],
|
| 608 |
+
"page_idx": 5
|
| 609 |
+
},
|
| 610 |
+
{
|
| 611 |
+
"type": "text",
|
| 612 |
+
"text": "Operations in Voxel Encoder. The voxel encoder in Section 3.1 aims to facilitate spatial feature interactions that are essential, especially in the camera-based manner. As presented in Table 2, camera-based network cannot converge if given no spatial interaction. For LiDAR-based network, it performs satisfactorily without the voxel encoder, which can be attributed to the established relations at the early stage. Overall, for both of them, 3D spatial interaction still plays a vital role and improves $2 . 6 \\%$ and $0 . 6 \\%$ NDS over 2D convolution for the camera- and LiDAR-based methods, respectively. ",
|
| 613 |
+
"bbox": [
|
| 614 |
+
174,
|
| 615 |
+
731,
|
| 616 |
+
825,
|
| 617 |
+
814
|
| 618 |
+
],
|
| 619 |
+
"page_idx": 5
|
| 620 |
+
},
|
| 621 |
+
{
|
| 622 |
+
"type": "text",
|
| 623 |
+
"text": "Effect of Multi-frame Input. To further release the potential of the designed paradigm, we input multi-frame sweeps and represent them in each voxel space, as shown in Table 3. With more input frames, networks for both modalities achieve consistent gains. The performance gap reaches $5 \\%$ and $1 8 . 1 \\%$ NDS for the camera- and LiDAR-based manner with 5 and 10 sweeps, respectively. ",
|
| 624 |
+
"bbox": [
|
| 625 |
+
174,
|
| 626 |
+
820,
|
| 627 |
+
825,
|
| 628 |
+
876
|
| 629 |
+
],
|
| 630 |
+
"page_idx": 5
|
| 631 |
+
},
|
| 632 |
+
{
|
| 633 |
+
"type": "text",
|
| 634 |
+
"text": "Networks for Voxel Space. In Table 4, we exploit the network for voxel space generation. It is clear that the deeper network with larger voxel space contributes more to the final result. With ResNet-101 and height 11 for space $\\mathbf { V } _ { I }$ , the camera-based method attains nearly $5 \\%$ gains over the baseline in both NDS and mAP. For LiDAR-based manner, the model performs slightly better with finer grid size that brings higher resolution to voxel space. It means the image-based voxel space requires strong backbones to extract expressive features, while the LiDAR-based one is less dependent on that. ",
|
| 635 |
+
"bbox": [
|
| 636 |
+
174,
|
| 637 |
+
883,
|
| 638 |
+
821,
|
| 639 |
+
911
|
| 640 |
+
],
|
| 641 |
+
"page_idx": 5
|
| 642 |
+
},
|
| 643 |
+
{
|
| 644 |
+
"type": "table",
|
| 645 |
+
"img_path": "images/e96d996b728c55a3a003fc761d7e0d2d817f7e334578a1992ea07d603e200b86.jpg",
|
| 646 |
+
"table_caption": [
|
| 647 |
+
"Table 1: Effect of different heights $Z$ in voxel space on nuScenes val set. "
|
| 648 |
+
],
|
| 649 |
+
"table_footnote": [],
|
| 650 |
+
"table_body": "<table><tr><td>modality</td><td>height</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan=\"4\">Camera</td><td>1</td><td>31.4</td><td>24.9</td></tr><tr><td>5</td><td>34.5</td><td>27.0</td></tr><tr><td>11</td><td>35.6</td><td>28.7</td></tr><tr><td>1</td><td>62.8</td><td>54.4</td></tr><tr><td rowspan=\"2\">LiDAR</td><td>5</td><td>63.8</td><td>55.5</td></tr><tr><td>11</td><td>63.8</td><td>56.3</td></tr></table>",
|
| 651 |
+
"bbox": [
|
| 652 |
+
178,
|
| 653 |
+
125,
|
| 654 |
+
475,
|
| 655 |
+
244
|
| 656 |
+
],
|
| 657 |
+
"page_idx": 6
|
| 658 |
+
},
|
| 659 |
+
{
|
| 660 |
+
"type": "table",
|
| 661 |
+
"img_path": "images/1299c4c0c7f89b76293abad66bbd2dfba6b5f22ffa0aba2c18aeb8c2a31a436f.jpg",
|
| 662 |
+
"table_caption": [
|
| 663 |
+
"Table 2: Effect of different operations in voxel encoder on nuScenes val set. "
|
| 664 |
+
],
|
| 665 |
+
"table_footnote": [],
|
| 666 |
+
"table_body": "<table><tr><td>modality</td><td>type</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan=\"3\">Camera</td><td>None</td><td>12.0</td><td>2.5</td></tr><tr><td>Conv2D</td><td>31.9</td><td>24.8</td></tr><tr><td>Conv3D</td><td>34.5</td><td>27.0</td></tr><tr><td rowspan=\"3\">LiDAR</td><td>None</td><td>63.1</td><td>54.3</td></tr><tr><td>Conv2D</td><td>63.2</td><td>54.6</td></tr><tr><td>Conv3D</td><td>63.8</td><td>55.5</td></tr></table>",
|
| 667 |
+
"bbox": [
|
| 668 |
+
516,
|
| 669 |
+
125,
|
| 670 |
+
825,
|
| 671 |
+
244
|
| 672 |
+
],
|
| 673 |
+
"page_idx": 6
|
| 674 |
+
},
|
| 675 |
+
{
|
| 676 |
+
"type": "table",
|
| 677 |
+
"img_path": "images/65301d373343bc137c1c334523c9127e3b2a83f1c7d37cdebb83e5a50b417fc7.jpg",
|
| 678 |
+
"table_caption": [
|
| 679 |
+
"Table 3: Effect of different number of frames on nuScenes val set. sweep denotes the sweep number of multi-frame input. "
|
| 680 |
+
],
|
| 681 |
+
"table_footnote": [],
|
| 682 |
+
"table_body": "<table><tr><td>modality</td><td>sweep</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan=\"3\">Camera</td><td>1</td><td>34.5</td><td>27.0</td></tr><tr><td>3</td><td>38.0</td><td>28.7</td></tr><tr><td>5</td><td>39.5</td><td>29.4</td></tr><tr><td rowspan=\"2\">LiDAR</td><td>1</td><td>45.7</td><td>42.8</td></tr><tr><td>10</td><td>63.8</td><td>55.5</td></tr></table>",
|
| 683 |
+
"bbox": [
|
| 684 |
+
179,
|
| 685 |
+
311,
|
| 686 |
+
475,
|
| 687 |
+
417
|
| 688 |
+
],
|
| 689 |
+
"page_idx": 6
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"type": "table",
|
| 693 |
+
"img_path": "images/5820f0db74a10d9ade2f9c897b18cef60cc31aacf6ad951f292ab3e6d2e33f35.jpg",
|
| 694 |
+
"table_caption": [
|
| 695 |
+
"Table 4: Effect of different models for voxel space construction on nuScenes val set. H and V denote space height and grid size. "
|
| 696 |
+
],
|
| 697 |
+
"table_footnote": [],
|
| 698 |
+
"table_body": "<table><tr><td>modality</td><td>voxel net</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan=\"3\">Camera</td><td>R50-H5</td><td>34.5</td><td>27.0</td></tr><tr><td>R50-H11</td><td>35.6</td><td>28.7</td></tr><tr><td>R101-H11</td><td>39.4</td><td>32.0</td></tr><tr><td rowspan=\"2\">LiDAR</td><td>v0.1</td><td>63.8</td><td>55.5</td></tr><tr><td>V0.075</td><td>64.3</td><td>56.3</td></tr></table>",
|
| 699 |
+
"bbox": [
|
| 700 |
+
519,
|
| 701 |
+
311,
|
| 702 |
+
820,
|
| 703 |
+
417
|
| 704 |
+
],
|
| 705 |
+
"page_idx": 6
|
| 706 |
+
},
|
| 707 |
+
{
|
| 708 |
+
"type": "table",
|
| 709 |
+
"img_path": "images/650b98c1b92c4803c8aebab5b1d9214b7f1571026e7b4d9b97a218037bde6e1b.jpg",
|
| 710 |
+
"table_caption": [
|
| 711 |
+
"Table 5: Effect of different knowledge transfer settings on nuScenes val set. CS denotes multiframe camera sweeps. "
|
| 712 |
+
],
|
| 713 |
+
"table_footnote": [],
|
| 714 |
+
"table_body": "<table><tr><td>student</td><td>teacher</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td rowspan=\"5\">Camera</td><td>1</td><td>34.5</td><td>27.0</td></tr><tr><td>CS</td><td>36.3</td><td>28.1</td></tr><tr><td>LiDAR</td><td>36.4</td><td>28.2</td></tr><tr><td>Multi-mod</td><td>37.1</td><td>28.8</td></tr><tr><td></td><td>63.8</td><td>55.5</td></tr><tr><td rowspan=\"2\">LiDAR</td><td></td><td></td><td></td></tr><tr><td>Multi-mod</td><td>64.4</td><td>56.1</td></tr></table>",
|
| 715 |
+
"bbox": [
|
| 716 |
+
176,
|
| 717 |
+
484,
|
| 718 |
+
478,
|
| 719 |
+
606
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 6
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "table",
|
| 725 |
+
"img_path": "images/dfb52fe767b138a68ddb953f57d9a38ce8e7103f50ef612afbeb728be9cd19d0.jpg",
|
| 726 |
+
"table_caption": [
|
| 727 |
+
"Table 6: Effect of different network settings for cross-modality fusion on nuScenes val set. V denotes the split voxel grid size. "
|
| 728 |
+
],
|
| 729 |
+
"table_footnote": [],
|
| 730 |
+
"table_body": "<table><tr><td>camera</td><td>lidar</td><td>NDS(%)</td><td>mAP(%)</td></tr><tr><td>R50</td><td>一</td><td>34.5</td><td>27.0</td></tr><tr><td>1</td><td>V0.1</td><td>63.8</td><td>55.5</td></tr><tr><td rowspan=\"2\">R50</td><td>V0.1</td><td>65.1</td><td>59.0</td></tr><tr><td>V0.075</td><td>65.6</td><td>60.1</td></tr><tr><td rowspan=\"2\">R101</td><td>V0.1</td><td>65.4</td><td>59.4</td></tr><tr><td>V0.075</td><td>66.3</td><td>61.0</td></tr></table>",
|
| 731 |
+
"bbox": [
|
| 732 |
+
522,
|
| 733 |
+
484,
|
| 734 |
+
818,
|
| 735 |
+
606
|
| 736 |
+
],
|
| 737 |
+
"page_idx": 6
|
| 738 |
+
},
|
| 739 |
+
{
|
| 740 |
+
"type": "text",
|
| 741 |
+
"text": "",
|
| 742 |
+
"bbox": [
|
| 743 |
+
174,
|
| 744 |
+
635,
|
| 745 |
+
825,
|
| 746 |
+
690
|
| 747 |
+
],
|
| 748 |
+
"page_idx": 6
|
| 749 |
+
},
|
| 750 |
+
{
|
| 751 |
+
"type": "text",
|
| 752 |
+
"text": "Knowledge Transfer. Benefiting from the unified representation, knowledge can be easily transferred across modalities. In Table 5, we compare combinations with different students and teachers. In particular, the camera-based student captures geometry-aware cues from camera sweeps or the LiDAR-based teacher, which brings up to $1 . 9 \\%$ NDS gain. If coupled with context features in the multi-modality setting, the gap is enlarged to $2 . 6 \\%$ NDS and $1 . 8 \\%$ mAP. For the LiDAR-based student, the increase brought by knowledge transfer saturated with $0 . 6 \\%$ NDS. This can be attributed to that rich context in images can not be well expressed only with sparse points during inference. Therefore, images are required as input if rich context is supplemented, like the following fusion part. ",
|
| 753 |
+
"bbox": [
|
| 754 |
+
173,
|
| 755 |
+
696,
|
| 756 |
+
825,
|
| 757 |
+
809
|
| 758 |
+
],
|
| 759 |
+
"page_idx": 6
|
| 760 |
+
},
|
| 761 |
+
{
|
| 762 |
+
"type": "text",
|
| 763 |
+
"text": "Cross-modality Fusion. In Table 6, we validate capability of UVTR with cross-modality fusion. As presented in the table, feature fusion brings significant gains over a single modality. And the LiDARbased representation dominates the final results, while the camera-based one provides supplementary context. It is reasonable because point clouds are more accurate in position and more representative in geometry expression. Cameras still provide sufficient context for better classification, which yields up to $1 . 6 \\%$ NDS and $3 . 9 \\%$ mAP gain compared with the LiDAR-based baseline. With finer voxel grids, the performance gap is enlarged to $2 . 5 \\%$ NDS and $5 . 5 \\%$ mAP. ",
|
| 764 |
+
"bbox": [
|
| 765 |
+
174,
|
| 766 |
+
814,
|
| 767 |
+
825,
|
| 768 |
+
911
|
| 769 |
+
],
|
| 770 |
+
"page_idx": 6
|
| 771 |
+
},
|
| 772 |
+
{
|
| 773 |
+
"type": "table",
|
| 774 |
+
"img_path": "images/a7385a1cbc3c584a91ca3dac86f1c923caff3e6cedfbcf48fac264e225ca8ece.jpg",
|
| 775 |
+
"table_caption": [
|
| 776 |
+
"Table 7: Comparisons of different methods with a single model on the nuScenes val set. We compare with classic methods on different modalities without test-time augmentation. † denotes the implementation from MMDetection3D [44]. L, C, CS, and M indicate the LiDAR, Camera, Camera Sweep, and Multi-modality input, respectively. L2 represents knowledge transfer from LiDAR. "
|
| 777 |
+
],
|
| 778 |
+
"table_footnote": [],
|
| 779 |
+
"table_body": "<table><tr><td>Method</td><td>Backbone</td><td>NDS(%)</td><td>mAP(%)</td><td>mATE↓</td><td>mASE↓</td><td>mAOE↓</td><td>mAVE↓</td><td>mAAE↓</td></tr><tr><td colspan=\"9\">LiDAR-based</td></tr><tr><td>CenterPoint [24] V0.1</td><td></td><td>64.9</td><td>56.6</td><td>0.291</td><td>0.252</td><td>0.324</td><td>0.284</td><td>0.189</td></tr><tr><td>UVTR-L</td><td>V0.1</td><td>66.4</td><td>59.3</td><td>0.345</td><td>0.259</td><td>0.313</td><td>0.218</td><td>0.185</td></tr><tr><td> UVTR-L</td><td>V0.075</td><td>67.7</td><td>60.9</td><td>0.334</td><td>0.257</td><td>0.300</td><td>0.204</td><td>0.182</td></tr><tr><td colspan=\"9\">Camera-based</td></tr><tr><td>DETR3D [8]</td><td>R101</td><td>42.5</td><td>34.6</td><td>0.773</td><td>0.268</td><td>0.383</td><td>0.842</td><td>0.216</td></tr><tr><td>UVTR-C</td><td>R50</td><td>41.9</td><td>33.3</td><td>0.793</td><td>0.276</td><td>0.454</td><td>0.760</td><td>0.196</td></tr><tr><td>UVTR-C</td><td>R101</td><td>44.1</td><td>36.2</td><td>0.758</td><td>0.272</td><td>0.410</td><td>0.758</td><td>0.203</td></tr><tr><td>UVTR-CS</td><td>R50</td><td>47.2</td><td>36.2</td><td>0.756</td><td>0.276</td><td>0.399</td><td>0.467</td><td>0.189</td></tr><tr><td> UVTR-CS</td><td>R101</td><td>48.3</td><td>37.9</td><td>0.731</td><td>0.267</td><td>0.350</td><td>0.510</td><td>0.200</td></tr><tr><td>UVTR-L2C</td><td>R101</td><td>45.0</td><td>37.2</td><td>0.735</td><td>0.269</td><td>0.397</td><td>0.761</td><td>0.193</td></tr><tr><td>UVTR-L2CS</td><td>R101</td><td>48.8</td><td>39.2</td><td>0.720</td><td>0.268</td><td>0.354</td><td>0.534</td><td>0.206</td></tr><tr><td colspan=\"9\">LiDAR+Camera</td></tr><tr><td>FUTR3D [9]</td><td>V0.075-R101</td><td>68.3</td><td>64.5</td><td>1</td><td>-</td><td>-</td><td>-</td><td>1</td></tr><tr><td>UVTR-M</td><td>V0.075-R101</td><td>70.2</td><td>65.4</td><td>0.332</td><td>0.258</td><td>0.268</td><td>0.212</td><td>0.177</td></tr></table>",
|
| 780 |
+
"bbox": [
|
| 781 |
+
176,
|
| 782 |
+
152,
|
| 783 |
+
821,
|
| 784 |
+
412
|
| 785 |
+
],
|
| 786 |
+
"page_idx": 7
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"type": "table",
|
| 790 |
+
"img_path": "images/eda2b1fe680a9b42c498c65f3b39b03fba05e87279c671c421256951fce1fd4d.jpg",
|
| 791 |
+
"table_caption": [
|
| 792 |
+
"Table 8: Comparisons of different distances, weather, and lighting conditions on nuScenes val set. "
|
| 793 |
+
],
|
| 794 |
+
"table_footnote": [],
|
| 795 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Modality</td><td colspan=\"3\">Distance: NDS(%)</td><td colspan=\"2\">Weather: NDS(%)</td><td colspan=\"2\">Lighting: NDS(%)</td></tr><tr><td><20m</td><td>20-30m</td><td>>30m</td><td>Sunny</td><td>Rainy</td><td>Day</td><td>Night</td></tr><tr><td>CenterPoint[ t[24]</td><td>LiDAR</td><td>74.1</td><td>62.1</td><td>34.6</td><td>64.6</td><td>64.4</td><td>65.1</td><td>40.1</td></tr><tr><td>UVTR-L</td><td>LiDAR</td><td>75.9</td><td>64.9</td><td>37.3</td><td>67.4</td><td>67.9</td><td>67.8</td><td>41.4</td></tr><tr><td>UVTR-C</td><td>Camera</td><td>52.8</td><td>39.7</td><td>20.4</td><td>43.1</td><td>48.3</td><td>44.5</td><td>23.5</td></tr><tr><td>UVTR-M</td><td>Multi-mod</td><td>77.2</td><td>68.2</td><td>38.9</td><td>69.7</td><td>72.0</td><td>70.3</td><td>42.6</td></tr></table>",
|
| 796 |
+
"bbox": [
|
| 797 |
+
176,
|
| 798 |
+
453,
|
| 799 |
+
820,
|
| 800 |
+
550
|
| 801 |
+
],
|
| 802 |
+
"page_idx": 7
|
| 803 |
+
},
|
| 804 |
+
{
|
| 805 |
+
"type": "text",
|
| 806 |
+
"text": "4.3 Main Results ",
|
| 807 |
+
"text_level": 1,
|
| 808 |
+
"bbox": [
|
| 809 |
+
174,
|
| 810 |
+
580,
|
| 811 |
+
305,
|
| 812 |
+
594
|
| 813 |
+
],
|
| 814 |
+
"page_idx": 7
|
| 815 |
+
},
|
| 816 |
+
{
|
| 817 |
+
"type": "text",
|
| 818 |
+
"text": "In this section, we first report results with various modalities that are optimized on the whole train set. Then, we give analyses of the framework robustness, including camera view drop and translational noise. Comparisons with leading methods on the nuScenes test set are presented in the end. ",
|
| 819 |
+
"bbox": [
|
| 820 |
+
174,
|
| 821 |
+
607,
|
| 822 |
+
825,
|
| 823 |
+
650
|
| 824 |
+
],
|
| 825 |
+
"page_idx": 7
|
| 826 |
+
},
|
| 827 |
+
{
|
| 828 |
+
"type": "text",
|
| 829 |
+
"text": "Results with Different Modalities. In Table 7, we carry out experiments with different modalities on the nuScenes val set. Compared with classic methods, UVTR achieves significant improvement. In particular, for LiDAR-based method, it attains $1 . 5 \\%$ NDS and $2 . 7 \\%$ mAP gain over CenterPoint [24]. And a finer resolution contributes better results with $6 7 . 7 \\%$ NDS. For camera-based manner, UVTR-C performs better in single frame setting with $1 . 6 \\%$ NDS gain over DETR3D [8]. If applied knowledge transfer in UVTR-L2C, the gap is enlarged to $2 . 5 \\%$ NDS. The performance improves with more frames and attains up to $4 8 . 8 \\%$ NDS. For multi-modality input, UVTR-M achieves $4 . 5 \\%$ mAP gain over UVTR-L and outperforms the contemporary FUTR3D [9] with $1 . 9 \\%$ NDS in a same setting. ",
|
| 830 |
+
"bbox": [
|
| 831 |
+
173,
|
| 832 |
+
655,
|
| 833 |
+
825,
|
| 834 |
+
767
|
| 835 |
+
],
|
| 836 |
+
"page_idx": 7
|
| 837 |
+
},
|
| 838 |
+
{
|
| 839 |
+
"type": "text",
|
| 840 |
+
"text": "Results in Different Conditions. In Table 8, we report the performance with different distances, weather conditions, and light situations. (1) Distance: For LiDAR-based approaches, the proposed UVTR-L achieves better performance in all situations compared with CenterPoint [24]. Equipped with both LiDAR and camera inputs in UVTR-M, the framework attains significant gains, especially in a relatively far distance $3 . 3 \\%$ NDS gain in $2 0 { - } 3 0 m$ ). If the object is too far $( > 3 0 m )$ , the performance gain decreases to $1 . 6 \\%$ NDS, but still much better than CenterPoint and UVTR-L. (2) Weather condition: It is clear that the proposed UVTR-L achieves significant gain compared with CenterPoint in both conditions. And additional camera input brings much better results, especially in rainy weather $4 . 1 \\%$ NDS gain). (3) Light situation: Compared with that in the daylight situation, both LiDAR-based and camera-based approaches perform inferior in the dark night. Compared with ",
|
| 841 |
+
"bbox": [
|
| 842 |
+
174,
|
| 843 |
+
772,
|
| 844 |
+
825,
|
| 845 |
+
911
|
| 846 |
+
],
|
| 847 |
+
"page_idx": 7
|
| 848 |
+
},
|
| 849 |
+
{
|
| 850 |
+
"type": "image",
|
| 851 |
+
"img_path": "images/e113a22a13b9d3491ead6bb6db766ef671282b7155c21af54c12dee852cd77d9.jpg",
|
| 852 |
+
"image_caption": [
|
| 853 |
+
"Figure 5: We validate the robustness of UVTR by adding two typical errors during inference. For dropped view in 5a, we randomly drop a fixed number of camera views to simulate the camera failure. For sensor noise in 5b, we randomly add translational noises in LiDAR to camera calibration matrix. "
|
| 854 |
+
],
|
| 855 |
+
"image_footnote": [],
|
| 856 |
+
"bbox": [
|
| 857 |
+
173,
|
| 858 |
+
93,
|
| 859 |
+
812,
|
| 860 |
+
273
|
| 861 |
+
],
|
| 862 |
+
"page_idx": 8
|
| 863 |
+
},
|
| 864 |
+
{
|
| 865 |
+
"type": "table",
|
| 866 |
+
"img_path": "images/a5b830ed933a98a6db8a0ebbff46c77a6670d6c5172104771bd9a4341ce86874.jpg",
|
| 867 |
+
"table_caption": [
|
| 868 |
+
"Table 9: Comparisons of leading methods with a single model on the nuScenes test set. L, C, CS, and M indicate the LiDAR, Camera, Camera Sweep, and Multi-modality input, respectively. L2 represents knowledge transfer from LiDAR. Flipping augmentation is adopted for LiDAR. It should be noted that the performance of UVTR-L2CS3 can be further improved with more than 3 sweeps. "
|
| 869 |
+
],
|
| 870 |
+
"table_footnote": [],
|
| 871 |
+
"table_body": "<table><tr><td>Method</td><td>Backbone</td><td>NDS(%) 1</td><td>mAP(%)</td><td>mATE↓</td><td>mASE↓</td><td>mAOE↓</td><td></td><td>mAVE↓mAAE↓</td></tr><tr><td colspan=\"9\">LiDAR-based</td></tr><tr><td>3DSSD [45]</td><td>Point-based</td><td>56.4</td><td>42.6</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>CenterPoint [24]</td><td>V0.075</td><td>65.5</td><td>58.0</td><td>1</td><td>=</td><td>-</td><td></td><td>1</td></tr><tr><td>HotSpotNet [46]</td><td>V0.1</td><td>66.0</td><td>59.3</td><td>0.274</td><td>0.239</td><td>0.384</td><td>0.333</td><td>0.133</td></tr><tr><td>AFDetV2 [47]</td><td>V0.075</td><td>68.5</td><td>62.4</td><td>0.257</td><td>0.234</td><td>0.341</td><td>0.299</td><td>0.137</td></tr><tr><td>UVTR-L</td><td>V0.075</td><td>69.7</td><td>63.9</td><td>0.302</td><td>0.246</td><td>0.350</td><td>0.207</td><td>0.123</td></tr><tr><td colspan=\"9\">Camera-based</td></tr><tr><td>FCOS3D [27]</td><td>R101</td><td>42.8</td><td>35.8</td><td>0.690</td><td>0.249</td><td>0.452</td><td>1.434</td><td>0.124</td></tr><tr><td>DD3D [48]</td><td>V2-99</td><td>47.7</td><td>41.8</td><td>0.572</td><td>0.249</td><td>0.368</td><td>1.014</td><td>0.124</td></tr><tr><td>DETR3D [8]</td><td>V2-99</td><td>47.9</td><td>41.2</td><td>0.641</td><td>0.255</td><td>0.394</td><td>0.845</td><td>0.133</td></tr><tr><td>BEVDet [6]</td><td>V2-99</td><td>48.8</td><td>42.4</td><td>0.524</td><td>0.242</td><td>0.373</td><td>0.950</td><td>0.148</td></tr><tr><td>PETR[10]</td><td>V2-99</td><td>50.4</td><td>44.1</td><td>0.593</td><td>0.249</td><td>0.383</td><td>0.808</td><td>0.132</td></tr><tr><td>UVTR-L2C</td><td>V2-99</td><td>52.2</td><td>45.2</td><td>0.612</td><td>0.256</td><td>0.385</td><td>0.664</td><td>0.125</td></tr><tr><td>UVTR-L2CS3</td><td>V2-99</td><td>55.1</td><td>47.2</td><td>0.577</td><td>0.253</td><td>0.391</td><td>0.508</td><td>0.123</td></tr><tr><td colspan=\"9\">LiDAR+Camera</td></tr><tr><td>FusionPainting [49]</td><td>V0.075-R50</td><td>70.4</td><td>66.3</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>MVP [32]</td><td>V0.075-DLA34</td><td>70.5</td><td>66.4</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>PointAugmenting [50]</td><td>V0.075-DLA34</td><td>71.0</td><td>66.8</td><td>1</td><td>=</td><td></td><td></td><td>=</td></tr><tr><td>UVTR-M</td><td>V0.075-R101</td><td>71.1</td><td>67.1</td><td>0.306</td><td>0.245</td><td>0.351</td><td>0.225</td><td>0.124</td></tr></table>",
|
| 872 |
+
"bbox": [
|
| 873 |
+
176,
|
| 874 |
+
407,
|
| 875 |
+
823,
|
| 876 |
+
699
|
| 877 |
+
],
|
| 878 |
+
"page_idx": 8
|
| 879 |
+
},
|
| 880 |
+
{
|
| 881 |
+
"type": "text",
|
| 882 |
+
"text": "CenterPoint, the proposed UVTR-L still performs better. And the camera inputs still bring significant gains in both situations, especially in a daylight environment ( $2 . 5 \\%$ NDS gain). ",
|
| 883 |
+
"bbox": [
|
| 884 |
+
174,
|
| 885 |
+
731,
|
| 886 |
+
823,
|
| 887 |
+
760
|
| 888 |
+
],
|
| 889 |
+
"page_idx": 8
|
| 890 |
+
},
|
| 891 |
+
{
|
| 892 |
+
"type": "text",
|
| 893 |
+
"text": "Robustness of the Framework. To validate the robustness of UVTR, we simulate two typical sensor errors during inference in Figure 5. For loss of view, the multi-modality manner achieves well robustness in Figure 5a. Because LiDAR can still capture surrounding scenes if cameras are offline. As for the camera-based manner, the model losses inputs from dropped scenes, but the network still works well and outputs predictions within the field of view. For sensor jitter in Figure 5b, the model performs stable especially in the multi-modality setting because of the accurate perception from LiDAR. Meanwhile, knowledge transfer consistently improves performance in both settings. ",
|
| 894 |
+
"bbox": [
|
| 895 |
+
174,
|
| 896 |
+
765,
|
| 897 |
+
825,
|
| 898 |
+
863
|
| 899 |
+
],
|
| 900 |
+
"page_idx": 8
|
| 901 |
+
},
|
| 902 |
+
{
|
| 903 |
+
"type": "text",
|
| 904 |
+
"text": "Comparison with Leading Methods. In Table 9, we present comparisons with leading methods on the nuScenes test set. For LiDAR-based method, UVTR-L surpasses AFDetV2 [47] with $1 . 2 \\%$ NDS and attains $6 9 . 7 \\%$ NDS. For camera-based manner, with single frame, UVTR-L2C outperforms ",
|
| 905 |
+
"bbox": [
|
| 906 |
+
176,
|
| 907 |
+
869,
|
| 908 |
+
825,
|
| 909 |
+
911
|
| 910 |
+
],
|
| 911 |
+
"page_idx": 8
|
| 912 |
+
},
|
| 913 |
+
{
|
| 914 |
+
"type": "table",
|
| 915 |
+
"img_path": "images/0ef9546cac803dab18ed69cc2265055cb88805eedf47d17cadf4c2141023ac90.jpg",
|
| 916 |
+
"table_caption": [
|
| 917 |
+
"Table 10: Comparisons of different leading tracking methods on nuScenes test set. \\* indicates the method at the leaderboard with no released publication. "
|
| 918 |
+
],
|
| 919 |
+
"table_footnote": [],
|
| 920 |
+
"table_body": "<table><tr><td>Method</td><td>Tracker</td><td>AMOTA(%)</td><td>AMOTP</td><td>Recall</td></tr><tr><td colspan=\"5\">LiDAR-based</td></tr><tr><td>CenterPoint [24]</td><td>Greedy</td><td>63.8</td><td>0.555</td><td>0.675</td></tr><tr><td>UVTR-L</td><td>Greedy</td><td>67.0</td><td>0.656</td><td>0.703</td></tr><tr><td colspan=\"5\">Camera-based</td></tr><tr><td>PolarDETR [51]</td><td>Transformer</td><td>27.3</td><td>1.185</td><td>0.404</td></tr><tr><td>BEVTrack*</td><td>Private</td><td>34.1</td><td>1.107</td><td>0.463</td></tr><tr><td>UVTR-L2CS3</td><td>Greedy</td><td>51.9</td><td>1.125</td><td>0.599</td></tr><tr><td colspan=\"5\">LiDAR+Camera</td></tr><tr><td>EagerMOT [52]</td><td>Two-stage</td><td>67.7</td><td>0.550</td><td>0.727</td></tr><tr><td>AlphaTrack [53]</td><td>Position+Appearance</td><td>69.3</td><td>0.585</td><td>0.723</td></tr><tr><td>UVTR-M</td><td>Greedy</td><td>70.1</td><td>0.686</td><td>0.750</td></tr></table>",
|
| 921 |
+
"bbox": [
|
| 922 |
+
232,
|
| 923 |
+
125,
|
| 924 |
+
766,
|
| 925 |
+
340
|
| 926 |
+
],
|
| 927 |
+
"page_idx": 9
|
| 928 |
+
},
|
| 929 |
+
{
|
| 930 |
+
"type": "text",
|
| 931 |
+
"text": "PETR [10] with $1 . 8 \\%$ NDS and $1 . 1 \\%$ mAP. With 3 camera sweeps, UVTR-L2CS3 obtains significant gain and attains $5 5 . 1 \\%$ NDS, which can be further improved with more frames. For the multi-modality setting, we directly adopt pretrained models from LiDAR- and camera-based manner with simple fine-tuning without bells-and-whistles. Compared with similar approaches without special module, UVTR-M achieves $7 1 . 1 \\%$ NDS and $6 7 . 1 \\%$ mAP, which is on par with leading approaches. ",
|
| 932 |
+
"bbox": [
|
| 933 |
+
173,
|
| 934 |
+
363,
|
| 935 |
+
825,
|
| 936 |
+
434
|
| 937 |
+
],
|
| 938 |
+
"page_idx": 9
|
| 939 |
+
},
|
| 940 |
+
{
|
| 941 |
+
"type": "text",
|
| 942 |
+
"text": "Tracking Extension. To better illustrate the capability and generality of the proposed UVTR, we further conduct experiments on the downstream tracking task. In particular, we follow the classic tracking-by-detection paradigm and apply the simple greedy tracker in CenterPoint. The only difference lies in that we adopt threshold 0.2 and NMS to remove low quality results. As presented in Table 10, the proposed UVTR achieves leading tracking performance with the greedy tracker in different settings. Specifically, in a camera-based setting, the proposed UVTR-L2CS3 surpasses previous SOTA at the leaderboard (BEVTrack) with $1 7 . 8 \\%$ AMOTA. It further proves the effectiveness and generality of the proposed cross-modality interaction in UVTR. ",
|
| 943 |
+
"bbox": [
|
| 944 |
+
173,
|
| 945 |
+
439,
|
| 946 |
+
825,
|
| 947 |
+
551
|
| 948 |
+
],
|
| 949 |
+
"page_idx": 9
|
| 950 |
+
},
|
| 951 |
+
{
|
| 952 |
+
"type": "text",
|
| 953 |
+
"text": "5 Discussion and Conclusion ",
|
| 954 |
+
"text_level": 1,
|
| 955 |
+
"bbox": [
|
| 956 |
+
176,
|
| 957 |
+
569,
|
| 958 |
+
428,
|
| 959 |
+
587
|
| 960 |
+
],
|
| 961 |
+
"page_idx": 9
|
| 962 |
+
},
|
| 963 |
+
{
|
| 964 |
+
"type": "text",
|
| 965 |
+
"text": "We presented the UVTR, a conceptually simple yet effective framework for multi-modality 3D object detection. The key innovation lies in that it unifies the voxel-based representation for different modalities and facilitates multi-level interactions. In particular, it uniformly encodes inputs from different sensors in the modality-specific space to reduce the semantic ambiguity and enable spatial interaction. With the unified representation, the cross-modality interaction can be easily conducted for knowledge transfer and modality fusion. Moreover, object-level interactions in the unified space are further supported by the transformer decoder for accurate and robust detection. Experiments on the nuScenes dataset prove the effectiveness of UVTR, which attains consistent improvements over various benchmarks and surpasses previous methods with leading performance. ",
|
| 966 |
+
"bbox": [
|
| 967 |
+
174,
|
| 968 |
+
599,
|
| 969 |
+
825,
|
| 970 |
+
726
|
| 971 |
+
],
|
| 972 |
+
"page_idx": 9
|
| 973 |
+
},
|
| 974 |
+
{
|
| 975 |
+
"type": "text",
|
| 976 |
+
"text": "There still exist certain limitations in the current method. First, to construct voxel space for multi-view images, we need to process all of them in the shared image backbone, which brings computational cost, especially for multi-frame setting. In the future, we plan to explore a new manner for voxel space construction at the early stage of the network, like that of point clouds. Moreover, we construct the voxel space with a spatial resolution $1 2 8 \\times 1 2 8$ . We believe a higher resolution and more image frames could bring stronger voxel space with better results, which remains to be explored. ",
|
| 977 |
+
"bbox": [
|
| 978 |
+
174,
|
| 979 |
+
731,
|
| 980 |
+
825,
|
| 981 |
+
815
|
| 982 |
+
],
|
| 983 |
+
"page_idx": 9
|
| 984 |
+
},
|
| 985 |
+
{
|
| 986 |
+
"type": "text",
|
| 987 |
+
"text": "Societal Impacts. The proposed method focuses on 3D object detection that can be used in autonomous driving. Theoretically, a better 3D detector leads to safer autonomous vehicles. However, in the short term, the current technique could not solve all the corner cases and extreme situations. It may bring potential risks to the decision process in real-world autonomous systems. ",
|
| 988 |
+
"bbox": [
|
| 989 |
+
174,
|
| 990 |
+
820,
|
| 991 |
+
825,
|
| 992 |
+
877
|
| 993 |
+
],
|
| 994 |
+
"page_idx": 9
|
| 995 |
+
},
|
| 996 |
+
{
|
| 997 |
+
"type": "text",
|
| 998 |
+
"text": "Acknowledgement. This work is supported by Shenzhen Science and Technology Program KQTD20210811090149095. ",
|
| 999 |
+
"bbox": [
|
| 1000 |
+
173,
|
| 1001 |
+
882,
|
| 1002 |
+
823,
|
| 1003 |
+
911
|
| 1004 |
+
],
|
| 1005 |
+
"page_idx": 9
|
| 1006 |
+
},
|
| 1007 |
+
{
|
| 1008 |
+
"type": "text",
|
| 1009 |
+
"text": "References \n[1] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. \n[2] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudo-lidar $^ { + + }$ : Accurate depth for 3d object detection in autonomous driving. In International Conference on Learning Representations, 2020. \n[3] Bo Li, Tianlei Zhang, and Tian Xia. Vehicle detection from 3d lidar using fully convolutional network. arXiv:1608.07916, 2016. \n[4] Lue Fan, Xuan Xiong, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Rangedet: In defense of range view for lidar-based 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2021. \n[5] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. \n[6] Junjie Huang, Guan Huang, Zheng Zhu, and Dalong Du. Bevdet: High-performance multi-camera 3d object detection in bird-eye-view. arXiv:2112.11790, 2021. \n[7] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. \n[8] Yue Wang, Vitor Campagnolo Guizilini, Tianyuan Zhang, Yilun Wang, Hang Zhao, and Justin Solomon. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on Robot Learning, 2022. \n[9] Xuanyao Chen, Tianyuan Zhang, Yue Wang, Yilun Wang, and Hang Zhao. Futr3d: A unified sensor fusion framework for 3d detection. arXiv:2203.10642, 2022. \n[10] Yingfei Liu, Tiancai Wang, Xiangyu Zhang, and Jian Sun. Petr: Position embedding transformation for multi-view 3d object detection. arXiv:2203.05625, 2022. \n[11] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. International Conference on Learning Representations, 2021. \n[12] Wenwei Zhang, Zhe Wang, and Chen Change Loy. Exploring data augmentation for multi-modality 3d object detection. arXiv:2012.12741, 2020. \n[13] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems, 2017. \n[14] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In IEEE/CVF International Conference on Computer Vision, 2019. \n[15] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. \n[16] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In IEEE/CVF International Conference on Computer Vision, 2019. \n[17] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017. \n[18] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. \n[19] Bin Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Real-time 3d object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2018. \n[20] Bin Yang, Ming Liang, and Raquel Urtasun. Hdnet: Exploiting hd maps for 3d object detection. In Conference on Robot Learning, 2018. \n[21] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. \n[22] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 2018. \n[23] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI Conference on Artificial Intelligence, 2021. \n[24] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. \n[25] Garrick Brazil and Xiaoming Liu. M3d-rpn: Monocular 3d region proposal network for object detection. In IEEE/CVF International Conference on Computer Vision, 2019. \n[26] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2019. \n[27] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. Fcos3d: Fully convolutional one-stage monocular 3d object detection. In IEEE/CVF International Conference on Computer Vision, 2021. \n[28] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In IEEE/CVF International Conference on Computer Vision, 2019. \n[29] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. \n[30] Ming Liang, Bin Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In European Conference on Computer Vision, 2018. \n[31] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. Epnet: Enhancing point features with image semantics for 3d object detection. In European Conference on Computer Vision, 2020. \n[32] Tianwei Yin, Xingyi Zhou, and Philipp Krähenbühl. Multimodal virtual point 3d detection. Advances in Neural Information Processing Systems, 2021. \n[33] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. \n[34] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven L Waslander. Joint 3d proposal generation and object detection from view aggregation. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2018. \n[35] Jin Hyeok Yoo, Yecheol Kim, Jisong Kim, and Jun Won Choi. 3d-cvf: Generating joint camera and lidar features using cross-view spatial feature fusion for 3d object detection. In European Conference on Computer Vision, 2020. \n[36] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. Distilling the knowledge in a neural network. arXiv:1503.02531, 2015. \n[37] Xiaoyang Guo, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Liga-stereo: Learning lidar geometry aware representations for stereo-based 3d detector. In IEEE/CVF International Conference on Computer Vision, 2021. \n[38] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. \n[39] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In European Conference on Computer Vision, 2020. \n[40] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, Nojun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In IEEE/CVF International Conference on Computer Vision, 2019. \n[41] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, 2020. \n[42] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. \n[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. \n[44] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020. \n[45] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. \n[46] Qi Chen, Lin Sun, Zhixin Wang, Kui Jia, and Alan Yuille. Object as hotspots: An anchor-free 3d object detection approach via firing of hotspots. In European Conference on Computer Vision, 2020. \n[47] Yihan Hu, Zhuangzhuang Ding, Runzhou Ge, Wenxin Shao, Li Huang, Kun Li, and Qiang Liu. Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. AAAI Conference on Artificial Intelligence, 2022. \n[48] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? In IEEE/CVF International Conference on Computer Vision, 2021. \n[49] Shaoqing Xu, Dingfu Zhou, Jin Fang, Junbo Yin, Zhou Bin, and Liangjun Zhang. Fusionpainting: Multimodal fusion with adaptive attention for 3d object detection. In IEEE International Conference on Intelligent Transportation Systems, 2021. \n[50] Chunwei Wang, Chao Ma, Ming Zhu, and Xiaokang Yang. Pointaugmenting: Cross-modal augmentation for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition, 2021. \n[51] Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Chang Huang, and Wenyu Liu. Polar parametrization for vision-based surround-view 3d detection. arXiv:2206.10965, 2022. \n[52] Aleksandr Kim, Aljoša Ošep, and Laura Leal-Taixé. Eagermot: 3d multi-object tracking via sensor fusion. In IEEE International Conference on Robotics and Automation, 2021. \n[53] Yihan Zeng, Chao Ma, Ming Zhu, Zhiming Fan, and Xiaokang Yang. Cross-modal 3d object detection and tracking for auto-driving. In IEEE/RSJ International Conference on Intelligent Robots and Systems, 2021. ",
|
| 1010 |
+
"bbox": [
|
| 1011 |
+
171,
|
| 1012 |
+
80,
|
| 1013 |
+
828,
|
| 1014 |
+
916
|
| 1015 |
+
],
|
| 1016 |
+
"page_idx": 10
|
| 1017 |
+
},
|
| 1018 |
+
{
|
| 1019 |
+
"type": "text",
|
| 1020 |
+
"text": "",
|
| 1021 |
+
"bbox": [
|
| 1022 |
+
168,
|
| 1023 |
+
68,
|
| 1024 |
+
828,
|
| 1025 |
+
917
|
| 1026 |
+
],
|
| 1027 |
+
"page_idx": 11
|
| 1028 |
+
},
|
| 1029 |
+
{
|
| 1030 |
+
"type": "text",
|
| 1031 |
+
"text": "",
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
171,
|
| 1034 |
+
92,
|
| 1035 |
+
828,
|
| 1036 |
+
511
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 12
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "text",
|
| 1042 |
+
"text": "Checklist ",
|
| 1043 |
+
"text_level": 1,
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
174,
|
| 1046 |
+
536,
|
| 1047 |
+
254,
|
| 1048 |
+
553
|
| 1049 |
+
],
|
| 1050 |
+
"page_idx": 12
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"type": "text",
|
| 1054 |
+
"text": "1. For all authors... ",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
214,
|
| 1057 |
+
564,
|
| 1058 |
+
339,
|
| 1059 |
+
578
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 12
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "text",
|
| 1065 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] \n(b) Did you describe the limitations of your work? [Yes] See Section 5. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] See Section 5. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 1066 |
+
"bbox": [
|
| 1067 |
+
238,
|
| 1068 |
+
583,
|
| 1069 |
+
825,
|
| 1070 |
+
688
|
| 1071 |
+
],
|
| 1072 |
+
"page_idx": 12
|
| 1073 |
+
},
|
| 1074 |
+
{
|
| 1075 |
+
"type": "text",
|
| 1076 |
+
"text": "2. If you are including theoretical results... ",
|
| 1077 |
+
"bbox": [
|
| 1078 |
+
214,
|
| 1079 |
+
693,
|
| 1080 |
+
493,
|
| 1081 |
+
707
|
| 1082 |
+
],
|
| 1083 |
+
"page_idx": 12
|
| 1084 |
+
},
|
| 1085 |
+
{
|
| 1086 |
+
"type": "text",
|
| 1087 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
|
| 1088 |
+
"bbox": [
|
| 1089 |
+
238,
|
| 1090 |
+
710,
|
| 1091 |
+
738,
|
| 1092 |
+
742
|
| 1093 |
+
],
|
| 1094 |
+
"page_idx": 12
|
| 1095 |
+
},
|
| 1096 |
+
{
|
| 1097 |
+
"type": "text",
|
| 1098 |
+
"text": "3. If you ran experiments... ",
|
| 1099 |
+
"bbox": [
|
| 1100 |
+
212,
|
| 1101 |
+
747,
|
| 1102 |
+
393,
|
| 1103 |
+
761
|
| 1104 |
+
],
|
| 1105 |
+
"page_idx": 12
|
| 1106 |
+
},
|
| 1107 |
+
{
|
| 1108 |
+
"type": "text",
|
| 1109 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] We give detailed experimental setup in Section 4.1 and publicly release the code. \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See Section 4.1 and supplementary material. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [No] \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] We report them in the supplementary material. Our experiments in this paper require about 8 NVIDIA Tesla V100 GPUs. ",
|
| 1110 |
+
"bbox": [
|
| 1111 |
+
238,
|
| 1112 |
+
765,
|
| 1113 |
+
825,
|
| 1114 |
+
911
|
| 1115 |
+
],
|
| 1116 |
+
"page_idx": 12
|
| 1117 |
+
},
|
| 1118 |
+
{
|
| 1119 |
+
"type": "text",
|
| 1120 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1121 |
+
"bbox": [
|
| 1122 |
+
212,
|
| 1123 |
+
90,
|
| 1124 |
+
825,
|
| 1125 |
+
106
|
| 1126 |
+
],
|
| 1127 |
+
"page_idx": 13
|
| 1128 |
+
},
|
| 1129 |
+
{
|
| 1130 |
+
"type": "text",
|
| 1131 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] See Section 4.1. \n(b) Did you mention the license of the assets? [Yes] We conduct experiments on nuScenes [43] dataset with custom non-commercial license. \n(c) Did you include any new assets either in the supplemental material or as a URL? [No] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [No] We conduct experiments only on public datasets. ",
|
| 1132 |
+
"bbox": [
|
| 1133 |
+
238,
|
| 1134 |
+
109,
|
| 1135 |
+
825,
|
| 1136 |
+
232
|
| 1137 |
+
],
|
| 1138 |
+
"page_idx": 13
|
| 1139 |
+
},
|
| 1140 |
+
{
|
| 1141 |
+
"type": "text",
|
| 1142 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1143 |
+
"bbox": [
|
| 1144 |
+
212,
|
| 1145 |
+
236,
|
| 1146 |
+
705,
|
| 1147 |
+
250
|
| 1148 |
+
],
|
| 1149 |
+
"page_idx": 13
|
| 1150 |
+
},
|
| 1151 |
+
{
|
| 1152 |
+
"type": "text",
|
| 1153 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
|
| 1154 |
+
"bbox": [
|
| 1155 |
+
238,
|
| 1156 |
+
253,
|
| 1157 |
+
826,
|
| 1158 |
+
343
|
| 1159 |
+
],
|
| 1160 |
+
"page_idx": 13
|
| 1161 |
+
}
|
| 1162 |
+
]
|
parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/XA4ru9mfxTP/XA4ru9mfxTP_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/XdDl3bFUNn5/XdDl3bFUNn5_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/dev/XtyeppctGgc/XtyeppctGgc.md
ADDED
|
@@ -0,0 +1,311 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning
|
| 2 |
+
|
| 3 |
+
Dongze Lian1∗ Daquan Zhou1,2∗ Jiashi Feng2 Xinchao Wang1 1National University of Singapore 2ByteDance {dongze,xinchao}@nus.edu.sg {zhoudaquan21,jshfeng}@gmail.com
|
| 4 |
+
|
| 5 |
+
# Abstract
|
| 6 |
+
|
| 7 |
+
Existing fine-tuning methods either tune all parameters of the pre-trained model (full fine-tuning), which is not efficient, or only tune the last linear layer (linear probing), which suffers a significant accuracy drop compared to the full fine-tuning. In this paper, we propose a new parameter-efficient fine-tuning method termed as SSF, representing that researchers only need to Scale and Shift the deep Features extracted by a pre-trained model to catch up with the performance of full finetuning. In this way, SSF also surprisingly outperforms other parameter-efficient fine-tuning approaches even with a smaller number of tunable parameters. Furthermore, different from some existing parameter-efficient fine-tuning methods (e.g., Adapter or VPT) that introduce the extra parameters and computational cost in the training and inference stages, SSF only adds learnable parameters during the training stage, and these additional parameters can be merged into the original pre-trained model weights via re-parameterization in the inference phase. With the proposed SSF, our model obtains $2 . 4 6 \%$ $( 9 0 . 7 2 \%$ vs. $8 8 . 5 4 \%$ ) and $1 1 . 4 8 \%$ $7 3 . 1 0 \%$ vs. $6 5 . 5 7 \%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy compared to the full fine-tuning but only fine-tuning about $0 . 3 { \bf M }$ parameters. We also conduct amounts of experiments in various model families (CNNs, Transformers, and MLPs) and datasets. Results on 26 image classification datasets in total and 3 robustness & out-of-distribution datasets show the effectiveness of SSF. Code is available at https://github.com/dongzelian/SSF.
|
| 8 |
+
|
| 9 |
+
# 1 Introduction
|
| 10 |
+
|
| 11 |
+
With the popularity of the data-driven methods in the deep learning community, the dataset scale and the model size have both got huge explosions. There is a tendency to explore large models and then adopt these pre-trained models in downstream tasks to achieve better performance and faster convergence, which gradually becomes a common way.
|
| 12 |
+
|
| 13 |
+
However, the current procedure depends on full fine-tuning heavily, where all the parameters of the model are updated. It inevitably causes the model to be over-fitted to the small target dataset and thus cannot be used for other tasks after the fine-tuning. As a result, the device will need to save a dedicated set of model parameters for each task, which causes a huge amount of storage space, especially for today’s large models (e.g., ViT-G/14 [11] 1.8G, CoAtNet [5] 2.4G).
|
| 14 |
+
|
| 15 |
+
A simple solution for the above problem is linear probing [16], where only the last head layer is fine-tuned. However, this practice usually yields inferior performance compared to the full fine-tuning proxy. Motivated by the success of the parameter-efficient fine-tuning strategy with prompt in the field of natural language processing (NLP) [21, 32, 23, 19], the recent work implements a similar proxy on vision tasks [29], termed as Visual Prompt Tuning (VPT). Specifically, VPT [29] proposes to insert learnable prompts as inputs and append them to the original image tokens. These prompts
|
| 16 |
+
|
| 17 |
+
<table><tr><td>Method</td><td>Acc.</td><td>Params. (M)</td><td>Unified parameter space</td><td>No extra inference params.</td></tr><tr><td>Full fine-tuning</td><td>93.82</td><td>85.88</td><td>√</td><td>√</td></tr><tr><td>Linear probing</td><td>88.70</td><td>0.08</td><td>√</td><td>√</td></tr><tr><td>Adapter [21] VPT [29]</td><td>93.34</td><td>0.31</td><td>√</td><td>×</td></tr><tr><td></td><td>93.17</td><td>0.54</td><td>×</td><td>×</td></tr><tr><td>SSF (ours)</td><td>93.99</td><td>0.28</td><td>√</td><td>√</td></tr></table>
|
| 18 |
+
|
| 19 |
+
Table 1: Characteristics of different finetuning methods. Acc. means the Top-1 accuracy $( \% )$ on CIFAR-100 with a pre-trained ViT-B/16 for tuning. Params. means the learnable parameters at fine-tuning. Our SSF has a unified learnable parameter space and does not require extra inference parameters while obtaining superior performance.
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
Figure 1: Performance comparisons of seven finetuning methods with a pre-trained ViT-B/16 model on the FGVC dataset and VTAB-1k benchmark. Our SSF (red dots) achieves state-of-the-art performance only with about $0 . 3 { \bf M }$ average learnable parameters.
|
| 23 |
+
|
| 24 |
+
will interact with the image tokens by performing self-attention and are updated during the fine-tuning process. In this manner, a significant performance improvement can be achieved in downstream tasks compared to a linear probing proxy. Nevertheless, compared to the full fine-tuning and linear probing, it additionally raises two issues: i) VPT tunes the number of prompts for different tasks, which introduces a task-dependent learnable parameter space. The fine-tuning performance is sensitive to the number of prompts for each task and needs to be carefully designed. Too few or too many prompts might either degrade the accuracy of fine-tuning or increase the redundancy of the computation (e.g., 200 prompts on Clevr/count vs. 1 prompt on Flowers102); ii) VPT [29], as well as other Adapter-based methods [21, 42], introduces additional parameters and computational cost in the inference phase compared to the original pre-trained model. For instance, VPT introduces additional inputs for self-attention with image tokens. Adapter-based methods insert additional modules into the pre-trained model. These methods change the specific backbone architecture or the input of the network, which might result in frequent structure modifications and heavy workload, especially for those models that are already deployed in edge devices (e.g., mobile phones).
|
| 25 |
+
|
| 26 |
+
To cope with the above issues, we attempt to find a general proxy for parameter-efficient finetuning, where the learnable parameter space is unified (task-independent) and no additional inference parameters are introduced. Inspired by some feature modulation methods [58, 25, 45], we propose a new parameter-efficient fine-tuning method named SSF, where you only need to Scale and Shift your deep Features extracted by a pre-trained model for fine-tuning. The intuition behind our approach come from the fact that the upstream datasets and downstream datasets have different data distributions [50]. Therefore, it is difficult to apply the model weights trained in the upstream dataset to the downstream dataset. For instance, a naive linear probing strategy with keeping the weights of backbone frozen will cause performance degradation. To alleviate the above problem, SSF introduces scale parameters and shift parameters, which could be considered as variance and mean to modulate the features of the downstream dataset extracted with the pre-trained model on the upstream dataset, such that the modulated feature falls in a discriminative space. These scale parameters and shift parameters do not depend on any input and have a unified learnable parameter space for different tasks. Another advantage of SSF is that it only introduces linear transformations because we scale and shift the extracted features. These linear transformations could be further merged into the original pre-trained weight via model re-parameterization [10] in the inference phase, thus avoiding the extra parameters and FLOPs for downstream tasks. For a deployed model in edge devices, only the updated weights after fine-tuning need to be uploaded instead of changing the backbone architecture. Table 1 shows the specific characteristics comparisons between SSF and other fine-tuning methods. SSF is simple, effective, and efficient, which also conforms to Occam’s Razor principle. Therefore, we explore this new baseline and find that it surprisingly outperforms all other parameter-efficient fine-tuning methods.
|
| 27 |
+
|
| 28 |
+
We evaluate our method on 26 classification datasets in total and 3 robustness & out-of-distribution datasets. SSF obtains state-of-the-art performance compared to other parameter-efficient fine-tuning methods with the trainable parameters and accuracy trade-off (Table 1 and Figure 1). Compared to the full fine-tuning, our method obtains $2 . 4 6 \%$ $9 0 . 7 2 \%$ vs. $8 8 . 5 4 \%$ ) and $1 1 . 4 8 \%$ ( $7 3 . 1 0 \%$ vs. $6 5 . 5 7 \%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy but only with about $0 . 3 { \bf M }$ trainable parameters. Furthermore, our SSF does not require additional parameters during the inference phase. It is plug-and-play and is very easy to extend to various model families (CNNs, Transformers, and MLPs). Our SSF establishes a new baseline and we hope that it brings more insight into the field of the efficient model tuning.
|
| 29 |
+
|
| 30 |
+
# 2 Related Work
|
| 31 |
+
|
| 32 |
+
# 2.1 Model Families
|
| 33 |
+
|
| 34 |
+
Convolution has been used for a long time as the main module to extract the image features in computer vision tasks, and CNN-based architectures have been studied [49, 18, 59, 39, 60, 37, 63] with extension on graph-based data [62, 61, 36]. Recently, another architecture family, Transformer, has gained widespread attention owing to its great success in NLP [56, 8, 23]. Following this direction, Dosovitskiy et al. [11] first employ a transformer in the domain of computer vision and introduce a new architecture paradigm, ViT, which achieves promising results [64, 48]. Subsequently, various transformer-based models, such as DeiT [53] and Swin Transformer [38], are introduced and shown to be effective on a variety of tasks such as object detection, semantic segmentation, action recognition [40], etc. In another line, Tolstikhin et al. [52] propose a pure MLP-based architecture, and subsequent papers [20, 33] have interestingly demonstrated that the MLP-based architectures can catch up to transformers. However, in addition to the well-designed modules, their excellent performance is also attributed to the deployment of large-scale models. Given a large-scale model pre-trained on a large dataset, how to perform parameter-efficient fine-tuning in downstream tasks is essential but is currently less explored. In this paper, we propose SSF as a new baseline and show its promising performance with comprehensive validation in a wide variety of tasks.
|
| 35 |
+
|
| 36 |
+
# 2.2 Pre-training and Fine-tuning
|
| 37 |
+
|
| 38 |
+
Early models [18, 24, 22, 59, 51] are usually pre-trained on the ImageNet-1K dataset, and then fine-tuned on downstream tasks to achieve faster convergence [17] or better performance. Such a procedure is called pre-training and fine-tuning, or transfer learning. Recent works tend to employ larger models (e.g., ViT [11] and Swin Transformer V2 [38]) and train them on larger datasets (e.g., ImageNet-21K and JFT-300M) in pursuit of better performance. Both in the domains of NLP and computer vision, these large models [8, 38, 47, 15, 69, 70] achieve enormous performance improvements compared to the small-scale models and provide pre-trained weights for downstream tasks. Some other works attempt to explore how to efficiently fine-tune the pre-trained models [13, 71] on the target tasks. For instance, given a target task, SpotTune [13] investigates which layers need to be fine-tuned. Touvron et al. [54] find that fine-tuning the weights of the attention layers and freezing weights of the other parts is sufficient to adapt the vision transformers to other downstream tasks. Some works also propose to insert adapters into the network to fine-tune in a parameter-efficient way. These adapters can be a small non-linear network [21], a hyper-network that generates model weights [43], or a compactor [42] which performs a low-rank decomposition to reduce the parameters. Some works have also tried to only update the bias term [2, 66]. More recently, VPT [29] proposes to insert a small number of learnable parameters (prompts) and optimize them while freezing the backbone, which achieves significant performance improvement compared to the full fine-tuning. During the submission of this work, some methods [3, 68] are also proposed for parameter-efficient fine-tuning, e.g., inserting a adapter module or neural prompt search. Different from all the above works, we propose to scale and shift deep features extracted by a pre-trained model, which is simple but effective and outperforms other parameter-efficient fine-tuning methods.
|
| 39 |
+
|
| 40 |
+
# 2.3 Feature Modulation
|
| 41 |
+
|
| 42 |
+
Many works have attempted to modulate features to obtain better performance. The most relevant ones to our work are various normalization methods [26, 1, 58]. BN, LN, and GN usually normalize the features and then transform them linearly with scale and shift factors to modulate feature distribution, which has been verified to be effective in amounts of tasks. STN [28] introduces a learnable module to spatially transform feature maps. In the field of image generation, AdaIN [25] generates scale and shift factors to characterize specific image styles. Self-modulation [4] shows GANs benefit from self-modulation layers in the generator. In vision-language tasks, Conditional BN [6] and FiLM [45] are often utilized to modulate the features of two modalities. Unlike some algorithms such as BN, our SSF is not limited to the modulation of normalization layer, and it has a different motivation that is to alleviate the distribution mismatch between upstream tasks and downstream tasks for parameter-efficient fine-tuning. As a comparison, we also conduct experiments in Sec. 4.3 and show that our SSF is more effective compared to only tuning the normalization layer. Compared to STN, AdaIN, FiLM and so on, our method is input-independent and these scale and shift parameters model the distribution of the whole dataset so that they can be absorbed into the original pre-trained model weights in the inference phase.
|
| 43 |
+
|
| 44 |
+
# 2.4 Model Re-parameterization
|
| 45 |
+
|
| 46 |
+
Model re-parameterization has been a common practice to improve inference efficiency. One of the representative techniques is batch normalization folding used in the model compression algorithms [27]. The parameters introduced by the batch normalization layers [26] are merged into the convolutional layers usually stacked before them. This technique is further utilized to merge different branches of networks into a new branch [65, 10, 9]. Similarly, our SSF fully adopts linear transformations, which allows the scale and shift parameters in the training phase to be merged into the original pre-trained model weights, thus avoiding the introduction of the extra parameters and computational cost during the inference phase.
|
| 47 |
+
|
| 48 |
+
# 3 Approach
|
| 49 |
+
|
| 50 |
+
# 3.1 Preliminaries
|
| 51 |
+
|
| 52 |
+
Transformers. In a vision transformer (ViT) [11], an RGB image $I \in \mathbb { R } ^ { 3 \times H \times W }$ is divided into $N \times N$ non-overlapping patches, and then these image patches appended a class token are fed into an embedding layer followed by the $L$ -layer vision transformer blocks with self-attention as the core operation. The input $\boldsymbol { x } \in \mathbb { R } ^ { ( N ^ { \hat { 2 } } + 1 ) \times d }$ , where $d$ is the embedding dimension, is first transformed to keys $K \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ , values $V \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ , and queries $\bar { Q ^ { \prime } } \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ . After that, we can calculate a global self-attention by
|
| 53 |
+
|
| 54 |
+
$$
|
| 55 |
+
\mathrm { A t t e n t i o n } ( Q , K , V ) = \mathrm { S o f t m a x } ( \frac { Q K ^ { T } } { \sqrt { d } } ) V .
|
| 56 |
+
$$
|
| 57 |
+
|
| 58 |
+
The output of the attention layer will be fed to a two-layer MLP to extract information in the channel dimension.
|
| 59 |
+
|
| 60 |
+
Adapter. Adapter [21] is inserted into the transformer layer for efficient fine-tuning. It is a bottleneck module with a few trainable parameters, which contains a down-projection to reduce the feature dimension, a non-linear activation function, and an up-projection to project back to the original dimension. Therefore, given the input $\boldsymbol { x } \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ , the output is calculated by
|
| 61 |
+
|
| 62 |
+
$$
|
| 63 |
+
\mathrm { o u t } = [ W ^ { \mathrm { u p } } \phi ( W ^ { \mathrm { d o w n } } x ^ { T } ) ] ^ { T } ,
|
| 64 |
+
$$
|
| 65 |
+
|
| 66 |
+
where $W ^ { \mathrm { d o w n } } \in \mathbb { R } ^ { d ^ { \prime } \times d }$ (where $d ^ { \prime } \ll d ,$ ), $\phi$ , and $W ^ { \mathrm { u p } } \in \mathbb { R } ^ { d \times d ^ { \prime } }$ represent the down-projection matrix, non-linear function, and up-projection matrix, respectively.
|
| 67 |
+
|
| 68 |
+
VPT. VPT [29] inserts some learnable parameters (i.e., prompts) into the input space after the embedding layer. These prompts interact with the original image tokens by performing self-attention. During the fine-tuning, the weights of the backbone network are kept frozen and only the parameters of the prompts are updated. VPT-Shallow inserts prompts in the first layer while VPT-Deep inserts prompts in all the layers of the transformer. Assuming that the input is $\overline { { x } } \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ , denote the inserted prompts as $\dot { \boldsymbol { p } } \in \mathbb { R } ^ { n \times d }$ , where $n$ is the number of prompts, the combined tokens $x ^ { \prime }$ is
|
| 69 |
+
|
| 70 |
+
$$
|
| 71 |
+
x ^ { \prime } = [ x ; p ] ,
|
| 72 |
+
$$
|
| 73 |
+
|
| 74 |
+
where $x ^ { \prime } \in \mathbb { R } ^ { ( N ^ { 2 } + n + 1 ) \times d }$ will be fed into the transformer block for self-attention (Eq. (1)).
|
| 75 |
+
|
| 76 |
+
# 3.2 Scaling and Shifting Your Features for Fine-tuning
|
| 77 |
+
|
| 78 |
+
Different from the above methods, we introduce both the scale and shift factors to modulate deep features extracted by a pre-trained model with linear transformation to match the distribution of a target dataset, as mentioned in Sec. 1. Five main properties are covered in our method: i) SSF achieves on-par performance with the full fine-tuning strategy; ii) all downstream tasks can be inputted to the model independently without relying on any other task; iii) the model only needs to fine-tune very few parameters; iv) unlike VPT [29], which adjusts the number of prompts for each task, the set of parameters for fine-tuning in SSF does not change as the task changes, making it feasible to further fine-tune the parameters later by adding more tasks for multi-task learning or continuous learning2; v) thanks to the linear transformation, SSF avoids the introduction of the extra parameters and computational cost during the inference phase, making our method zero overhead.
|
| 79 |
+
|
| 80 |
+
The design of SSF. SSF performs the linear transformation to modulate the features for parameter-efficient fine-tuning as shown in Figure 2. In Figure 2 (a), given a model pre-trained in the upstream task, we insert SSF-ADA3 after each operation (OP) of the network to modulate features. There are $K$ OPs in total and these operations might contain multi-head self-attention (MSA), MLP and layer normalization (LN), etc. During the fine-tuning, the pre-trained weights in these operations are kept frozen and the SSFADA parameters are kept updated. The specific SSF-ADA structure is shown in Figure2 (c), where the features output from the previous operation are performed dot product with a scale factor and then summed with a shift factor, which are input-independent. Formally, given the input $\boldsymbol { x } \in \bar { \mathbb { R } } ^ { ( N ^ { 2 } + 1 ) \times d }$ , the output $\boldsymbol { y } \in \mathbb { R } ^ { ( N ^ { 2 } + 1 ) \times d }$ (is also the input of the next operation) is calculated by
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
Figure 2: The overall pipeline of SSF. (a) Training pipeline via SSF, where an OP means an operation, e.g., MSA, MLP or LN. (b) A pre-trained model or inference pipeline. (c) Our SSF-ADA.
|
| 84 |
+
|
| 85 |
+
$$
|
| 86 |
+
y = \gamma \odot x + \beta ,
|
| 87 |
+
$$
|
| 88 |
+
|
| 89 |
+
where $\gamma \in \mathbb { R } ^ { d }$ and $\beta \in \mathbb { R } ^ { d }$ are the scale and shift factors, respectively. $\odot$ is the dot product.
|
| 90 |
+
|
| 91 |
+
Re-parameterization. Since SSF-ADA is a completely linear transformation, we can re-parameterize it by absorbing the scale and shift terms into the previous linear layer as follows
|
| 92 |
+
|
| 93 |
+
$$
|
| 94 |
+
y = \gamma \odot x + \beta = \gamma \odot ( w * t + b ) + \beta = ( \gamma \odot w ) * t + \gamma \odot b + \beta ,
|
| 95 |
+
$$
|
| 96 |
+
|
| 97 |
+
where $w$ and $b$ are the weight and bias terms, respectively. $^ *$ represents the ‘convolution’ operation in the convolutional layer or the ‘multiplication’ operation in the MLP layer. $t$ is the input of the previous linear layer. Since $w$ and $b$ are frozen and $\gamma$ and $\beta$ are updated in the fine-tuning, $\gamma$ and $\beta$ can be merged into the original parameter space ( $\dot { } w$ and $b$ ) in the inference stage through the above formulation. From this perspective, our SSF-ADA makes it possible to perform downstream tasks without adding any extra parameters and computational costs, as shown in Figure2 (b).
|
| 98 |
+
|
| 99 |
+
Discussion. The first question is why we want the input $\gamma$ and $\beta$ to be input-independent. As FiLM [45] and AdaIN [25] show, we could obtain $\gamma$ and $\beta$ by conditioning an image sample, however, this might cause two shortcomings. One is that we want $\gamma$ and $\beta$ to be input-independent to represent the distribution of the whole downstream dataset so that we can modify the previous weight distribution to fit the downstream dataset by modulating the feature. Secondly, the conditional input requires the introduction of some additional networks (e.g., MLPs) to generate $\gamma$ and $\beta$ , which introduces more trainable parameters. More importantly, to better generate $\gamma$ and $\beta$ , a non-linear activation function might be required, which will lead to the intractability of the re-parameterization. Therefore, we directly perform a fully linear transformation to merge the $\gamma$ and $\beta$ factors into the original pre-trained weights, so that weights can be easily uploaded to the edge devices without any modification of the backbone architecture.
|
| 100 |
+
|
| 101 |
+
The second question is which operations should be followed by SSF-ADA. Our experience is that you can insert SSF-ADA after each operation with a linear coefficient in ViT. Although we can search for some optimal layers or operations with Neural Architecture Search (NAS) [46, 35, 14, 34], to reduce the number of the trainable parameters, we believe that our method will produce better results (or not worse than NAS) without introducing too many trainable parameters that can be merged for inference, as will be shown in Sec. 4.3.
|
| 102 |
+
|
| 103 |
+
# 3.3 Complexity Analysis
|
| 104 |
+
|
| 105 |
+
We also compare the complexity of Adapter, VPT and our SSF. Take a ViT as an example, the dimension and number of the tokens are $d$ and $N ^ { 2 }$ . Assuming that Adapter projects features from $d$ -dim to $d ^ { \prime }$ -dim (where $d ^ { \prime } \ll d _ { , }$ ) so that the extra trainable parameters are $\bar { 2 } d d ^ { \prime }$ in each layer,
|
| 106 |
+
|
| 107 |
+
VPT inserts $n$ prompts to obtain nd extra parameters in each layer, and SSF inserts SSF-ADA after each operation with a linear coefficient to obtain md extra parameters in each layer, when the total number of layers is $L$ , the complexity of Adapter, VPT and SSF is shown in Table 2. The specific number of additional parameters used by
|
| 108 |
+
|
| 109 |
+
Table 2: The complexity comparisons of Adapter [21], VPT [29] and our SSF. ‘ $( 1 ) ^ { \prime }$ : the same parameters and FLOPs for training and inference; $\mathbf { \eta } ^ { \mathrm { ( 0 ) } } \mathbf { : }$ no additional parameters and FLOPs are required for inference.
|
| 110 |
+
|
| 111 |
+
<table><tr><td>Method</td><td>Adapter</td><td>VPT-Shallow</td><td>VPT-Deep</td><td>SSF (ours)</td></tr><tr><td># Extra Params.</td><td>2Ldd' (1)</td><td>nd(1)</td><td>nLd(1)</td><td>mLd (0)</td></tr><tr><td>#Extra FLOPs</td><td>2N² Ldd' (1)</td><td></td><td>2n(2N² +n)d(1) | 2n(2N²+n)Ld(1)</td><td>mN²Ld (0)</td></tr></table>
|
| 112 |
+
|
| 113 |
+
Adapter, VPT and SSF depends on the values of $d ^ { \prime }$ , $n$ and $m$ . However, in practice, SSF outperforms Adapter and VPT-Deep even with slightly fewer parameters in the training stage as we will see in Sec. 4. Further, in the inference stage, borrowing the model re-parameterization strategy, the extra parameters and FLOPs of SSF are zero. However, the complexity of Adapter and VPT remain the same compared to the training, which establishes the strengths of our approach.
|
| 114 |
+
|
| 115 |
+
# 4 Experiments
|
| 116 |
+
|
| 117 |
+
# 4.1 Experimental Settings
|
| 118 |
+
|
| 119 |
+
Datasets. We mainly conduct our experiments on a series of datasets that can be categorized into three types as detailed below:
|
| 120 |
+
|
| 121 |
+
FGVC. Following VPT [29], we employ five Fine-Grained Visual Classification (FGVC) datasets to evaluate the effectiveness of our proposed SSF, which consists of CUB-200-2011 [57], NABirds [55], Oxford Flowers [44], Stanford Dogs [30] and Stanford Cars [12].
|
| 122 |
+
|
| 123 |
+
VTAB-1k. VTAB-1k benchmark is introduced in [67], which contains 19 tasks from diverse domains: i) Natural images that are captured by standard cameras; ii) Specialized images that are captured by non-standard cameras, e.g., remote sensing and medical cameras; iii) Structured images that are synthesized from simulated environments. This benchmark contains a variety of tasks (e.g., object counting, depth estimation) from different image domains and each task only contains 1,000 training samples, thus is extremely challenging.
|
| 124 |
+
|
| 125 |
+
General Image Classification Datasets. We also validate the effectiveness of SSF on general image classification tasks. We choose the CIFAR-100 [31] and ImageNet-1K [7] datasets as evaluation datasets, where CIFAR-100 contains 60,000 images with 100 categories. ImageNet-1K contains 1.28M training images and 50K validation images with 1,000 categories, which are very large datasets for object recognition.
|
| 126 |
+
|
| 127 |
+
Models. For a fair comparison, we follow VPT [29] and mainly select ViT-B/16 [11] model pretrained on ImageNet-21K as the initialization for fine-tuning. In addition, we also generalize our method to backbones of different model families, including the recent Swin Transformer [38] (SwinB), ConvNeXt-B [39] and AS-MLP-B [33]. The former builds a hierarchical transformer-based architecture, and the latter two belong to CNN-based architecture and MLP-based architecture respectively.
|
| 128 |
+
|
| 129 |
+
Baselines. We first compare our method with the two basic fine-tuning methods: i) full fine-tuning, where all parameters of the models are updated at fine-tuning; ii) linear probing, where only the parameters of the classification head (an MLP layer) are updated. We also compare our method with recent parameter-efficient fine-tuning methods: iii) Adapter [21], where a new adapter structure with up-projection, non-linear function, and down-projection is inserted into the transformer and only the parameters of this new module are updated; iv) Bias [66], where all the bias terms of parameters are updated; v) VPT [29], where the prompts are inserted into transformers as the input tokens and they are updated at fine-tuning.
|
| 130 |
+
|
| 131 |
+
Implementation Details. For the FGVC datasets, we process the image with a randomly resize crop to $2 2 4 \times 2 2 4$ and a random horizontal flip for data augmentation. For VTAB-1k, we directly resize the image to $2 2 4 \times 2 2 4$ , following the default settings in VTAB [67]. For CIFAR-100 and ImageNet1K, we follow the fine-tuning setting of ViT-B/16 in [11], where the stronger data augmentation strategies are adopted. We employ the AdamW [41] optimizer to fine-tune models for 100 epochs for CIFAR-100, and 30 epochs for ImageNet-1K. The cosine decay strategy is adopted for the learning rate schedule, and the linear warm-up is used in the first 10 epochs for CIFAR-100 and 5 epochs for ImageNet-1K.
|
| 132 |
+
|
| 133 |
+
<table><tr><td rowspan=1 colspan=2>DatasetMethod</td><td rowspan=1 colspan=1>CUB-200-2011</td><td rowspan=1 colspan=1>NABirds</td><td rowspan=1 colspan=1>OxfordFlowers</td><td rowspan=1 colspan=1>StanfordDogs</td><td rowspan=1 colspan=1>StanfordCars</td><td rowspan=1 colspan=1>Mean</td><td rowspan=1 colspan=1>Params.(M)</td></tr><tr><td rowspan=1 colspan=2>Full fine-tuningLinear probing</td><td rowspan=1 colspan=1>87.385.3</td><td rowspan=1 colspan=1>82.775.9</td><td rowspan=1 colspan=1>98.897.9</td><td rowspan=1 colspan=1>89.486.2</td><td rowspan=1 colspan=1>84.551.3</td><td rowspan=1 colspan=1>88.5479.32</td><td rowspan=1 colspan=1>85.980.18</td></tr><tr><td rowspan=2 colspan=2>Adapter [21]Bias [66]VPT-Shallow [29]</td><td rowspan=1 colspan=1>87.188.4</td><td rowspan=1 colspan=1>84.384.2</td><td rowspan=1 colspan=1>98.598.8</td><td rowspan=1 colspan=1>89.891.2</td><td rowspan=1 colspan=1>68.679.4</td><td rowspan=1 colspan=1>85.6788.41</td><td rowspan=1 colspan=1>0.410.28</td></tr><tr><td rowspan=1 colspan=1>VPT-Shallow [29]</td><td rowspan=2 colspan=1>86.788.5</td><td rowspan=2 colspan=1>78.884.2</td><td rowspan=2 colspan=1>98.499.0</td><td rowspan=2 colspan=1>90.790.2</td><td rowspan=2 colspan=1>68.783.6</td><td rowspan=2 colspan=1>84.6289.11</td><td rowspan=2 colspan=1>0.250.85</td></tr><tr><td rowspan=1 colspan=2>VPT-Deep [29]</td></tr><tr><td rowspan=1 colspan=2>SSF (ours)</td><td rowspan=1 colspan=1>89.5</td><td rowspan=1 colspan=1>85.7</td><td rowspan=1 colspan=1>99.6</td><td rowspan=1 colspan=1>89.6</td><td rowspan=1 colspan=1>89.2</td><td rowspan=1 colspan=1>90.72</td><td rowspan=1 colspan=1>0.39</td></tr></table>
|
| 134 |
+
|
| 135 |
+
Table 3: Performance comparisons on five FGVC datasets with ViT-B/16 models pre-trained on ImageNet-21K.
|
| 136 |
+
|
| 137 |
+
<table><tr><td></td><td colspan="5">Natural Specialized</td><td colspan="5">Structured</td></tr><tr><td>Dataset Method</td><td>CEIATIP0 Crreeaer 0</td><td>rmeii</td><td>L63uns NHAS 3</td><td>rrarar gan Bto.AI P5ss55</td><td>Phrroiar Ceeretlrt</td><td>TTsisrlsr Cresisttla poresidsp qTTa</td><td>szhTI1111 puorssitst</td><td>SrHTIPt</td><td></td><td>) sr</td></tr><tr><td>Full fine-tuning [29] Linear probing [29] Adapter [21]</td><td>68.9 87.7 63.485.0</td><td>64.3 97.2 63.2 97.0 63.2 97.7</td><td>86.987.438.8 86.336.651.0 87.0 34.6 50.8</td><td>79.7 95.7 84.2 78.587.5 76.3 88.0 73.1</td><td>73.9 68.6 74.0 70.5</td><td>56.358.6 41.7 65.5 34.330.6 33.2 55.4 37.4 31.2 53.2</td><td>57.546.7 12.520.0 30.3</td><td>25.729.1 9.619.2 22.1</td><td>52.94</td><td>65.57 85.84 0.04</td></tr><tr><td>Bias[66] VPT-Shallow [29] VPT-Deep [29]</td><td>74.1 86.1 72.887.059.2 77.7 86.9</td><td>97.5 62.6 97.5</td><td>85.359.9 51.4 87.374.5 51.2</td><td>78.7 91.672.9 78.2 92.075.6</td><td>45.7 69.8 72.9 50.5 68.4 68.5</td><td>61.5 55.6 32.4 55.9 58.640.5 67.1</td><td>25.4 13.8 66.640.0 68.7 36.1 20.2</td><td>15.7 25.1 34.1</td><td>55.82 62.05 64.85</td><td>0.27 0.14 0.11</td></tr><tr><td>SSF (ours)</td><td>78.890.8</td><td>65.8 98.0 88.3 78.1</td><td>49.6</td><td>81.896.1 83.4</td><td>60.046.5 69.0 92.6 75.199.4 91.8 90.2 52.9|87.4 95.987.4 75.5|75.9 62.3 53.380.6 77.3 54.9 29.5 37.9|73.10</td><td>72.8 73.647.9</td><td>32.9</td><td>37.8</td><td>69.43</td><td>0.60 0.24</td></tr></table>
|
| 138 |
+
|
| 139 |
+
Table 4: Performance comparisons on the VTAB-1k benchmark with ViT-B/16 models pre-trained on ImageNet-21K.
|
| 140 |
+
|
| 141 |
+
# 4.2 Performance Comparisons on Image Classification
|
| 142 |
+
|
| 143 |
+
We compare the performance of our SSF and other baseline methods in 26 image classification tasks and the results on FGVC and VTAB-1k are shown in Table 3 and Table 4 (also see Figure 1), respectively, and the results on CIFAR-100 and ImageNet-1K are shown in Table 5, which are evaluated in Top-1 accuracy $( \% )$ . In these three tables, the bold font shows the best accuracy of all methods and the underline font shows the second best accuracy.
|
| 144 |
+
|
| 145 |
+
We have the following findings by observing them: i) In Table 3 and Table 4, where the last column is the average of the fine-tuned parameters for each method on the corresponding datasets, our SSF outperforms VPT [29] and other parameter-efficient fine-tuning methods, and even achieves better performance than full fine-tuning, which is mainly owing to the linear transformation applied on the features. Specifically, SSF obtains $1 . 8 1 \%$ $( 9 0 . 7 2 \%$ vs. $8 9 . 1 1 \%$ and $2 . 4 6 \%$ $9 0 . 7 2 \%$ vs. $8 8 . 5 4 \%$ ) accuracy improvement on five FGVC datasets, and $5 . 2 9 \%$ $7 3 . 1 0 \%$ vs. $6 9 . 4 3 \%$ ) and $1 1 . 4 8 \%$ $7 3 . 1 0 \%$ vs. $6 5 . 5 7 \%$ ) improvement on the VTAB-1k benchmark compared to VPT and full fine-tuning. Meanwhile, SSF also uses fewer trainable parameters compared to VPT-Deep in both datasets (0.39M vs. 0.85M, 0.24M vs. 0.60M). SSF maintains a unified learnable parameter space for different tasks with a few parameters while VPT [29] needs to design the different number of prompts for each task, which also shows the conciseness of our approach; ii) In Table 5, i.e., in CIFAR-100 and ImageNet-1K, SSF and other parameter-efficient fine-tuning methods have difficulty in achieving the similar performance to the full fine-tuning, probably because these datasets have sufficient data to prevent over-fitting of the model, especially in ImageNet-1K. In contrast, in the VTAB-1k benchmark, the amount of data is not very large (e.g., only 1,000 training images), which might cause over-fitting of the model for the full fine-tuning. Nevertheless, in CIFAR-100 and ImageNet-1K, our SSF still outperforms previous parameter-efficient fine-tuning methods (Adapter, Bias, and VPT), which shows the effectiveness of our method; iii) In Table 5, the results of our SSF with Swin Transformer, ConvNeXt, and AS-MLP models consistently outperform those of other parameter-efficient fine-tuning methods, which also verifies the effectiveness of SSF on a wide variety of models.
|
| 146 |
+
|
| 147 |
+
Computational cost. To validate the efficiency of our method, we show the computational cost of SSF in Figure 3. We employ a batch size of 16 for the training stage and inference stage, and use mixed precision training. All running results in Figure 3 are measured in a single GeForce RTX 2080Ti GPU. We can see that SSF has similar training time and training memory with VPT but with less inference time and inference memory. Here, we show the computational cost of VPT with 200/50 prompts (the same number of prompts to obtain the performance in Table 5) for VPT-Shallow and VPT-Deep, respectively. When adding the number of prompts, the time cost and memory will be larger but our SSF achieves zero-overhead inference, which is more advantageous.
|
| 148 |
+
|
| 149 |
+
<table><tr><td>Model</td><td colspan="3">ViT-B/16 [11]</td><td colspan="4">Swin-B [38]</td><td colspan="4">ConvNeXt-B [39]</td><td colspan="2">|AS-MLP-B [33]</td></tr><tr><td>Dataset</td><td>CEIPPIII0 ) :seaed</td><td>Teege</td><td>) 'seed</td><td>CEI-AIPI1</td><td>Jr) 'sr</td><td>Teee</td><td>) 'sirN</td><td>CEIPPIII0</td><td>Jr) 'rN</td><td>Trenege ) 'seed</td><td>CEIPPIPI0</td><td></td><td>n) 'se.d</td></tr><tr><td>Method Full fine-tuning Linear probing</td><td>93.82 85.88 88.70</td><td>0.08 82.04</td><td>83.58 86.57 0.77</td><td>89.27</td><td>93.85 86.858 0.10</td><td>83.25</td><td>1.03</td><td>85.20 88.03|94.14 87.67 89.20</td><td>0.10</td><td>85.80 88.85 84.05</td><td>1.03</td><td>89.96 79.04</td><td>86.83 0.10</td></tr><tr><td>Adapter [21] Bias[66] VPT-Shallow [29]</td><td>93.34 93.39 90.38</td><td>0.31 82.72 0.18 82.74 0.23 82.08</td><td>1.00 0.87 0.92</td><td>92.49 92.19 90.02</td><td>0.33 0.24 0.13</td><td>83.82 83.92 83.29</td><td>1.26 1.16 1.05</td><td>92.86 92.80 -</td><td>0.45 0.23</td><td>84.49 84.63 -</td><td>1.37 31.16 -</td><td>88.01 87.46 ·</td><td>0.33 0.26 ·</td></tr><tr><td>VPT-Deep [29] SSF (ours)</td><td>93.17 93.99</td><td>0.54 82.45 0.28 83.10</td><td>1.23 0.97</td><td>92.62 193.06</td><td>0.70 0.37</td><td>83.44 84.40</td><td>1.63 1.29</td><td>- 193.45</td><td>1 0.36</td><td>- 84.85</td><td>- 1.28</td><td>： 88.28</td><td>- 0.37</td></tr></table>
|
| 150 |
+
|
| 151 |
+
Table 5: Performance comparisons on CIFAR-100 and ImageNet-1K with various model families, where ViT-B/16, Swin-B, and ConvNeXt-B are pre-trained on ImageNet-21K, and AS-MLP-B is pre-trained on ImageNet-1K.
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
Figure 3: Computational cost of different tuning methods. From left to right: training time, training memory, test time, and test memory.
|
| 155 |
+
|
| 156 |
+
# 4.3 The Impacts of Different Designs
|
| 157 |
+
|
| 158 |
+
As the core operation of SSF, we thoroughly evaluate how SSF-ADA affects results, e.g., the insertion locations, the initialization of SSF-ADA and its components. We conduct experiments to analyze the impacts of different designs for fine-tuning. All experiments are implemented with pre-trained ViT-B/16 models on CIFAR-100 and the results are shown in Table 6.
|
| 159 |
+
|
| 160 |
+
The impact of the number of layers. We directly insert SSF-ADA into different layers to evaluate the effect of inserting layers, and the results are shown in table 6a. The values in the #layers column indicate the number of layers with SSF-ADA, where #layers-0 represents linear probing. From the first and second rows, we find that the results will improve from $8 8 . 7 0 \%$ to $9 2 . 6 9 \%$ and grow with a small number of trainable parameters (0.08M vs. 0.11M) when only inserting SSF-ADA into the first two layers. Keep adding SSF-ADA in the subsequent layers will make the results better. The growth of the results is almost linear with the number of layers of inserted SSF-ADA. Therefore, we directly choose to insert SSF-ADA into all (12) layers of vision transformer to bring the best results $( 9 3 . 9 9 \% )$ with 0.28M trainable parameters.
|
| 161 |
+
|
| 162 |
+
The impact of the different insertion locations. Based on the different operations of ViT, we evaluate the impact of the insertion locations of SSF-ADA. We separately remove SSF-ADA after these operations and the results are shown in Table 6b. We find that removing the SSF-ADA in the MLP operation achieves inferior results than removing those in the Attention operation $( 9 3 . 4 6 \%$ vs. $9 3 . 6 9 \%$ with comparable trainable parameters (0.19M vs. 0.21M), which suggests that performing feature modulation for the MLP operation might be more important. Although one can use NAS to search for the importance of different operations and thereby insert SSF-ADA in specific locations, the results might not be better than inserting SSF-ADA in all operations. Therefore, in order to obtain excellent performance, we do not perform NAS but directly insert SSF-ADA into all operations.
|
| 163 |
+
|
| 164 |
+
<table><tr><td>#layers</td><td>Acc.</td><td>Params.</td><td>location</td><td>Acc. Params.</td><td>initialization</td><td>Acc.</td><td></td><td>case</td><td>Acc. Params.</td></tr><tr><td>0</td><td>88.70</td><td>0.08</td><td>w/o. mlp</td><td>93.46 0.19</td><td>random</td><td>[90.11</td><td>w/o. scale</td><td>93.49</td><td>0.18</td></tr><tr><td>2</td><td>92.69</td><td>0.11</td><td>w/o.attn</td><td>93.69 0.21</td><td>constant</td><td>93.91</td><td>w/o. shift</td><td>93.74</td><td>0.18</td></tr><tr><td>4</td><td>93.30</td><td>0.15</td><td>w/o.embed</td><td>93.91 0.28</td><td>uniform</td><td>93.87</td><td>only norm</td><td>93.26</td><td>0.11</td></tr><tr><td>8</td><td>93.60</td><td>0.22</td><td>w/o.norm</td><td>93.80 0.25</td><td>trunc_normal</td><td>93.93</td><td>scalar scale</td><td>93.59</td><td>0.18</td></tr><tr><td>12 (ours)</td><td>93.99</td><td>0.28</td><td>ours</td><td>93.99 0.28</td><td>normal (ours)</td><td>93.99</td><td>ours</td><td>93.99</td><td>0.28</td></tr><tr><td colspan="2">(a)</td><td></td><td>(b)</td><td></td><td>(c)</td><td></td><td></td><td>(d)</td><td></td></tr></table>
|
| 165 |
+
|
| 166 |
+
Table 6: The impacts of different designs. (a) The impact of the number of layers with SSF-ADA. (b) The impacts of the different insertion locations of SSF-ADA. (c) The impacts of initialization. (d) The impacts of different components. Acc.: Top-1 accuracy $( \% )$ ; Params.: parameters (M).
|
| 167 |
+
|
| 168 |
+
The impact of initialization. We also investigate how different ways of initializing the scale and shift factors affect performance in Table 6c. In our experiments, we first randomly initialize both scale and shift parameters with a mean value of zero, but find that the performance is inferior $( 9 0 . 1 1 \% )$ and cannot converge in some experiments. After that, we randomly initialize the scale factor with a mean value of one and find better performance, which implies that the weights of a pre-trained model should not be completely disrupted in the fine-tuning, instead, we should start from this pre-trained model to optimize our model. Experiments show that using the normal initialization achieves the best performance, where the mean values of the scale factor and shift factor are one and zero, respectively.
|
| 169 |
+
|
| 170 |
+
The impact of different components. We also evaluate the impacts of different components in SSF-ADA and the results are shown in Table 6d. We find that removing the scale term yields worse performance than removing the shift term with the same trainable parameters, which shows that the scale term might be more important than the shift term. Also, note that the difference between ‘w/o. scale’ and the ‘Bias’ method in Table 5 is that we fine-tune the model with an additional shift term in ‘w/o. scale’, while ‘Bias’ fine-tunes the model based on the original biases, suggesting that fine-tuning the model in a res-like manner can obtain slightly better performance $( 9 3 . 4 9 \%$ vs. $9 3 . 3 9 \%$ ). We also try to only fine-tune all scale and shift factors in the normalization layer (LN), or fine-tune the model with SSF but set the scale term as a scalar. These experiments yield inferior performance than SSF $9 3 . 2 6 \%$ vs. $9 3 . 9 9 \%$ , $9 3 . 5 9 \%$ vs. $9 3 . 9 9 \%$ ), but could probably be considered as an alternative due to the fact that they only use about half of the trainable parameters of SSF.
|
| 171 |
+
|
| 172 |
+
# 4.4 Performance Comparisons on Robustness and OOD Datasets
|
| 173 |
+
|
| 174 |
+
We also conduct experiments to analyze the robustness and Out-Of-Distribution (OOD) ability of our SSF method with the following datasets: ImageNet-A, ImageNet-R and ImageNet-C. Please refer to Appendix for their details. We perform the robustness and OOD evaluation on these three datasets with the fine-tuned models on ImageNet-1K. All experimental results are listed in Table 7.
|
| 175 |
+
|
| 176 |
+
From this table, we can see that our SSF obtains better performance than VPT and other parameter-efficient fine-tuning methods on three datasets, which shows our fine-tuning method has stronger robust
|
| 177 |
+
|
| 178 |
+
Table 7: Performance comparisons on robustness and out-of-distribution datasets. ‘IN’ means ImageNet. The performance on IN-1K, IN-A and IN-R is evaluated in Top-1 accuracy $( \% )$ . The performance on IN-C is evaluated in mCE (mean corruption error). The lower $( \downarrow )$ , the better.
|
| 179 |
+
|
| 180 |
+
<table><tr><td></td><td>Dataset</td><td rowspan="2">IN-1K (↑)</td><td rowspan="2">IN-A (↑)</td><td rowspan="2">IN-R (↑)</td><td rowspan="2">IN-C (↓)</td></tr><tr><td>Method</td><td></td></tr><tr><td colspan="2">Full fine-tuning</td><td>83.58</td><td>34.49</td><td>51.29</td><td>46.47</td></tr><tr><td colspan="2">Linear probing</td><td>82.04</td><td>33.91</td><td>52.87</td><td>46.91</td></tr><tr><td colspan="2">Adapter [21]</td><td>82.72</td><td>42.21</td><td>54.13</td><td>42.65</td></tr><tr><td colspan="2">Bias [66]</td><td>82.74</td><td>42.12</td><td>55.94</td><td>41.90</td></tr><tr><td colspan="2">VPT-Shallow [29]</td><td>82.08</td><td>30.93</td><td>53.72</td><td>46.88</td></tr><tr><td colspan="2">VPT-Deep [29]</td><td>82.45</td><td>39.10</td><td>53.54</td><td>43.10</td></tr><tr><td colspan="2">SSF (ours)</td><td>83.10</td><td>45.88</td><td>56.77</td><td>41.47</td></tr></table>
|
| 181 |
+
|
| 182 |
+
ness and out-of-distribution generalization. Furthermore, although SSF has lower accuracy than full fine-tuning on ImageNet-1K, the performance on ImageNet-A, ImageNet-R and ImageNet-C is better, which also shows the performance between ImageNet-1K and ImageNet-A/R/C is not absolutely positive relevant. Such improvements in robustness and OOD datasets might come from the fact that SSF freezes most of the pre-trained parameters, which maximally preserves the knowledge learned from the large-scale dataset and thus maintains a better generalization ability.
|
| 183 |
+
|
| 184 |
+
# 4.5 Visualization and Analysis
|
| 185 |
+
|
| 186 |
+
Although our goal is to modulate the features extracted by a pre-trained model, the scale and shift parameters are input-independent indeed. Therefore, these parameters can also be regarded as encoding information of the whole downstream dataset. After re-parameterization, these scale and shift parameters are absorbed into the original model weights. To better understand information learned by the SSF, we visualize the distributions of weights and biases before and after finetuning via SSF in Figure 4a. We can see that the scale and shift parameters adjust the original weights and biases, and change the distribution of weights and biases to fit the downstream task.
|
| 187 |
+
|
| 188 |
+
|
| 189 |
+
Figure 4: Comparisons of parameter distribution between the original pre-trained model and different fine-tuning methods. The first row shows weight distribution and the second row is bias distribution. The blue histograms show the original pre-trained model, and the orange ones show the fine-tuned model via SSF in (a) and full fine-tuned model in (b).
|
| 190 |
+
|
| 191 |
+
As a comparison, we also visualize the original weight distribution and the weight distribution after full fine-tuning in Figure 4b, from which we can find an interesting phenomenon that full fine-tuning does not change the distribution of weights and biases much, but probably only a small portion of the values is changed. It is worth noting that although SSF does not match the weight distribution of full fine-tuning, it achieves better performance $9 3 . 9 9 \%$ vs. $9 3 . 8 2 \%$ in Table 5) on CIFAR-100.
|
| 192 |
+
|
| 193 |
+
To further investigate why SSF can achieve superior performance, beyond weight distribution, we also visualize the feature similarities between full fine-tuning and linear probing, full fine-tuning and VPT-Deep, full fine-tuning and SSF, as shown in Figure 5. In the last layer, SSF has the most similar feature to full fine-tuning and the accuracy is also the closest. This shows that even if the weight distribution learned by SSF is different from full fine-tuning, SSF is also able to extract the features of the images in the downstream task very well, which validates the effectiveness of our method.
|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
Figure 5: The visualization of the feature similarities between full fine-tuning and linear probing, full fine-tuning and VPT-Deep, full finetuning and SSF, in different layers of ViT-B/16.
|
| 197 |
+
|
| 198 |
+
# 5 Conclusion
|
| 199 |
+
|
| 200 |
+
In this paper, we focus on parameter-efficient fine-tuning and propose an SSF method to scale and shift the features extracted by a pre-trained model. The intuition behind our method comes from alleviating the distribution mismatch between upstream tasks and downstream tasks by modulating deep features. SSF surprisingly outperforms other parameter-efficient fine-tuning approaches with a small number of learnable parameters. Besides, the introduced scale and shift parameters during the fine-tuning can be merged into the original pre-trained model weights via re-parameterization in the inference phase, thereby avoiding extra parameters and FLOPs. With the proposed SSF method, our model obtains $2 . 4 6 \%$ $9 0 . 7 2 \%$ vs. $8 8 . 5 4 \%$ ) and $1 1 . 4 8 \%$ $7 3 . 1 0 \%$ vs. $6 5 . 5 7 \%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy compared to the full fine-tuning but only fine-tuning about $0 . 3 { \bf M }$ parameters. Experiments on 26 image classification datasets in total and 3 robustness & out-of-distribution datasets with various model families (CNNs, Transformers, and MLPs) show the effectiveness of SSF, which establishes a new baseline.
|
| 201 |
+
|
| 202 |
+
# Acknowledgement
|
| 203 |
+
|
| 204 |
+
The authors acknowledge the support from the Singapore National Research Foundation (“CogniVision – Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption” grant NRF-CRP20-2017-0003) – www.green-ic.org/ CogniVision. Xinchao Wang is the corresponding author.
|
| 205 |
+
|
| 206 |
+
# References
|
| 207 |
+
|
| 208 |
+
[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.
|
| 209 |
+
[2] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11285–11297. Curran Associates, Inc., 2020.
|
| 210 |
+
[3] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022.
|
| 211 |
+
[4] Ting Chen, Mario Lucic, Neil Houlsby, and Sylvain Gelly. On self modulation for generative adversarial networks. In International Conference on Learning Representations (ICLR), 2019.
|
| 212 |
+
[5] Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems, 34:3965–3977, 2021.
|
| 213 |
+
[6] Harm De Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C Courville. Modulating early visual processing by language. Advances in Neural Information Processing Systems, 30, 2017.
|
| 214 |
+
[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009.
|
| 215 |
+
[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.
|
| 216 |
+
[9] Xiaohan Ding, Chunlong Xia, Xiangyu Zhang, Xiaojie Chu, Jungong Han, and Guiguang Ding. Repmlp: Re-parameterizing convolutions into fully-connected layers for image recognition. arXiv preprint arXiv:2105.01883, 2021.
|
| 217 |
+
[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733–13742, 2021.
|
| 218 |
+
[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021.
|
| 219 |
+
[12] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Finegrained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017.
|
| 220 |
+
[13] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4805–4814, 2019.
|
| 221 |
+
[14] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544–560. Springer, 2020.
|
| 222 |
+
[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000–16009, 2022.
|
| 223 |
+
[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020.
|
| 224 |
+
[17] Kaiming He, Ross Girshick, and Piotr Dollár. Rethinking imagenet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918–4927, 2019.
|
| 225 |
+
[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016.
|
| 226 |
+
[19] Xuehai He, Chunyuan Li, Pengchuan Zhang, Jianwei Yang, and Xin Eric Wang. Parameterefficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329, 2022.
|
| 227 |
+
[20] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. arXiv preprint arXiv:2106.12368, 2021.
|
| 228 |
+
[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790–2799. PMLR, 2019.
|
| 229 |
+
[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, To Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.
|
| 230 |
+
[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021.
|
| 231 |
+
[24] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017.
|
| 232 |
+
[25] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017.
|
| 233 |
+
[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448–456. PMLR, 2015.
|
| 234 |
+
[27] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704–2713, 2018.
|
| 235 |
+
[28] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015.
|
| 236 |
+
[29] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022.
|
| 237 |
+
[30] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR Workshop on Fine-Grained Visual Categorization (FGVC), volume 2. Citeseer, 2011.
|
| 238 |
+
[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009.
|
| 239 |
+
[32] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021.
|
| 240 |
+
[33] Dongze Lian, Zehao Yu, Xing Sun, and Shenghua Gao. As-mlp: An axial shifted mlp architecture for vision. In International Conference on Learning Representations (ICLR), 2022.
|
| 241 |
+
[34] Dongze Lian, Yin Zheng, Yintao Xu, Yanxiong Lu, Leyu Lin, Peilin Zhao, Junzhou Huang, and Shenghua Gao. Towards fast adaptation of neural architectures with meta learning. In International Conference on Learning Representations (ICLR), 2020.
|
| 242 |
+
[35] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations (ICLR), 2019.
|
| 243 |
+
[36] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming catastrophic forgetting in graph neural networks. In AAAI Conference on Artificial Intelligence, 2021.
|
| 244 |
+
[37] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In Proceedings of the European Conference on Computer Vision, 2022.
|
| 245 |
+
[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021.
|
| 246 |
+
[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
|
| 247 |
+
[40] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021.
|
| 248 |
+
[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019.
|
| 249 |
+
[42] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. arXiv preprint arXiv:2106.04647, 2021.
|
| 250 |
+
[43] Rabeeh Karimi Mahabadi, Sebastian Ruder, Mostafa Dehghani, and James Henderson. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. arXiv preprint arXiv:2106.04489, 2021. number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722–729. IEEE, 2008.
|
| 251 |
+
[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018.
|
| 252 |
+
[46] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095– 4104. PMLR, 2018.
|
| 253 |
+
[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021.
|
| 254 |
+
[48] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted selfattention via multi-scale token aggregation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022.
|
| 255 |
+
[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015.
|
| 256 |
+
[50] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016.
|
| 257 |
+
[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (ICML), pages 6105–6114, 2019.
|
| 258 |
+
[52] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. MLP-Mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021.
|
| 259 |
+
[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021.
|
| 260 |
+
[54] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022.
|
| 261 |
+
[55] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595–604, 2015.
|
| 262 |
+
[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017.
|
| 263 |
+
[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011.
|
| 264 |
+
[58] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018.
|
| 265 |
+
[59] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017.
|
| 266 |
+
[60] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In Proceedings of the European Conference on Computer Vision, 2022.
|
| 267 |
+
[61] Yiding Yang, Zunlei Feng, Mingli Song, and Xinchao Wang. Factorizable graph convolutional networks. In Conference on Neural Information Processing Systems, 2020.
|
| 268 |
+
[62] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2020.
|
| 269 |
+
[63] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Proceedings of the European Conference on Computer Vision, 2022.
|
| 270 |
+
[64] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conference on Computer Vision and Pattern Recognition, 2022.
|
| 271 |
+
[65] Sergey Zagoruyko and Nikos Komodakis. Diracnets: Training very deep neural networks without skip-connections. arXiv preprint arXiv:1706.00388, 2017.
|
| 272 |
+
[66] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021.
|
| 273 |
+
[67] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019.
|
| 274 |
+
[68] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022.
|
| 275 |
+
[69] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. DeepViT: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021.
|
| 276 |
+
[70] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Anima Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. arXiv preprint arXiv:2204.12451, 2022.
|
| 277 |
+
[71] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021.
|
| 278 |
+
|
| 279 |
+
# Checklist
|
| 280 |
+
|
| 281 |
+
1. For all authors...
|
| 282 |
+
|
| 283 |
+
(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] See abstract, introduction and experiments.
|
| 284 |
+
(b) Did you describe the limitations of your work? [Yes] See appendix.
|
| 285 |
+
(c) Did you discuss any potential negative societal impacts of your work? [Yes] See appendix.
|
| 286 |
+
(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes]
|
| 287 |
+
|
| 288 |
+
2. If you are including theoretical results...
|
| 289 |
+
|
| 290 |
+
(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A]
|
| 291 |
+
|
| 292 |
+
3. If you ran experiments...
|
| 293 |
+
|
| 294 |
+
(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See abstract and experiments.
|
| 295 |
+
(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See experiments.
|
| 296 |
+
(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [N/A] A part of the experiments is tested several times.
|
| 297 |
+
(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See experiments.
|
| 298 |
+
|
| 299 |
+
4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
|
| 300 |
+
|
| 301 |
+
(a) If your work uses existing assets, did you cite the creators? [Yes]
|
| 302 |
+
(b) Did you mention the license of the assets? [Yes]
|
| 303 |
+
(c) Did you include any new assets either in the supplemental material or as a URL? [Yes]
|
| 304 |
+
(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A]
|
| 305 |
+
(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] See appendix.
|
| 306 |
+
|
| 307 |
+
5. If you used crowdsourcing or conducted research with human subjects...
|
| 308 |
+
|
| 309 |
+
(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A]
|
| 310 |
+
(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A]
|
| 311 |
+
(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A]
|
parse/dev/XtyeppctGgc/XtyeppctGgc_content_list.json
ADDED
|
@@ -0,0 +1,1264 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
235,
|
| 8 |
+
122,
|
| 9 |
+
754,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Dongze Lian1∗ Daquan Zhou1,2∗ Jiashi Feng2 Xinchao Wang1 1National University of Singapore 2ByteDance {dongze,xinchao}@nus.edu.sg {zhoudaquan21,jshfeng}@gmail.com ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
236,
|
| 19 |
+
224,
|
| 20 |
+
761,
|
| 21 |
+
270
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Abstract ",
|
| 28 |
+
"text_level": 1,
|
| 29 |
+
"bbox": [
|
| 30 |
+
462,
|
| 31 |
+
305,
|
| 32 |
+
535,
|
| 33 |
+
321
|
| 34 |
+
],
|
| 35 |
+
"page_idx": 0
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"type": "text",
|
| 39 |
+
"text": "Existing fine-tuning methods either tune all parameters of the pre-trained model (full fine-tuning), which is not efficient, or only tune the last linear layer (linear probing), which suffers a significant accuracy drop compared to the full fine-tuning. In this paper, we propose a new parameter-efficient fine-tuning method termed as SSF, representing that researchers only need to Scale and Shift the deep Features extracted by a pre-trained model to catch up with the performance of full finetuning. In this way, SSF also surprisingly outperforms other parameter-efficient fine-tuning approaches even with a smaller number of tunable parameters. Furthermore, different from some existing parameter-efficient fine-tuning methods (e.g., Adapter or VPT) that introduce the extra parameters and computational cost in the training and inference stages, SSF only adds learnable parameters during the training stage, and these additional parameters can be merged into the original pre-trained model weights via re-parameterization in the inference phase. With the proposed SSF, our model obtains $2 . 4 6 \\%$ $( 9 0 . 7 2 \\%$ vs. $8 8 . 5 4 \\%$ ) and $1 1 . 4 8 \\%$ $7 3 . 1 0 \\%$ vs. $6 5 . 5 7 \\%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy compared to the full fine-tuning but only fine-tuning about $0 . 3 { \\bf M }$ parameters. We also conduct amounts of experiments in various model families (CNNs, Transformers, and MLPs) and datasets. Results on 26 image classification datasets in total and 3 robustness & out-of-distribution datasets show the effectiveness of SSF. Code is available at https://github.com/dongzelian/SSF. ",
|
| 40 |
+
"bbox": [
|
| 41 |
+
232,
|
| 42 |
+
335,
|
| 43 |
+
766,
|
| 44 |
+
612
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "1 Introduction ",
|
| 51 |
+
"text_level": 1,
|
| 52 |
+
"bbox": [
|
| 53 |
+
174,
|
| 54 |
+
635,
|
| 55 |
+
310,
|
| 56 |
+
651
|
| 57 |
+
],
|
| 58 |
+
"page_idx": 0
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "text",
|
| 62 |
+
"text": "With the popularity of the data-driven methods in the deep learning community, the dataset scale and the model size have both got huge explosions. There is a tendency to explore large models and then adopt these pre-trained models in downstream tasks to achieve better performance and faster convergence, which gradually becomes a common way. ",
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
659,
|
| 66 |
+
823,
|
| 67 |
+
714
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "However, the current procedure depends on full fine-tuning heavily, where all the parameters of the model are updated. It inevitably causes the model to be over-fitted to the small target dataset and thus cannot be used for other tasks after the fine-tuning. As a result, the device will need to save a dedicated set of model parameters for each task, which causes a huge amount of storage space, especially for today’s large models (e.g., ViT-G/14 [11] 1.8G, CoAtNet [5] 2.4G). ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
720,
|
| 77 |
+
825,
|
| 78 |
+
790
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "A simple solution for the above problem is linear probing [16], where only the last head layer is fine-tuned. However, this practice usually yields inferior performance compared to the full fine-tuning proxy. Motivated by the success of the parameter-efficient fine-tuning strategy with prompt in the field of natural language processing (NLP) [21, 32, 23, 19], the recent work implements a similar proxy on vision tasks [29], termed as Visual Prompt Tuning (VPT). Specifically, VPT [29] proposes to insert learnable prompts as inputs and append them to the original image tokens. These prompts ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
796,
|
| 88 |
+
825,
|
| 89 |
+
880
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "table",
|
| 95 |
+
"img_path": "images/ec52410c8f12dad9228e3981aabf09684ebe3c36874303a6a9de5b7dd7f839dc.jpg",
|
| 96 |
+
"table_caption": [],
|
| 97 |
+
"table_footnote": [],
|
| 98 |
+
"table_body": "<table><tr><td>Method</td><td>Acc.</td><td>Params. (M)</td><td>Unified parameter space</td><td>No extra inference params.</td></tr><tr><td>Full fine-tuning</td><td>93.82</td><td>85.88</td><td>√</td><td>√</td></tr><tr><td>Linear probing</td><td>88.70</td><td>0.08</td><td>√</td><td>√</td></tr><tr><td>Adapter [21] VPT [29]</td><td>93.34</td><td>0.31</td><td>√</td><td>×</td></tr><tr><td></td><td>93.17</td><td>0.54</td><td>×</td><td>×</td></tr><tr><td>SSF (ours)</td><td>93.99</td><td>0.28</td><td>√</td><td>√</td></tr></table>",
|
| 99 |
+
"bbox": [
|
| 100 |
+
174,
|
| 101 |
+
82,
|
| 102 |
+
468,
|
| 103 |
+
154
|
| 104 |
+
],
|
| 105 |
+
"page_idx": 1
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"type": "text",
|
| 109 |
+
"text": "Table 1: Characteristics of different finetuning methods. Acc. means the Top-1 accuracy $( \\% )$ on CIFAR-100 with a pre-trained ViT-B/16 for tuning. Params. means the learnable parameters at fine-tuning. Our SSF has a unified learnable parameter space and does not require extra inference parameters while obtaining superior performance. ",
|
| 110 |
+
"bbox": [
|
| 111 |
+
174,
|
| 112 |
+
160,
|
| 113 |
+
470,
|
| 114 |
+
270
|
| 115 |
+
],
|
| 116 |
+
"page_idx": 1
|
| 117 |
+
},
|
| 118 |
+
{
|
| 119 |
+
"type": "image",
|
| 120 |
+
"img_path": "images/613ec6be12fb156c003ed7e7f147c704dda8789d3d95f22888e0c4be8b147b49.jpg",
|
| 121 |
+
"image_caption": [
|
| 122 |
+
"Figure 1: Performance comparisons of seven finetuning methods with a pre-trained ViT-B/16 model on the FGVC dataset and VTAB-1k benchmark. Our SSF (red dots) achieves state-of-the-art performance only with about $0 . 3 { \\bf M }$ average learnable parameters. "
|
| 123 |
+
],
|
| 124 |
+
"image_footnote": [],
|
| 125 |
+
"bbox": [
|
| 126 |
+
496,
|
| 127 |
+
80,
|
| 128 |
+
816,
|
| 129 |
+
179
|
| 130 |
+
],
|
| 131 |
+
"page_idx": 1
|
| 132 |
+
},
|
| 133 |
+
{
|
| 134 |
+
"type": "text",
|
| 135 |
+
"text": "will interact with the image tokens by performing self-attention and are updated during the fine-tuning process. In this manner, a significant performance improvement can be achieved in downstream tasks compared to a linear probing proxy. Nevertheless, compared to the full fine-tuning and linear probing, it additionally raises two issues: i) VPT tunes the number of prompts for different tasks, which introduces a task-dependent learnable parameter space. The fine-tuning performance is sensitive to the number of prompts for each task and needs to be carefully designed. Too few or too many prompts might either degrade the accuracy of fine-tuning or increase the redundancy of the computation (e.g., 200 prompts on Clevr/count vs. 1 prompt on Flowers102); ii) VPT [29], as well as other Adapter-based methods [21, 42], introduces additional parameters and computational cost in the inference phase compared to the original pre-trained model. For instance, VPT introduces additional inputs for self-attention with image tokens. Adapter-based methods insert additional modules into the pre-trained model. These methods change the specific backbone architecture or the input of the network, which might result in frequent structure modifications and heavy workload, especially for those models that are already deployed in edge devices (e.g., mobile phones). ",
|
| 136 |
+
"bbox": [
|
| 137 |
+
174,
|
| 138 |
+
276,
|
| 139 |
+
825,
|
| 140 |
+
470
|
| 141 |
+
],
|
| 142 |
+
"page_idx": 1
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"type": "text",
|
| 146 |
+
"text": "To cope with the above issues, we attempt to find a general proxy for parameter-efficient finetuning, where the learnable parameter space is unified (task-independent) and no additional inference parameters are introduced. Inspired by some feature modulation methods [58, 25, 45], we propose a new parameter-efficient fine-tuning method named SSF, where you only need to Scale and Shift your deep Features extracted by a pre-trained model for fine-tuning. The intuition behind our approach come from the fact that the upstream datasets and downstream datasets have different data distributions [50]. Therefore, it is difficult to apply the model weights trained in the upstream dataset to the downstream dataset. For instance, a naive linear probing strategy with keeping the weights of backbone frozen will cause performance degradation. To alleviate the above problem, SSF introduces scale parameters and shift parameters, which could be considered as variance and mean to modulate the features of the downstream dataset extracted with the pre-trained model on the upstream dataset, such that the modulated feature falls in a discriminative space. These scale parameters and shift parameters do not depend on any input and have a unified learnable parameter space for different tasks. Another advantage of SSF is that it only introduces linear transformations because we scale and shift the extracted features. These linear transformations could be further merged into the original pre-trained weight via model re-parameterization [10] in the inference phase, thus avoiding the extra parameters and FLOPs for downstream tasks. For a deployed model in edge devices, only the updated weights after fine-tuning need to be uploaded instead of changing the backbone architecture. Table 1 shows the specific characteristics comparisons between SSF and other fine-tuning methods. SSF is simple, effective, and efficient, which also conforms to Occam’s Razor principle. Therefore, we explore this new baseline and find that it surprisingly outperforms all other parameter-efficient fine-tuning methods. ",
|
| 147 |
+
"bbox": [
|
| 148 |
+
174,
|
| 149 |
+
477,
|
| 150 |
+
825,
|
| 151 |
+
780
|
| 152 |
+
],
|
| 153 |
+
"page_idx": 1
|
| 154 |
+
},
|
| 155 |
+
{
|
| 156 |
+
"type": "text",
|
| 157 |
+
"text": "We evaluate our method on 26 classification datasets in total and 3 robustness & out-of-distribution datasets. SSF obtains state-of-the-art performance compared to other parameter-efficient fine-tuning methods with the trainable parameters and accuracy trade-off (Table 1 and Figure 1). Compared to the full fine-tuning, our method obtains $2 . 4 6 \\%$ $9 0 . 7 2 \\%$ vs. $8 8 . 5 4 \\%$ ) and $1 1 . 4 8 \\%$ ( $7 3 . 1 0 \\%$ vs. $6 5 . 5 7 \\%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy but only with about $0 . 3 { \\bf M }$ trainable parameters. Furthermore, our SSF does not require additional parameters during the inference phase. It is plug-and-play and is very easy to extend to various model families (CNNs, Transformers, and MLPs). Our SSF establishes a new baseline and we hope that it brings more insight into the field of the efficient model tuning. ",
|
| 158 |
+
"bbox": [
|
| 159 |
+
174,
|
| 160 |
+
786,
|
| 161 |
+
825,
|
| 162 |
+
911
|
| 163 |
+
],
|
| 164 |
+
"page_idx": 1
|
| 165 |
+
},
|
| 166 |
+
{
|
| 167 |
+
"type": "text",
|
| 168 |
+
"text": "2 Related Work ",
|
| 169 |
+
"text_level": 1,
|
| 170 |
+
"bbox": [
|
| 171 |
+
174,
|
| 172 |
+
89,
|
| 173 |
+
321,
|
| 174 |
+
106
|
| 175 |
+
],
|
| 176 |
+
"page_idx": 2
|
| 177 |
+
},
|
| 178 |
+
{
|
| 179 |
+
"type": "text",
|
| 180 |
+
"text": "2.1 Model Families ",
|
| 181 |
+
"text_level": 1,
|
| 182 |
+
"bbox": [
|
| 183 |
+
174,
|
| 184 |
+
116,
|
| 185 |
+
320,
|
| 186 |
+
131
|
| 187 |
+
],
|
| 188 |
+
"page_idx": 2
|
| 189 |
+
},
|
| 190 |
+
{
|
| 191 |
+
"type": "text",
|
| 192 |
+
"text": "Convolution has been used for a long time as the main module to extract the image features in computer vision tasks, and CNN-based architectures have been studied [49, 18, 59, 39, 60, 37, 63] with extension on graph-based data [62, 61, 36]. Recently, another architecture family, Transformer, has gained widespread attention owing to its great success in NLP [56, 8, 23]. Following this direction, Dosovitskiy et al. [11] first employ a transformer in the domain of computer vision and introduce a new architecture paradigm, ViT, which achieves promising results [64, 48]. Subsequently, various transformer-based models, such as DeiT [53] and Swin Transformer [38], are introduced and shown to be effective on a variety of tasks such as object detection, semantic segmentation, action recognition [40], etc. In another line, Tolstikhin et al. [52] propose a pure MLP-based architecture, and subsequent papers [20, 33] have interestingly demonstrated that the MLP-based architectures can catch up to transformers. However, in addition to the well-designed modules, their excellent performance is also attributed to the deployment of large-scale models. Given a large-scale model pre-trained on a large dataset, how to perform parameter-efficient fine-tuning in downstream tasks is essential but is currently less explored. In this paper, we propose SSF as a new baseline and show its promising performance with comprehensive validation in a wide variety of tasks. ",
|
| 193 |
+
"bbox": [
|
| 194 |
+
173,
|
| 195 |
+
137,
|
| 196 |
+
825,
|
| 197 |
+
343
|
| 198 |
+
],
|
| 199 |
+
"page_idx": 2
|
| 200 |
+
},
|
| 201 |
+
{
|
| 202 |
+
"type": "text",
|
| 203 |
+
"text": "2.2 Pre-training and Fine-tuning ",
|
| 204 |
+
"text_level": 1,
|
| 205 |
+
"bbox": [
|
| 206 |
+
176,
|
| 207 |
+
358,
|
| 208 |
+
415,
|
| 209 |
+
373
|
| 210 |
+
],
|
| 211 |
+
"page_idx": 2
|
| 212 |
+
},
|
| 213 |
+
{
|
| 214 |
+
"type": "text",
|
| 215 |
+
"text": "Early models [18, 24, 22, 59, 51] are usually pre-trained on the ImageNet-1K dataset, and then fine-tuned on downstream tasks to achieve faster convergence [17] or better performance. Such a procedure is called pre-training and fine-tuning, or transfer learning. Recent works tend to employ larger models (e.g., ViT [11] and Swin Transformer V2 [38]) and train them on larger datasets (e.g., ImageNet-21K and JFT-300M) in pursuit of better performance. Both in the domains of NLP and computer vision, these large models [8, 38, 47, 15, 69, 70] achieve enormous performance improvements compared to the small-scale models and provide pre-trained weights for downstream tasks. Some other works attempt to explore how to efficiently fine-tune the pre-trained models [13, 71] on the target tasks. For instance, given a target task, SpotTune [13] investigates which layers need to be fine-tuned. Touvron et al. [54] find that fine-tuning the weights of the attention layers and freezing weights of the other parts is sufficient to adapt the vision transformers to other downstream tasks. Some works also propose to insert adapters into the network to fine-tune in a parameter-efficient way. These adapters can be a small non-linear network [21], a hyper-network that generates model weights [43], or a compactor [42] which performs a low-rank decomposition to reduce the parameters. Some works have also tried to only update the bias term [2, 66]. More recently, VPT [29] proposes to insert a small number of learnable parameters (prompts) and optimize them while freezing the backbone, which achieves significant performance improvement compared to the full fine-tuning. During the submission of this work, some methods [3, 68] are also proposed for parameter-efficient fine-tuning, e.g., inserting a adapter module or neural prompt search. Different from all the above works, we propose to scale and shift deep features extracted by a pre-trained model, which is simple but effective and outperforms other parameter-efficient fine-tuning methods. ",
|
| 216 |
+
"bbox": [
|
| 217 |
+
173,
|
| 218 |
+
378,
|
| 219 |
+
825,
|
| 220 |
+
669
|
| 221 |
+
],
|
| 222 |
+
"page_idx": 2
|
| 223 |
+
},
|
| 224 |
+
{
|
| 225 |
+
"type": "text",
|
| 226 |
+
"text": "2.3 Feature Modulation ",
|
| 227 |
+
"text_level": 1,
|
| 228 |
+
"bbox": [
|
| 229 |
+
174,
|
| 230 |
+
684,
|
| 231 |
+
351,
|
| 232 |
+
698
|
| 233 |
+
],
|
| 234 |
+
"page_idx": 2
|
| 235 |
+
},
|
| 236 |
+
{
|
| 237 |
+
"type": "text",
|
| 238 |
+
"text": "Many works have attempted to modulate features to obtain better performance. The most relevant ones to our work are various normalization methods [26, 1, 58]. BN, LN, and GN usually normalize the features and then transform them linearly with scale and shift factors to modulate feature distribution, which has been verified to be effective in amounts of tasks. STN [28] introduces a learnable module to spatially transform feature maps. In the field of image generation, AdaIN [25] generates scale and shift factors to characterize specific image styles. Self-modulation [4] shows GANs benefit from self-modulation layers in the generator. In vision-language tasks, Conditional BN [6] and FiLM [45] are often utilized to modulate the features of two modalities. Unlike some algorithms such as BN, our SSF is not limited to the modulation of normalization layer, and it has a different motivation that is to alleviate the distribution mismatch between upstream tasks and downstream tasks for parameter-efficient fine-tuning. As a comparison, we also conduct experiments in Sec. 4.3 and show that our SSF is more effective compared to only tuning the normalization layer. Compared to STN, AdaIN, FiLM and so on, our method is input-independent and these scale and shift parameters model the distribution of the whole dataset so that they can be absorbed into the original pre-trained model weights in the inference phase. ",
|
| 239 |
+
"bbox": [
|
| 240 |
+
174,
|
| 241 |
+
705,
|
| 242 |
+
825,
|
| 243 |
+
911
|
| 244 |
+
],
|
| 245 |
+
"page_idx": 2
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"type": "text",
|
| 249 |
+
"text": "2.4 Model Re-parameterization ",
|
| 250 |
+
"text_level": 1,
|
| 251 |
+
"bbox": [
|
| 252 |
+
174,
|
| 253 |
+
90,
|
| 254 |
+
406,
|
| 255 |
+
106
|
| 256 |
+
],
|
| 257 |
+
"page_idx": 3
|
| 258 |
+
},
|
| 259 |
+
{
|
| 260 |
+
"type": "text",
|
| 261 |
+
"text": "Model re-parameterization has been a common practice to improve inference efficiency. One of the representative techniques is batch normalization folding used in the model compression algorithms [27]. The parameters introduced by the batch normalization layers [26] are merged into the convolutional layers usually stacked before them. This technique is further utilized to merge different branches of networks into a new branch [65, 10, 9]. Similarly, our SSF fully adopts linear transformations, which allows the scale and shift parameters in the training phase to be merged into the original pre-trained model weights, thus avoiding the introduction of the extra parameters and computational cost during the inference phase. ",
|
| 262 |
+
"bbox": [
|
| 263 |
+
173,
|
| 264 |
+
109,
|
| 265 |
+
825,
|
| 266 |
+
222
|
| 267 |
+
],
|
| 268 |
+
"page_idx": 3
|
| 269 |
+
},
|
| 270 |
+
{
|
| 271 |
+
"type": "text",
|
| 272 |
+
"text": "3 Approach ",
|
| 273 |
+
"text_level": 1,
|
| 274 |
+
"bbox": [
|
| 275 |
+
173,
|
| 276 |
+
234,
|
| 277 |
+
289,
|
| 278 |
+
252
|
| 279 |
+
],
|
| 280 |
+
"page_idx": 3
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"type": "text",
|
| 284 |
+
"text": "3.1 Preliminaries ",
|
| 285 |
+
"text_level": 1,
|
| 286 |
+
"bbox": [
|
| 287 |
+
174,
|
| 288 |
+
258,
|
| 289 |
+
307,
|
| 290 |
+
273
|
| 291 |
+
],
|
| 292 |
+
"page_idx": 3
|
| 293 |
+
},
|
| 294 |
+
{
|
| 295 |
+
"type": "text",
|
| 296 |
+
"text": "Transformers. In a vision transformer (ViT) [11], an RGB image $I \\in \\mathbb { R } ^ { 3 \\times H \\times W }$ is divided into $N \\times N$ non-overlapping patches, and then these image patches appended a class token are fed into an embedding layer followed by the $L$ -layer vision transformer blocks with self-attention as the core operation. The input $\\boldsymbol { x } \\in \\mathbb { R } ^ { ( N ^ { \\hat { 2 } } + 1 ) \\times d }$ , where $d$ is the embedding dimension, is first transformed to keys $K \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ , values $V \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ , and queries $\\bar { Q ^ { \\prime } } \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ . After that, we can calculate a global self-attention by ",
|
| 297 |
+
"bbox": [
|
| 298 |
+
173,
|
| 299 |
+
276,
|
| 300 |
+
825,
|
| 301 |
+
367
|
| 302 |
+
],
|
| 303 |
+
"page_idx": 3
|
| 304 |
+
},
|
| 305 |
+
{
|
| 306 |
+
"type": "equation",
|
| 307 |
+
"img_path": "images/22fd4a15fd2492e996f4c48ced92c1f1290c47c4edc37fca2fc6f970f09877a1.jpg",
|
| 308 |
+
"text": "$$\n\\mathrm { A t t e n t i o n } ( Q , K , V ) = \\mathrm { S o f t m a x } ( \\frac { Q K ^ { T } } { \\sqrt { d } } ) V .\n$$",
|
| 309 |
+
"text_format": "latex",
|
| 310 |
+
"bbox": [
|
| 311 |
+
352,
|
| 312 |
+
375,
|
| 313 |
+
645,
|
| 314 |
+
410
|
| 315 |
+
],
|
| 316 |
+
"page_idx": 3
|
| 317 |
+
},
|
| 318 |
+
{
|
| 319 |
+
"type": "text",
|
| 320 |
+
"text": "The output of the attention layer will be fed to a two-layer MLP to extract information in the channel dimension. ",
|
| 321 |
+
"bbox": [
|
| 322 |
+
171,
|
| 323 |
+
415,
|
| 324 |
+
823,
|
| 325 |
+
445
|
| 326 |
+
],
|
| 327 |
+
"page_idx": 3
|
| 328 |
+
},
|
| 329 |
+
{
|
| 330 |
+
"type": "text",
|
| 331 |
+
"text": "Adapter. Adapter [21] is inserted into the transformer layer for efficient fine-tuning. It is a bottleneck module with a few trainable parameters, which contains a down-projection to reduce the feature dimension, a non-linear activation function, and an up-projection to project back to the original dimension. Therefore, given the input $\\boldsymbol { x } \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ , the output is calculated by ",
|
| 332 |
+
"bbox": [
|
| 333 |
+
173,
|
| 334 |
+
450,
|
| 335 |
+
825,
|
| 336 |
+
510
|
| 337 |
+
],
|
| 338 |
+
"page_idx": 3
|
| 339 |
+
},
|
| 340 |
+
{
|
| 341 |
+
"type": "equation",
|
| 342 |
+
"img_path": "images/b198aeac61967bae8600c9d29d97c47446088408057d764d7798ad4a15fc62be.jpg",
|
| 343 |
+
"text": "$$\n\\mathrm { o u t } = [ W ^ { \\mathrm { u p } } \\phi ( W ^ { \\mathrm { d o w n } } x ^ { T } ) ] ^ { T } ,\n$$",
|
| 344 |
+
"text_format": "latex",
|
| 345 |
+
"bbox": [
|
| 346 |
+
400,
|
| 347 |
+
517,
|
| 348 |
+
596,
|
| 349 |
+
536
|
| 350 |
+
],
|
| 351 |
+
"page_idx": 3
|
| 352 |
+
},
|
| 353 |
+
{
|
| 354 |
+
"type": "text",
|
| 355 |
+
"text": "where $W ^ { \\mathrm { d o w n } } \\in \\mathbb { R } ^ { d ^ { \\prime } \\times d }$ (where $d ^ { \\prime } \\ll d ,$ ), $\\phi$ , and $W ^ { \\mathrm { u p } } \\in \\mathbb { R } ^ { d \\times d ^ { \\prime } }$ represent the down-projection matrix, non-linear function, and up-projection matrix, respectively. ",
|
| 356 |
+
"bbox": [
|
| 357 |
+
176,
|
| 358 |
+
544,
|
| 359 |
+
825,
|
| 360 |
+
575
|
| 361 |
+
],
|
| 362 |
+
"page_idx": 3
|
| 363 |
+
},
|
| 364 |
+
{
|
| 365 |
+
"type": "text",
|
| 366 |
+
"text": "VPT. VPT [29] inserts some learnable parameters (i.e., prompts) into the input space after the embedding layer. These prompts interact with the original image tokens by performing self-attention. During the fine-tuning, the weights of the backbone network are kept frozen and only the parameters of the prompts are updated. VPT-Shallow inserts prompts in the first layer while VPT-Deep inserts prompts in all the layers of the transformer. Assuming that the input is $\\overline { { x } } \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ , denote the inserted prompts as $\\dot { \\boldsymbol { p } } \\in \\mathbb { R } ^ { n \\times d }$ , where $n$ is the number of prompts, the combined tokens $x ^ { \\prime }$ is ",
|
| 367 |
+
"bbox": [
|
| 368 |
+
173,
|
| 369 |
+
580,
|
| 370 |
+
825,
|
| 371 |
+
667
|
| 372 |
+
],
|
| 373 |
+
"page_idx": 3
|
| 374 |
+
},
|
| 375 |
+
{
|
| 376 |
+
"type": "equation",
|
| 377 |
+
"img_path": "images/b0e1574240223ff1889e89fdfd1c77a38a38fa0216e95265f2765c302876dda9.jpg",
|
| 378 |
+
"text": "$$\nx ^ { \\prime } = [ x ; p ] ,\n$$",
|
| 379 |
+
"text_format": "latex",
|
| 380 |
+
"bbox": [
|
| 381 |
+
460,
|
| 382 |
+
674,
|
| 383 |
+
537,
|
| 384 |
+
691
|
| 385 |
+
],
|
| 386 |
+
"page_idx": 3
|
| 387 |
+
},
|
| 388 |
+
{
|
| 389 |
+
"type": "text",
|
| 390 |
+
"text": "where $x ^ { \\prime } \\in \\mathbb { R } ^ { ( N ^ { 2 } + n + 1 ) \\times d }$ will be fed into the transformer block for self-attention (Eq. (1)). ",
|
| 391 |
+
"bbox": [
|
| 392 |
+
173,
|
| 393 |
+
702,
|
| 394 |
+
771,
|
| 395 |
+
718
|
| 396 |
+
],
|
| 397 |
+
"page_idx": 3
|
| 398 |
+
},
|
| 399 |
+
{
|
| 400 |
+
"type": "text",
|
| 401 |
+
"text": "3.2 Scaling and Shifting Your Features for Fine-tuning ",
|
| 402 |
+
"text_level": 1,
|
| 403 |
+
"bbox": [
|
| 404 |
+
173,
|
| 405 |
+
727,
|
| 406 |
+
566,
|
| 407 |
+
742
|
| 408 |
+
],
|
| 409 |
+
"page_idx": 3
|
| 410 |
+
},
|
| 411 |
+
{
|
| 412 |
+
"type": "text",
|
| 413 |
+
"text": "Different from the above methods, we introduce both the scale and shift factors to modulate deep features extracted by a pre-trained model with linear transformation to match the distribution of a target dataset, as mentioned in Sec. 1. Five main properties are covered in our method: i) SSF achieves on-par performance with the full fine-tuning strategy; ii) all downstream tasks can be inputted to the model independently without relying on any other task; iii) the model only needs to fine-tune very few parameters; iv) unlike VPT [29], which adjusts the number of prompts for each task, the set of parameters for fine-tuning in SSF does not change as the task changes, making it feasible to further fine-tune the parameters later by adding more tasks for multi-task learning or continuous learning2; v) thanks to the linear transformation, SSF avoids the introduction of the extra parameters and computational cost during the inference phase, making our method zero overhead. ",
|
| 414 |
+
"bbox": [
|
| 415 |
+
173,
|
| 416 |
+
746,
|
| 417 |
+
826,
|
| 418 |
+
886
|
| 419 |
+
],
|
| 420 |
+
"page_idx": 3
|
| 421 |
+
},
|
| 422 |
+
{
|
| 423 |
+
"type": "text",
|
| 424 |
+
"text": "The design of SSF. SSF performs the linear transformation to modulate the features for parameter-efficient fine-tuning as shown in Figure 2. In Figure 2 (a), given a model pre-trained in the upstream task, we insert SSF-ADA3 after each operation (OP) of the network to modulate features. There are $K$ OPs in total and these operations might contain multi-head self-attention (MSA), MLP and layer normalization (LN), etc. During the fine-tuning, the pre-trained weights in these operations are kept frozen and the SSFADA parameters are kept updated. The specific SSF-ADA structure is shown in Figure2 (c), where the features output from the previous operation are performed dot product with a scale factor and then summed with a shift factor, which are input-independent. Formally, given the input $\\boldsymbol { x } \\in \\bar { \\mathbb { R } } ^ { ( N ^ { 2 } + 1 ) \\times d }$ , the output $\\boldsymbol { y } \\in \\mathbb { R } ^ { ( N ^ { 2 } + 1 ) \\times d }$ (is also the input of the next operation) is calculated by ",
|
| 425 |
+
"bbox": [
|
| 426 |
+
174,
|
| 427 |
+
90,
|
| 428 |
+
485,
|
| 429 |
+
369
|
| 430 |
+
],
|
| 431 |
+
"page_idx": 4
|
| 432 |
+
},
|
| 433 |
+
{
|
| 434 |
+
"type": "image",
|
| 435 |
+
"img_path": "images/247e0c69b62a1b61b9db9dfd09e25ffebf382c56ff9d7c30d74421196ce8672e.jpg",
|
| 436 |
+
"image_caption": [
|
| 437 |
+
"Figure 2: The overall pipeline of SSF. (a) Training pipeline via SSF, where an OP means an operation, e.g., MSA, MLP or LN. (b) A pre-trained model or inference pipeline. (c) Our SSF-ADA. "
|
| 438 |
+
],
|
| 439 |
+
"image_footnote": [],
|
| 440 |
+
"bbox": [
|
| 441 |
+
504,
|
| 442 |
+
94,
|
| 443 |
+
818,
|
| 444 |
+
284
|
| 445 |
+
],
|
| 446 |
+
"page_idx": 4
|
| 447 |
+
},
|
| 448 |
+
{
|
| 449 |
+
"type": "equation",
|
| 450 |
+
"img_path": "images/64be1b33c811ee487fedc0135689eb3a5e2334d6987fb3e0281f3a0ce6ae326b.jpg",
|
| 451 |
+
"text": "$$\ny = \\gamma \\odot x + \\beta ,\n$$",
|
| 452 |
+
"text_format": "latex",
|
| 453 |
+
"bbox": [
|
| 454 |
+
444,
|
| 455 |
+
372,
|
| 456 |
+
552,
|
| 457 |
+
388
|
| 458 |
+
],
|
| 459 |
+
"page_idx": 4
|
| 460 |
+
},
|
| 461 |
+
{
|
| 462 |
+
"type": "text",
|
| 463 |
+
"text": "where $\\gamma \\in \\mathbb { R } ^ { d }$ and $\\beta \\in \\mathbb { R } ^ { d }$ are the scale and shift factors, respectively. $\\odot$ is the dot product. ",
|
| 464 |
+
"bbox": [
|
| 465 |
+
166,
|
| 466 |
+
397,
|
| 467 |
+
771,
|
| 468 |
+
412
|
| 469 |
+
],
|
| 470 |
+
"page_idx": 4
|
| 471 |
+
},
|
| 472 |
+
{
|
| 473 |
+
"type": "text",
|
| 474 |
+
"text": "Re-parameterization. Since SSF-ADA is a completely linear transformation, we can re-parameterize it by absorbing the scale and shift terms into the previous linear layer as follows ",
|
| 475 |
+
"bbox": [
|
| 476 |
+
169,
|
| 477 |
+
417,
|
| 478 |
+
823,
|
| 479 |
+
446
|
| 480 |
+
],
|
| 481 |
+
"page_idx": 4
|
| 482 |
+
},
|
| 483 |
+
{
|
| 484 |
+
"type": "equation",
|
| 485 |
+
"img_path": "images/ce8d9e3a4355666b2f3c37b2d3f46c1817891040b22b9e4ab21b433873324997.jpg",
|
| 486 |
+
"text": "$$\ny = \\gamma \\odot x + \\beta = \\gamma \\odot ( w * t + b ) + \\beta = ( \\gamma \\odot w ) * t + \\gamma \\odot b + \\beta ,\n$$",
|
| 487 |
+
"text_format": "latex",
|
| 488 |
+
"bbox": [
|
| 489 |
+
276,
|
| 490 |
+
455,
|
| 491 |
+
720,
|
| 492 |
+
473
|
| 493 |
+
],
|
| 494 |
+
"page_idx": 4
|
| 495 |
+
},
|
| 496 |
+
{
|
| 497 |
+
"type": "text",
|
| 498 |
+
"text": "where $w$ and $b$ are the weight and bias terms, respectively. $^ *$ represents the ‘convolution’ operation in the convolutional layer or the ‘multiplication’ operation in the MLP layer. $t$ is the input of the previous linear layer. Since $w$ and $b$ are frozen and $\\gamma$ and $\\beta$ are updated in the fine-tuning, $\\gamma$ and $\\beta$ can be merged into the original parameter space ( $\\dot { } w$ and $b$ ) in the inference stage through the above formulation. From this perspective, our SSF-ADA makes it possible to perform downstream tasks without adding any extra parameters and computational costs, as shown in Figure2 (b). ",
|
| 499 |
+
"bbox": [
|
| 500 |
+
174,
|
| 501 |
+
482,
|
| 502 |
+
825,
|
| 503 |
+
566
|
| 504 |
+
],
|
| 505 |
+
"page_idx": 4
|
| 506 |
+
},
|
| 507 |
+
{
|
| 508 |
+
"type": "text",
|
| 509 |
+
"text": "Discussion. The first question is why we want the input $\\gamma$ and $\\beta$ to be input-independent. As FiLM [45] and AdaIN [25] show, we could obtain $\\gamma$ and $\\beta$ by conditioning an image sample, however, this might cause two shortcomings. One is that we want $\\gamma$ and $\\beta$ to be input-independent to represent the distribution of the whole downstream dataset so that we can modify the previous weight distribution to fit the downstream dataset by modulating the feature. Secondly, the conditional input requires the introduction of some additional networks (e.g., MLPs) to generate $\\gamma$ and $\\beta$ , which introduces more trainable parameters. More importantly, to better generate $\\gamma$ and $\\beta$ , a non-linear activation function might be required, which will lead to the intractability of the re-parameterization. Therefore, we directly perform a fully linear transformation to merge the $\\gamma$ and $\\beta$ factors into the original pre-trained weights, so that weights can be easily uploaded to the edge devices without any modification of the backbone architecture. ",
|
| 510 |
+
"bbox": [
|
| 511 |
+
173,
|
| 512 |
+
571,
|
| 513 |
+
825,
|
| 514 |
+
724
|
| 515 |
+
],
|
| 516 |
+
"page_idx": 4
|
| 517 |
+
},
|
| 518 |
+
{
|
| 519 |
+
"type": "text",
|
| 520 |
+
"text": "The second question is which operations should be followed by SSF-ADA. Our experience is that you can insert SSF-ADA after each operation with a linear coefficient in ViT. Although we can search for some optimal layers or operations with Neural Architecture Search (NAS) [46, 35, 14, 34], to reduce the number of the trainable parameters, we believe that our method will produce better results (or not worse than NAS) without introducing too many trainable parameters that can be merged for inference, as will be shown in Sec. 4.3. ",
|
| 521 |
+
"bbox": [
|
| 522 |
+
173,
|
| 523 |
+
731,
|
| 524 |
+
825,
|
| 525 |
+
814
|
| 526 |
+
],
|
| 527 |
+
"page_idx": 4
|
| 528 |
+
},
|
| 529 |
+
{
|
| 530 |
+
"type": "text",
|
| 531 |
+
"text": "3.3 Complexity Analysis ",
|
| 532 |
+
"text_level": 1,
|
| 533 |
+
"bbox": [
|
| 534 |
+
174,
|
| 535 |
+
827,
|
| 536 |
+
356,
|
| 537 |
+
842
|
| 538 |
+
],
|
| 539 |
+
"page_idx": 4
|
| 540 |
+
},
|
| 541 |
+
{
|
| 542 |
+
"type": "text",
|
| 543 |
+
"text": "We also compare the complexity of Adapter, VPT and our SSF. Take a ViT as an example, the dimension and number of the tokens are $d$ and $N ^ { 2 }$ . Assuming that Adapter projects features from $d$ -dim to $d ^ { \\prime }$ -dim (where $d ^ { \\prime } \\ll d _ { , }$ ) so that the extra trainable parameters are $\\bar { 2 } d d ^ { \\prime }$ in each layer, ",
|
| 544 |
+
"bbox": [
|
| 545 |
+
176,
|
| 546 |
+
847,
|
| 547 |
+
825,
|
| 548 |
+
888
|
| 549 |
+
],
|
| 550 |
+
"page_idx": 4
|
| 551 |
+
},
|
| 552 |
+
{
|
| 553 |
+
"type": "text",
|
| 554 |
+
"text": "VPT inserts $n$ prompts to obtain nd extra parameters in each layer, and SSF inserts SSF-ADA after each operation with a linear coefficient to obtain md extra parameters in each layer, when the total number of layers is $L$ , the complexity of Adapter, VPT and SSF is shown in Table 2. The specific number of additional parameters used by ",
|
| 555 |
+
"bbox": [
|
| 556 |
+
174,
|
| 557 |
+
92,
|
| 558 |
+
423,
|
| 559 |
+
215
|
| 560 |
+
],
|
| 561 |
+
"page_idx": 5
|
| 562 |
+
},
|
| 563 |
+
{
|
| 564 |
+
"type": "table",
|
| 565 |
+
"img_path": "images/a98f7a5d7ba722e86a8e856faee92396d6b1395da9529bb0bc9dfc92aa178c84.jpg",
|
| 566 |
+
"table_caption": [
|
| 567 |
+
"Table 2: The complexity comparisons of Adapter [21], VPT [29] and our SSF. ‘ $( 1 ) ^ { \\prime }$ : the same parameters and FLOPs for training and inference; $\\mathbf { \\eta } ^ { \\mathrm { ( 0 ) } } \\mathbf { : }$ no additional parameters and FLOPs are required for inference. "
|
| 568 |
+
],
|
| 569 |
+
"table_footnote": [],
|
| 570 |
+
"table_body": "<table><tr><td>Method</td><td>Adapter</td><td>VPT-Shallow</td><td>VPT-Deep</td><td>SSF (ours)</td></tr><tr><td># Extra Params.</td><td>2Ldd' (1)</td><td>nd(1)</td><td>nLd(1)</td><td>mLd (0)</td></tr><tr><td>#Extra FLOPs</td><td>2N² Ldd' (1)</td><td></td><td>2n(2N² +n)d(1) | 2n(2N²+n)Ld(1)</td><td>mN²Ld (0)</td></tr></table>",
|
| 571 |
+
"bbox": [
|
| 572 |
+
437,
|
| 573 |
+
94,
|
| 574 |
+
828,
|
| 575 |
+
136
|
| 576 |
+
],
|
| 577 |
+
"page_idx": 5
|
| 578 |
+
},
|
| 579 |
+
{
|
| 580 |
+
"type": "text",
|
| 581 |
+
"text": "Adapter, VPT and SSF depends on the values of $d ^ { \\prime }$ , $n$ and $m$ . However, in practice, SSF outperforms Adapter and VPT-Deep even with slightly fewer parameters in the training stage as we will see in Sec. 4. Further, in the inference stage, borrowing the model re-parameterization strategy, the extra parameters and FLOPs of SSF are zero. However, the complexity of Adapter and VPT remain the same compared to the training, which establishes the strengths of our approach. ",
|
| 582 |
+
"bbox": [
|
| 583 |
+
174,
|
| 584 |
+
215,
|
| 585 |
+
825,
|
| 586 |
+
285
|
| 587 |
+
],
|
| 588 |
+
"page_idx": 5
|
| 589 |
+
},
|
| 590 |
+
{
|
| 591 |
+
"type": "text",
|
| 592 |
+
"text": "4 Experiments ",
|
| 593 |
+
"text_level": 1,
|
| 594 |
+
"bbox": [
|
| 595 |
+
174,
|
| 596 |
+
300,
|
| 597 |
+
312,
|
| 598 |
+
318
|
| 599 |
+
],
|
| 600 |
+
"page_idx": 5
|
| 601 |
+
},
|
| 602 |
+
{
|
| 603 |
+
"type": "text",
|
| 604 |
+
"text": "4.1 Experimental Settings ",
|
| 605 |
+
"text_level": 1,
|
| 606 |
+
"bbox": [
|
| 607 |
+
174,
|
| 608 |
+
327,
|
| 609 |
+
367,
|
| 610 |
+
342
|
| 611 |
+
],
|
| 612 |
+
"page_idx": 5
|
| 613 |
+
},
|
| 614 |
+
{
|
| 615 |
+
"type": "text",
|
| 616 |
+
"text": "Datasets. We mainly conduct our experiments on a series of datasets that can be categorized into three types as detailed below: ",
|
| 617 |
+
"bbox": [
|
| 618 |
+
173,
|
| 619 |
+
345,
|
| 620 |
+
823,
|
| 621 |
+
373
|
| 622 |
+
],
|
| 623 |
+
"page_idx": 5
|
| 624 |
+
},
|
| 625 |
+
{
|
| 626 |
+
"type": "text",
|
| 627 |
+
"text": "FGVC. Following VPT [29], we employ five Fine-Grained Visual Classification (FGVC) datasets to evaluate the effectiveness of our proposed SSF, which consists of CUB-200-2011 [57], NABirds [55], Oxford Flowers [44], Stanford Dogs [30] and Stanford Cars [12]. ",
|
| 628 |
+
"bbox": [
|
| 629 |
+
176,
|
| 630 |
+
380,
|
| 631 |
+
823,
|
| 632 |
+
422
|
| 633 |
+
],
|
| 634 |
+
"page_idx": 5
|
| 635 |
+
},
|
| 636 |
+
{
|
| 637 |
+
"type": "text",
|
| 638 |
+
"text": "VTAB-1k. VTAB-1k benchmark is introduced in [67], which contains 19 tasks from diverse domains: i) Natural images that are captured by standard cameras; ii) Specialized images that are captured by non-standard cameras, e.g., remote sensing and medical cameras; iii) Structured images that are synthesized from simulated environments. This benchmark contains a variety of tasks (e.g., object counting, depth estimation) from different image domains and each task only contains 1,000 training samples, thus is extremely challenging. ",
|
| 639 |
+
"bbox": [
|
| 640 |
+
174,
|
| 641 |
+
428,
|
| 642 |
+
825,
|
| 643 |
+
512
|
| 644 |
+
],
|
| 645 |
+
"page_idx": 5
|
| 646 |
+
},
|
| 647 |
+
{
|
| 648 |
+
"type": "text",
|
| 649 |
+
"text": "General Image Classification Datasets. We also validate the effectiveness of SSF on general image classification tasks. We choose the CIFAR-100 [31] and ImageNet-1K [7] datasets as evaluation datasets, where CIFAR-100 contains 60,000 images with 100 categories. ImageNet-1K contains 1.28M training images and 50K validation images with 1,000 categories, which are very large datasets for object recognition. ",
|
| 650 |
+
"bbox": [
|
| 651 |
+
174,
|
| 652 |
+
517,
|
| 653 |
+
825,
|
| 654 |
+
588
|
| 655 |
+
],
|
| 656 |
+
"page_idx": 5
|
| 657 |
+
},
|
| 658 |
+
{
|
| 659 |
+
"type": "text",
|
| 660 |
+
"text": "Models. For a fair comparison, we follow VPT [29] and mainly select ViT-B/16 [11] model pretrained on ImageNet-21K as the initialization for fine-tuning. In addition, we also generalize our method to backbones of different model families, including the recent Swin Transformer [38] (SwinB), ConvNeXt-B [39] and AS-MLP-B [33]. The former builds a hierarchical transformer-based architecture, and the latter two belong to CNN-based architecture and MLP-based architecture respectively. ",
|
| 661 |
+
"bbox": [
|
| 662 |
+
173,
|
| 663 |
+
593,
|
| 664 |
+
825,
|
| 665 |
+
678
|
| 666 |
+
],
|
| 667 |
+
"page_idx": 5
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"type": "text",
|
| 671 |
+
"text": "Baselines. We first compare our method with the two basic fine-tuning methods: i) full fine-tuning, where all parameters of the models are updated at fine-tuning; ii) linear probing, where only the parameters of the classification head (an MLP layer) are updated. We also compare our method with recent parameter-efficient fine-tuning methods: iii) Adapter [21], where a new adapter structure with up-projection, non-linear function, and down-projection is inserted into the transformer and only the parameters of this new module are updated; iv) Bias [66], where all the bias terms of parameters are updated; v) VPT [29], where the prompts are inserted into transformers as the input tokens and they are updated at fine-tuning. ",
|
| 672 |
+
"bbox": [
|
| 673 |
+
174,
|
| 674 |
+
683,
|
| 675 |
+
825,
|
| 676 |
+
795
|
| 677 |
+
],
|
| 678 |
+
"page_idx": 5
|
| 679 |
+
},
|
| 680 |
+
{
|
| 681 |
+
"type": "text",
|
| 682 |
+
"text": "Implementation Details. For the FGVC datasets, we process the image with a randomly resize crop to $2 2 4 \\times 2 2 4$ and a random horizontal flip for data augmentation. For VTAB-1k, we directly resize the image to $2 2 4 \\times 2 2 4$ , following the default settings in VTAB [67]. For CIFAR-100 and ImageNet1K, we follow the fine-tuning setting of ViT-B/16 in [11], where the stronger data augmentation strategies are adopted. We employ the AdamW [41] optimizer to fine-tune models for 100 epochs for CIFAR-100, and 30 epochs for ImageNet-1K. The cosine decay strategy is adopted for the learning rate schedule, and the linear warm-up is used in the first 10 epochs for CIFAR-100 and 5 epochs for ImageNet-1K. ",
|
| 683 |
+
"bbox": [
|
| 684 |
+
173,
|
| 685 |
+
800,
|
| 686 |
+
825,
|
| 687 |
+
911
|
| 688 |
+
],
|
| 689 |
+
"page_idx": 5
|
| 690 |
+
},
|
| 691 |
+
{
|
| 692 |
+
"type": "table",
|
| 693 |
+
"img_path": "images/c3b9d55e0dcfd19381e6ff7fcbb7884d18c9c981cd2a9dcac91904746e8d5625.jpg",
|
| 694 |
+
"table_caption": [],
|
| 695 |
+
"table_footnote": [
|
| 696 |
+
"Table 3: Performance comparisons on five FGVC datasets with ViT-B/16 models pre-trained on ImageNet-21K. "
|
| 697 |
+
],
|
| 698 |
+
"table_body": "<table><tr><td rowspan=1 colspan=2>DatasetMethod</td><td rowspan=1 colspan=1>CUB-200-2011</td><td rowspan=1 colspan=1>NABirds</td><td rowspan=1 colspan=1>OxfordFlowers</td><td rowspan=1 colspan=1>StanfordDogs</td><td rowspan=1 colspan=1>StanfordCars</td><td rowspan=1 colspan=1>Mean</td><td rowspan=1 colspan=1>Params.(M)</td></tr><tr><td rowspan=1 colspan=2>Full fine-tuningLinear probing</td><td rowspan=1 colspan=1>87.385.3</td><td rowspan=1 colspan=1>82.775.9</td><td rowspan=1 colspan=1>98.897.9</td><td rowspan=1 colspan=1>89.486.2</td><td rowspan=1 colspan=1>84.551.3</td><td rowspan=1 colspan=1>88.5479.32</td><td rowspan=1 colspan=1>85.980.18</td></tr><tr><td rowspan=2 colspan=2>Adapter [21]Bias [66]VPT-Shallow [29]</td><td rowspan=1 colspan=1>87.188.4</td><td rowspan=1 colspan=1>84.384.2</td><td rowspan=1 colspan=1>98.598.8</td><td rowspan=1 colspan=1>89.891.2</td><td rowspan=1 colspan=1>68.679.4</td><td rowspan=1 colspan=1>85.6788.41</td><td rowspan=1 colspan=1>0.410.28</td></tr><tr><td rowspan=1 colspan=1>VPT-Shallow [29]</td><td rowspan=2 colspan=1>86.788.5</td><td rowspan=2 colspan=1>78.884.2</td><td rowspan=2 colspan=1>98.499.0</td><td rowspan=2 colspan=1>90.790.2</td><td rowspan=2 colspan=1>68.783.6</td><td rowspan=2 colspan=1>84.6289.11</td><td rowspan=2 colspan=1>0.250.85</td></tr><tr><td rowspan=1 colspan=2>VPT-Deep [29]</td></tr><tr><td rowspan=1 colspan=2>SSF (ours)</td><td rowspan=1 colspan=1>89.5</td><td rowspan=1 colspan=1>85.7</td><td rowspan=1 colspan=1>99.6</td><td rowspan=1 colspan=1>89.6</td><td rowspan=1 colspan=1>89.2</td><td rowspan=1 colspan=1>90.72</td><td rowspan=1 colspan=1>0.39</td></tr></table>",
|
| 699 |
+
"bbox": [
|
| 700 |
+
173,
|
| 701 |
+
78,
|
| 702 |
+
823,
|
| 703 |
+
227
|
| 704 |
+
],
|
| 705 |
+
"page_idx": 6
|
| 706 |
+
},
|
| 707 |
+
{
|
| 708 |
+
"type": "table",
|
| 709 |
+
"img_path": "images/f9a73210d987b88756bb0f829d38ee98635f18daea8db94c39afb87a6330c7ea.jpg",
|
| 710 |
+
"table_caption": [],
|
| 711 |
+
"table_footnote": [
|
| 712 |
+
"Table 4: Performance comparisons on the VTAB-1k benchmark with ViT-B/16 models pre-trained on ImageNet-21K. "
|
| 713 |
+
],
|
| 714 |
+
"table_body": "<table><tr><td></td><td colspan=\"5\">Natural Specialized</td><td colspan=\"5\">Structured</td></tr><tr><td>Dataset Method</td><td>CEIATIP0 Crreeaer 0</td><td>rmeii</td><td>L63uns NHAS 3</td><td>rrarar gan Bto.AI P5ss55</td><td>Phrroiar Ceeretlrt</td><td>TTsisrlsr Cresisttla poresidsp qTTa</td><td>szhTI1111 puorssitst</td><td>SrHTIPt</td><td></td><td>) sr</td></tr><tr><td>Full fine-tuning [29] Linear probing [29] Adapter [21]</td><td>68.9 87.7 63.485.0</td><td>64.3 97.2 63.2 97.0 63.2 97.7</td><td>86.987.438.8 86.336.651.0 87.0 34.6 50.8</td><td>79.7 95.7 84.2 78.587.5 76.3 88.0 73.1</td><td>73.9 68.6 74.0 70.5</td><td>56.358.6 41.7 65.5 34.330.6 33.2 55.4 37.4 31.2 53.2</td><td>57.546.7 12.520.0 30.3</td><td>25.729.1 9.619.2 22.1</td><td>52.94</td><td>65.57 85.84 0.04</td></tr><tr><td>Bias[66] VPT-Shallow [29] VPT-Deep [29]</td><td>74.1 86.1 72.887.059.2 77.7 86.9</td><td>97.5 62.6 97.5</td><td>85.359.9 51.4 87.374.5 51.2</td><td>78.7 91.672.9 78.2 92.075.6</td><td>45.7 69.8 72.9 50.5 68.4 68.5</td><td>61.5 55.6 32.4 55.9 58.640.5 67.1</td><td>25.4 13.8 66.640.0 68.7 36.1 20.2</td><td>15.7 25.1 34.1</td><td>55.82 62.05 64.85</td><td>0.27 0.14 0.11</td></tr><tr><td>SSF (ours)</td><td>78.890.8</td><td>65.8 98.0 88.3 78.1</td><td>49.6</td><td>81.896.1 83.4</td><td>60.046.5 69.0 92.6 75.199.4 91.8 90.2 52.9|87.4 95.987.4 75.5|75.9 62.3 53.380.6 77.3 54.9 29.5 37.9|73.10</td><td>72.8 73.647.9</td><td>32.9</td><td>37.8</td><td>69.43</td><td>0.60 0.24</td></tr></table>",
|
| 715 |
+
"bbox": [
|
| 716 |
+
173,
|
| 717 |
+
262,
|
| 718 |
+
825,
|
| 719 |
+
426
|
| 720 |
+
],
|
| 721 |
+
"page_idx": 6
|
| 722 |
+
},
|
| 723 |
+
{
|
| 724 |
+
"type": "text",
|
| 725 |
+
"text": "4.2 Performance Comparisons on Image Classification ",
|
| 726 |
+
"text_level": 1,
|
| 727 |
+
"bbox": [
|
| 728 |
+
173,
|
| 729 |
+
464,
|
| 730 |
+
565,
|
| 731 |
+
479
|
| 732 |
+
],
|
| 733 |
+
"page_idx": 6
|
| 734 |
+
},
|
| 735 |
+
{
|
| 736 |
+
"type": "text",
|
| 737 |
+
"text": "We compare the performance of our SSF and other baseline methods in 26 image classification tasks and the results on FGVC and VTAB-1k are shown in Table 3 and Table 4 (also see Figure 1), respectively, and the results on CIFAR-100 and ImageNet-1K are shown in Table 5, which are evaluated in Top-1 accuracy $( \\% )$ . In these three tables, the bold font shows the best accuracy of all methods and the underline font shows the second best accuracy. ",
|
| 738 |
+
"bbox": [
|
| 739 |
+
174,
|
| 740 |
+
483,
|
| 741 |
+
823,
|
| 742 |
+
553
|
| 743 |
+
],
|
| 744 |
+
"page_idx": 6
|
| 745 |
+
},
|
| 746 |
+
{
|
| 747 |
+
"type": "text",
|
| 748 |
+
"text": "We have the following findings by observing them: i) In Table 3 and Table 4, where the last column is the average of the fine-tuned parameters for each method on the corresponding datasets, our SSF outperforms VPT [29] and other parameter-efficient fine-tuning methods, and even achieves better performance than full fine-tuning, which is mainly owing to the linear transformation applied on the features. Specifically, SSF obtains $1 . 8 1 \\%$ $( 9 0 . 7 2 \\%$ vs. $8 9 . 1 1 \\%$ and $2 . 4 6 \\%$ $9 0 . 7 2 \\%$ vs. $8 8 . 5 4 \\%$ ) accuracy improvement on five FGVC datasets, and $5 . 2 9 \\%$ $7 3 . 1 0 \\%$ vs. $6 9 . 4 3 \\%$ ) and $1 1 . 4 8 \\%$ $7 3 . 1 0 \\%$ vs. $6 5 . 5 7 \\%$ ) improvement on the VTAB-1k benchmark compared to VPT and full fine-tuning. Meanwhile, SSF also uses fewer trainable parameters compared to VPT-Deep in both datasets (0.39M vs. 0.85M, 0.24M vs. 0.60M). SSF maintains a unified learnable parameter space for different tasks with a few parameters while VPT [29] needs to design the different number of prompts for each task, which also shows the conciseness of our approach; ii) In Table 5, i.e., in CIFAR-100 and ImageNet-1K, SSF and other parameter-efficient fine-tuning methods have difficulty in achieving the similar performance to the full fine-tuning, probably because these datasets have sufficient data to prevent over-fitting of the model, especially in ImageNet-1K. In contrast, in the VTAB-1k benchmark, the amount of data is not very large (e.g., only 1,000 training images), which might cause over-fitting of the model for the full fine-tuning. Nevertheless, in CIFAR-100 and ImageNet-1K, our SSF still outperforms previous parameter-efficient fine-tuning methods (Adapter, Bias, and VPT), which shows the effectiveness of our method; iii) In Table 5, the results of our SSF with Swin Transformer, ConvNeXt, and AS-MLP models consistently outperform those of other parameter-efficient fine-tuning methods, which also verifies the effectiveness of SSF on a wide variety of models. ",
|
| 749 |
+
"bbox": [
|
| 750 |
+
173,
|
| 751 |
+
559,
|
| 752 |
+
825,
|
| 753 |
+
849
|
| 754 |
+
],
|
| 755 |
+
"page_idx": 6
|
| 756 |
+
},
|
| 757 |
+
{
|
| 758 |
+
"type": "text",
|
| 759 |
+
"text": "Computational cost. To validate the efficiency of our method, we show the computational cost of SSF in Figure 3. We employ a batch size of 16 for the training stage and inference stage, and use mixed precision training. All running results in Figure 3 are measured in a single GeForce RTX 2080Ti GPU. We can see that SSF has similar training time and training memory with VPT but with less inference time and inference memory. Here, we show the computational cost of VPT with 200/50 prompts (the same number of prompts to obtain the performance in Table 5) for VPT-Shallow and VPT-Deep, respectively. When adding the number of prompts, the time cost and memory will be larger but our SSF achieves zero-overhead inference, which is more advantageous. ",
|
| 760 |
+
"bbox": [
|
| 761 |
+
176,
|
| 762 |
+
856,
|
| 763 |
+
823,
|
| 764 |
+
911
|
| 765 |
+
],
|
| 766 |
+
"page_idx": 6
|
| 767 |
+
},
|
| 768 |
+
{
|
| 769 |
+
"type": "table",
|
| 770 |
+
"img_path": "images/44b9ce52a3d8e9d2edbeac21b2b63b3dfa82cd3a02289f53b670b798de62bbd9.jpg",
|
| 771 |
+
"table_caption": [],
|
| 772 |
+
"table_footnote": [
|
| 773 |
+
"Table 5: Performance comparisons on CIFAR-100 and ImageNet-1K with various model families, where ViT-B/16, Swin-B, and ConvNeXt-B are pre-trained on ImageNet-21K, and AS-MLP-B is pre-trained on ImageNet-1K. "
|
| 774 |
+
],
|
| 775 |
+
"table_body": "<table><tr><td>Model</td><td colspan=\"3\">ViT-B/16 [11]</td><td colspan=\"4\">Swin-B [38]</td><td colspan=\"4\">ConvNeXt-B [39]</td><td colspan=\"2\">|AS-MLP-B [33]</td></tr><tr><td>Dataset</td><td>CEIPPIII0 ) :seaed</td><td>Teege</td><td>) 'seed</td><td>CEI-AIPI1</td><td>Jr) 'sr</td><td>Teee</td><td>) 'sirN</td><td>CEIPPIII0</td><td>Jr) 'rN</td><td>Trenege ) 'seed</td><td>CEIPPIPI0</td><td></td><td>n) 'se.d</td></tr><tr><td>Method Full fine-tuning Linear probing</td><td>93.82 85.88 88.70</td><td>0.08 82.04</td><td>83.58 86.57 0.77</td><td>89.27</td><td>93.85 86.858 0.10</td><td>83.25</td><td>1.03</td><td>85.20 88.03|94.14 87.67 89.20</td><td>0.10</td><td>85.80 88.85 84.05</td><td>1.03</td><td>89.96 79.04</td><td>86.83 0.10</td></tr><tr><td>Adapter [21] Bias[66] VPT-Shallow [29]</td><td>93.34 93.39 90.38</td><td>0.31 82.72 0.18 82.74 0.23 82.08</td><td>1.00 0.87 0.92</td><td>92.49 92.19 90.02</td><td>0.33 0.24 0.13</td><td>83.82 83.92 83.29</td><td>1.26 1.16 1.05</td><td>92.86 92.80 -</td><td>0.45 0.23</td><td>84.49 84.63 -</td><td>1.37 31.16 -</td><td>88.01 87.46 ·</td><td>0.33 0.26 ·</td></tr><tr><td>VPT-Deep [29] SSF (ours)</td><td>93.17 93.99</td><td>0.54 82.45 0.28 83.10</td><td>1.23 0.97</td><td>92.62 193.06</td><td>0.70 0.37</td><td>83.44 84.40</td><td>1.63 1.29</td><td>- 193.45</td><td>1 0.36</td><td>- 84.85</td><td>- 1.28</td><td>： 88.28</td><td>- 0.37</td></tr></table>",
|
| 776 |
+
"bbox": [
|
| 777 |
+
173,
|
| 778 |
+
88,
|
| 779 |
+
825,
|
| 780 |
+
272
|
| 781 |
+
],
|
| 782 |
+
"page_idx": 7
|
| 783 |
+
},
|
| 784 |
+
{
|
| 785 |
+
"type": "image",
|
| 786 |
+
"img_path": "images/89f4621c4d080085b8619165d09e34921423f5f00a635bd4afb1c6da319be0a2.jpg",
|
| 787 |
+
"image_caption": [
|
| 788 |
+
"Figure 3: Computational cost of different tuning methods. From left to right: training time, training memory, test time, and test memory. "
|
| 789 |
+
],
|
| 790 |
+
"image_footnote": [],
|
| 791 |
+
"bbox": [
|
| 792 |
+
176,
|
| 793 |
+
315,
|
| 794 |
+
825,
|
| 795 |
+
460
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 7
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "text",
|
| 801 |
+
"text": "",
|
| 802 |
+
"bbox": [
|
| 803 |
+
174,
|
| 804 |
+
507,
|
| 805 |
+
825,
|
| 806 |
+
563
|
| 807 |
+
],
|
| 808 |
+
"page_idx": 7
|
| 809 |
+
},
|
| 810 |
+
{
|
| 811 |
+
"type": "text",
|
| 812 |
+
"text": "4.3 The Impacts of Different Designs ",
|
| 813 |
+
"text_level": 1,
|
| 814 |
+
"bbox": [
|
| 815 |
+
174,
|
| 816 |
+
574,
|
| 817 |
+
442,
|
| 818 |
+
589
|
| 819 |
+
],
|
| 820 |
+
"page_idx": 7
|
| 821 |
+
},
|
| 822 |
+
{
|
| 823 |
+
"type": "text",
|
| 824 |
+
"text": "As the core operation of SSF, we thoroughly evaluate how SSF-ADA affects results, e.g., the insertion locations, the initialization of SSF-ADA and its components. We conduct experiments to analyze the impacts of different designs for fine-tuning. All experiments are implemented with pre-trained ViT-B/16 models on CIFAR-100 and the results are shown in Table 6. ",
|
| 825 |
+
"bbox": [
|
| 826 |
+
174,
|
| 827 |
+
594,
|
| 828 |
+
825,
|
| 829 |
+
648
|
| 830 |
+
],
|
| 831 |
+
"page_idx": 7
|
| 832 |
+
},
|
| 833 |
+
{
|
| 834 |
+
"type": "text",
|
| 835 |
+
"text": "The impact of the number of layers. We directly insert SSF-ADA into different layers to evaluate the effect of inserting layers, and the results are shown in table 6a. The values in the #layers column indicate the number of layers with SSF-ADA, where #layers-0 represents linear probing. From the first and second rows, we find that the results will improve from $8 8 . 7 0 \\%$ to $9 2 . 6 9 \\%$ and grow with a small number of trainable parameters (0.08M vs. 0.11M) when only inserting SSF-ADA into the first two layers. Keep adding SSF-ADA in the subsequent layers will make the results better. The growth of the results is almost linear with the number of layers of inserted SSF-ADA. Therefore, we directly choose to insert SSF-ADA into all (12) layers of vision transformer to bring the best results $( 9 3 . 9 9 \\% )$ with 0.28M trainable parameters. ",
|
| 836 |
+
"bbox": [
|
| 837 |
+
173,
|
| 838 |
+
655,
|
| 839 |
+
825,
|
| 840 |
+
780
|
| 841 |
+
],
|
| 842 |
+
"page_idx": 7
|
| 843 |
+
},
|
| 844 |
+
{
|
| 845 |
+
"type": "text",
|
| 846 |
+
"text": "The impact of the different insertion locations. Based on the different operations of ViT, we evaluate the impact of the insertion locations of SSF-ADA. We separately remove SSF-ADA after these operations and the results are shown in Table 6b. We find that removing the SSF-ADA in the MLP operation achieves inferior results than removing those in the Attention operation $( 9 3 . 4 6 \\%$ vs. $9 3 . 6 9 \\%$ with comparable trainable parameters (0.19M vs. 0.21M), which suggests that performing feature modulation for the MLP operation might be more important. Although one can use NAS to search for the importance of different operations and thereby insert SSF-ADA in specific locations, the results might not be better than inserting SSF-ADA in all operations. Therefore, in order to obtain excellent performance, we do not perform NAS but directly insert SSF-ADA into all operations. ",
|
| 847 |
+
"bbox": [
|
| 848 |
+
174,
|
| 849 |
+
786,
|
| 850 |
+
826,
|
| 851 |
+
911
|
| 852 |
+
],
|
| 853 |
+
"page_idx": 7
|
| 854 |
+
},
|
| 855 |
+
{
|
| 856 |
+
"type": "table",
|
| 857 |
+
"img_path": "images/deb0ba75455cc31fb498355e3a4c57f0a9e3eaef36db1f69003e18d55b41d124.jpg",
|
| 858 |
+
"table_caption": [],
|
| 859 |
+
"table_footnote": [
|
| 860 |
+
"Table 6: The impacts of different designs. (a) The impact of the number of layers with SSF-ADA. (b) The impacts of the different insertion locations of SSF-ADA. (c) The impacts of initialization. (d) The impacts of different components. Acc.: Top-1 accuracy $( \\% )$ ; Params.: parameters (M). "
|
| 861 |
+
],
|
| 862 |
+
"table_body": "<table><tr><td>#layers</td><td>Acc.</td><td>Params.</td><td>location</td><td>Acc. Params.</td><td>initialization</td><td>Acc.</td><td></td><td>case</td><td>Acc. Params.</td></tr><tr><td>0</td><td>88.70</td><td>0.08</td><td>w/o. mlp</td><td>93.46 0.19</td><td>random</td><td>[90.11</td><td>w/o. scale</td><td>93.49</td><td>0.18</td></tr><tr><td>2</td><td>92.69</td><td>0.11</td><td>w/o.attn</td><td>93.69 0.21</td><td>constant</td><td>93.91</td><td>w/o. shift</td><td>93.74</td><td>0.18</td></tr><tr><td>4</td><td>93.30</td><td>0.15</td><td>w/o.embed</td><td>93.91 0.28</td><td>uniform</td><td>93.87</td><td>only norm</td><td>93.26</td><td>0.11</td></tr><tr><td>8</td><td>93.60</td><td>0.22</td><td>w/o.norm</td><td>93.80 0.25</td><td>trunc_normal</td><td>93.93</td><td>scalar scale</td><td>93.59</td><td>0.18</td></tr><tr><td>12 (ours)</td><td>93.99</td><td>0.28</td><td>ours</td><td>93.99 0.28</td><td>normal (ours)</td><td>93.99</td><td>ours</td><td>93.99</td><td>0.28</td></tr><tr><td colspan=\"2\">(a)</td><td></td><td>(b)</td><td></td><td>(c)</td><td></td><td></td><td>(d)</td><td></td></tr></table>",
|
| 863 |
+
"bbox": [
|
| 864 |
+
183,
|
| 865 |
+
101,
|
| 866 |
+
821,
|
| 867 |
+
204
|
| 868 |
+
],
|
| 869 |
+
"page_idx": 8
|
| 870 |
+
},
|
| 871 |
+
{
|
| 872 |
+
"type": "text",
|
| 873 |
+
"text": "The impact of initialization. We also investigate how different ways of initializing the scale and shift factors affect performance in Table 6c. In our experiments, we first randomly initialize both scale and shift parameters with a mean value of zero, but find that the performance is inferior $( 9 0 . 1 1 \\% )$ and cannot converge in some experiments. After that, we randomly initialize the scale factor with a mean value of one and find better performance, which implies that the weights of a pre-trained model should not be completely disrupted in the fine-tuning, instead, we should start from this pre-trained model to optimize our model. Experiments show that using the normal initialization achieves the best performance, where the mean values of the scale factor and shift factor are one and zero, respectively. ",
|
| 874 |
+
"bbox": [
|
| 875 |
+
173,
|
| 876 |
+
248,
|
| 877 |
+
825,
|
| 878 |
+
359
|
| 879 |
+
],
|
| 880 |
+
"page_idx": 8
|
| 881 |
+
},
|
| 882 |
+
{
|
| 883 |
+
"type": "text",
|
| 884 |
+
"text": "The impact of different components. We also evaluate the impacts of different components in SSF-ADA and the results are shown in Table 6d. We find that removing the scale term yields worse performance than removing the shift term with the same trainable parameters, which shows that the scale term might be more important than the shift term. Also, note that the difference between ‘w/o. scale’ and the ‘Bias’ method in Table 5 is that we fine-tune the model with an additional shift term in ‘w/o. scale’, while ‘Bias’ fine-tunes the model based on the original biases, suggesting that fine-tuning the model in a res-like manner can obtain slightly better performance $( 9 3 . 4 9 \\%$ vs. $9 3 . 3 9 \\%$ ). We also try to only fine-tune all scale and shift factors in the normalization layer (LN), or fine-tune the model with SSF but set the scale term as a scalar. These experiments yield inferior performance than SSF $9 3 . 2 6 \\%$ vs. $9 3 . 9 9 \\%$ , $9 3 . 5 9 \\%$ vs. $9 3 . 9 9 \\%$ ), but could probably be considered as an alternative due to the fact that they only use about half of the trainable parameters of SSF. ",
|
| 885 |
+
"bbox": [
|
| 886 |
+
173,
|
| 887 |
+
366,
|
| 888 |
+
825,
|
| 889 |
+
517
|
| 890 |
+
],
|
| 891 |
+
"page_idx": 8
|
| 892 |
+
},
|
| 893 |
+
{
|
| 894 |
+
"type": "text",
|
| 895 |
+
"text": "4.4 Performance Comparisons on Robustness and OOD Datasets ",
|
| 896 |
+
"text_level": 1,
|
| 897 |
+
"bbox": [
|
| 898 |
+
174,
|
| 899 |
+
526,
|
| 900 |
+
637,
|
| 901 |
+
541
|
| 902 |
+
],
|
| 903 |
+
"page_idx": 8
|
| 904 |
+
},
|
| 905 |
+
{
|
| 906 |
+
"type": "text",
|
| 907 |
+
"text": "We also conduct experiments to analyze the robustness and Out-Of-Distribution (OOD) ability of our SSF method with the following datasets: ImageNet-A, ImageNet-R and ImageNet-C. Please refer to Appendix for their details. We perform the robustness and OOD evaluation on these three datasets with the fine-tuned models on ImageNet-1K. All experimental results are listed in Table 7. ",
|
| 908 |
+
"bbox": [
|
| 909 |
+
173,
|
| 910 |
+
545,
|
| 911 |
+
447,
|
| 912 |
+
683
|
| 913 |
+
],
|
| 914 |
+
"page_idx": 8
|
| 915 |
+
},
|
| 916 |
+
{
|
| 917 |
+
"type": "text",
|
| 918 |
+
"text": "From this table, we can see that our SSF obtains better performance than VPT and other parameter-efficient fine-tuning methods on three datasets, which shows our fine-tuning method has stronger robust",
|
| 919 |
+
"bbox": [
|
| 920 |
+
174,
|
| 921 |
+
690,
|
| 922 |
+
447,
|
| 923 |
+
760
|
| 924 |
+
],
|
| 925 |
+
"page_idx": 8
|
| 926 |
+
},
|
| 927 |
+
{
|
| 928 |
+
"type": "table",
|
| 929 |
+
"img_path": "images/8aeb4da590cb67018f9073a41655584abd5d8d04f746eeb79e8d72c6cafe8a41.jpg",
|
| 930 |
+
"table_caption": [
|
| 931 |
+
"Table 7: Performance comparisons on robustness and out-of-distribution datasets. ‘IN’ means ImageNet. The performance on IN-1K, IN-A and IN-R is evaluated in Top-1 accuracy $( \\% )$ . The performance on IN-C is evaluated in mCE (mean corruption error). The lower $( \\downarrow )$ , the better. "
|
| 932 |
+
],
|
| 933 |
+
"table_footnote": [],
|
| 934 |
+
"table_body": "<table><tr><td></td><td>Dataset</td><td rowspan=\"2\">IN-1K (↑)</td><td rowspan=\"2\">IN-A (↑)</td><td rowspan=\"2\">IN-R (↑)</td><td rowspan=\"2\">IN-C (↓)</td></tr><tr><td>Method</td><td></td></tr><tr><td colspan=\"2\">Full fine-tuning</td><td>83.58</td><td>34.49</td><td>51.29</td><td>46.47</td></tr><tr><td colspan=\"2\">Linear probing</td><td>82.04</td><td>33.91</td><td>52.87</td><td>46.91</td></tr><tr><td colspan=\"2\">Adapter [21]</td><td>82.72</td><td>42.21</td><td>54.13</td><td>42.65</td></tr><tr><td colspan=\"2\">Bias [66]</td><td>82.74</td><td>42.12</td><td>55.94</td><td>41.90</td></tr><tr><td colspan=\"2\">VPT-Shallow [29]</td><td>82.08</td><td>30.93</td><td>53.72</td><td>46.88</td></tr><tr><td colspan=\"2\">VPT-Deep [29]</td><td>82.45</td><td>39.10</td><td>53.54</td><td>43.10</td></tr><tr><td colspan=\"2\">SSF (ours)</td><td>83.10</td><td>45.88</td><td>56.77</td><td>41.47</td></tr></table>",
|
| 935 |
+
"bbox": [
|
| 936 |
+
464,
|
| 937 |
+
551,
|
| 938 |
+
816,
|
| 939 |
+
662
|
| 940 |
+
],
|
| 941 |
+
"page_idx": 8
|
| 942 |
+
},
|
| 943 |
+
{
|
| 944 |
+
"type": "text",
|
| 945 |
+
"text": "ness and out-of-distribution generalization. Furthermore, although SSF has lower accuracy than full fine-tuning on ImageNet-1K, the performance on ImageNet-A, ImageNet-R and ImageNet-C is better, which also shows the performance between ImageNet-1K and ImageNet-A/R/C is not absolutely positive relevant. Such improvements in robustness and OOD datasets might come from the fact that SSF freezes most of the pre-trained parameters, which maximally preserves the knowledge learned from the large-scale dataset and thus maintains a better generalization ability. ",
|
| 946 |
+
"bbox": [
|
| 947 |
+
174,
|
| 948 |
+
758,
|
| 949 |
+
825,
|
| 950 |
+
842
|
| 951 |
+
],
|
| 952 |
+
"page_idx": 8
|
| 953 |
+
},
|
| 954 |
+
{
|
| 955 |
+
"type": "text",
|
| 956 |
+
"text": "4.5 Visualization and Analysis ",
|
| 957 |
+
"text_level": 1,
|
| 958 |
+
"bbox": [
|
| 959 |
+
174,
|
| 960 |
+
852,
|
| 961 |
+
397,
|
| 962 |
+
866
|
| 963 |
+
],
|
| 964 |
+
"page_idx": 8
|
| 965 |
+
},
|
| 966 |
+
{
|
| 967 |
+
"type": "text",
|
| 968 |
+
"text": "Although our goal is to modulate the features extracted by a pre-trained model, the scale and shift parameters are input-independent indeed. Therefore, these parameters can also be regarded as encoding information of the whole downstream dataset. After re-parameterization, these scale and shift parameters are absorbed into the original model weights. To better understand information learned by the SSF, we visualize the distributions of weights and biases before and after finetuning via SSF in Figure 4a. We can see that the scale and shift parameters adjust the original weights and biases, and change the distribution of weights and biases to fit the downstream task. ",
|
| 969 |
+
"bbox": [
|
| 970 |
+
174,
|
| 971 |
+
869,
|
| 972 |
+
823,
|
| 973 |
+
911
|
| 974 |
+
],
|
| 975 |
+
"page_idx": 8
|
| 976 |
+
},
|
| 977 |
+
{
|
| 978 |
+
"type": "image",
|
| 979 |
+
"img_path": "images/1732345a26a5fa74985966322a92d0ce4ad77eec09595f26ab1354fb42428702.jpg",
|
| 980 |
+
"image_caption": [
|
| 981 |
+
"Figure 4: Comparisons of parameter distribution between the original pre-trained model and different fine-tuning methods. The first row shows weight distribution and the second row is bias distribution. The blue histograms show the original pre-trained model, and the orange ones show the fine-tuned model via SSF in (a) and full fine-tuned model in (b). "
|
| 982 |
+
],
|
| 983 |
+
"image_footnote": [],
|
| 984 |
+
"bbox": [
|
| 985 |
+
179,
|
| 986 |
+
102,
|
| 987 |
+
818,
|
| 988 |
+
236
|
| 989 |
+
],
|
| 990 |
+
"page_idx": 9
|
| 991 |
+
},
|
| 992 |
+
{
|
| 993 |
+
"type": "text",
|
| 994 |
+
"text": "",
|
| 995 |
+
"bbox": [
|
| 996 |
+
174,
|
| 997 |
+
308,
|
| 998 |
+
825,
|
| 999 |
+
363
|
| 1000 |
+
],
|
| 1001 |
+
"page_idx": 9
|
| 1002 |
+
},
|
| 1003 |
+
{
|
| 1004 |
+
"type": "text",
|
| 1005 |
+
"text": "As a comparison, we also visualize the original weight distribution and the weight distribution after full fine-tuning in Figure 4b, from which we can find an interesting phenomenon that full fine-tuning does not change the distribution of weights and biases much, but probably only a small portion of the values is changed. It is worth noting that although SSF does not match the weight distribution of full fine-tuning, it achieves better performance $9 3 . 9 9 \\%$ vs. $9 3 . 8 2 \\%$ in Table 5) on CIFAR-100. ",
|
| 1006 |
+
"bbox": [
|
| 1007 |
+
174,
|
| 1008 |
+
364,
|
| 1009 |
+
583,
|
| 1010 |
+
474
|
| 1011 |
+
],
|
| 1012 |
+
"page_idx": 9
|
| 1013 |
+
},
|
| 1014 |
+
{
|
| 1015 |
+
"type": "text",
|
| 1016 |
+
"text": "To further investigate why SSF can achieve superior performance, beyond weight distribution, we also visualize the feature similarities between full fine-tuning and linear probing, full fine-tuning and VPT-Deep, full fine-tuning and SSF, as shown in Figure 5. In the last layer, SSF has the most similar feature to full fine-tuning and the accuracy is also the closest. This shows that even if the weight distribution learned by SSF is different from full fine-tuning, SSF is also able to extract the features of the images in the downstream task very well, which validates the effectiveness of our method. ",
|
| 1017 |
+
"bbox": [
|
| 1018 |
+
174,
|
| 1019 |
+
481,
|
| 1020 |
+
581,
|
| 1021 |
+
618
|
| 1022 |
+
],
|
| 1023 |
+
"page_idx": 9
|
| 1024 |
+
},
|
| 1025 |
+
{
|
| 1026 |
+
"type": "image",
|
| 1027 |
+
"img_path": "images/d51c3872ae3cde7269450362c50ffb3e382bf0e22bd6e5dc549ce6434b9cc2ea.jpg",
|
| 1028 |
+
"image_caption": [
|
| 1029 |
+
"Figure 5: The visualization of the feature similarities between full fine-tuning and linear probing, full fine-tuning and VPT-Deep, full finetuning and SSF, in different layers of ViT-B/16. "
|
| 1030 |
+
],
|
| 1031 |
+
"image_footnote": [],
|
| 1032 |
+
"bbox": [
|
| 1033 |
+
594,
|
| 1034 |
+
375,
|
| 1035 |
+
821,
|
| 1036 |
+
508
|
| 1037 |
+
],
|
| 1038 |
+
"page_idx": 9
|
| 1039 |
+
},
|
| 1040 |
+
{
|
| 1041 |
+
"type": "text",
|
| 1042 |
+
"text": "5 Conclusion ",
|
| 1043 |
+
"text_level": 1,
|
| 1044 |
+
"bbox": [
|
| 1045 |
+
174,
|
| 1046 |
+
631,
|
| 1047 |
+
299,
|
| 1048 |
+
647
|
| 1049 |
+
],
|
| 1050 |
+
"page_idx": 9
|
| 1051 |
+
},
|
| 1052 |
+
{
|
| 1053 |
+
"type": "text",
|
| 1054 |
+
"text": "In this paper, we focus on parameter-efficient fine-tuning and propose an SSF method to scale and shift the features extracted by a pre-trained model. The intuition behind our method comes from alleviating the distribution mismatch between upstream tasks and downstream tasks by modulating deep features. SSF surprisingly outperforms other parameter-efficient fine-tuning approaches with a small number of learnable parameters. Besides, the introduced scale and shift parameters during the fine-tuning can be merged into the original pre-trained model weights via re-parameterization in the inference phase, thereby avoiding extra parameters and FLOPs. With the proposed SSF method, our model obtains $2 . 4 6 \\%$ $9 0 . 7 2 \\%$ vs. $8 8 . 5 4 \\%$ ) and $1 1 . 4 8 \\%$ $7 3 . 1 0 \\%$ vs. $6 5 . 5 7 \\%$ ) performance improvement on FGVC and VTAB-1k in terms of Top-1 accuracy compared to the full fine-tuning but only fine-tuning about $0 . 3 { \\bf M }$ parameters. Experiments on 26 image classification datasets in total and 3 robustness & out-of-distribution datasets with various model families (CNNs, Transformers, and MLPs) show the effectiveness of SSF, which establishes a new baseline. ",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
173,
|
| 1057 |
+
655,
|
| 1058 |
+
825,
|
| 1059 |
+
820
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 9
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "text",
|
| 1065 |
+
"text": "Acknowledgement ",
|
| 1066 |
+
"text_level": 1,
|
| 1067 |
+
"bbox": [
|
| 1068 |
+
176,
|
| 1069 |
+
830,
|
| 1070 |
+
331,
|
| 1071 |
+
847
|
| 1072 |
+
],
|
| 1073 |
+
"page_idx": 9
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"type": "text",
|
| 1077 |
+
"text": "The authors acknowledge the support from the Singapore National Research Foundation (“CogniVision – Energy-autonomous always-on cognitive and attentive cameras for distributed real-time vision with milliwatt power consumption” grant NRF-CRP20-2017-0003) – www.green-ic.org/ CogniVision. Xinchao Wang is the corresponding author. ",
|
| 1078 |
+
"bbox": [
|
| 1079 |
+
174,
|
| 1080 |
+
851,
|
| 1081 |
+
825,
|
| 1082 |
+
906
|
| 1083 |
+
],
|
| 1084 |
+
"page_idx": 9
|
| 1085 |
+
},
|
| 1086 |
+
{
|
| 1087 |
+
"type": "text",
|
| 1088 |
+
"text": "References ",
|
| 1089 |
+
"text_level": 1,
|
| 1090 |
+
"bbox": [
|
| 1091 |
+
174,
|
| 1092 |
+
90,
|
| 1093 |
+
266,
|
| 1094 |
+
106
|
| 1095 |
+
],
|
| 1096 |
+
"page_idx": 10
|
| 1097 |
+
},
|
| 1098 |
+
{
|
| 1099 |
+
"type": "text",
|
| 1100 |
+
"text": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. \n[2] Han Cai, Chuang Gan, Ligeng Zhu, and Song Han. Tinytl: Reduce memory, not parameters for efficient on-device learning. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 11285–11297. Curran Associates, Inc., 2020. \n[3] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. arXiv preprint arXiv:2205.13535, 2022. \n[4] Ting Chen, Mario Lucic, Neil Houlsby, and Sylvain Gelly. On self modulation for generative adversarial networks. In International Conference on Learning Representations (ICLR), 2019. \n[5] Zihang Dai, Hanxiao Liu, Quoc V Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes. Advances in Neural Information Processing Systems, 34:3965–3977, 2021. \n[6] Harm De Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C Courville. Modulating early visual processing by language. Advances in Neural Information Processing Systems, 30, 2017. \n[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. \n[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. \n[9] Xiaohan Ding, Chunlong Xia, Xiangyu Zhang, Xiaojie Chu, Jungong Han, and Guiguang Ding. Repmlp: Re-parameterizing convolutions into fully-connected layers for image recognition. arXiv preprint arXiv:2105.01883, 2021. \n[10] Xiaohan Ding, Xiangyu Zhang, Ningning Ma, Jungong Han, Guiguang Ding, and Jian Sun. Repvgg: Making vgg-style convnets great again. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733–13742, 2021. \n[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. \n[12] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Finegrained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31, 2017. \n[13] Yunhui Guo, Honghui Shi, Abhishek Kumar, Kristen Grauman, Tajana Rosing, and Rogerio Feris. Spottune: transfer learning through adaptive fine-tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4805–4814, 2019. \n[14] Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, and Jian Sun. Single path one-shot neural architecture search with uniform sampling. In European Conference on Computer Vision, pages 544–560. Springer, 2020. \n[15] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000–16009, 2022. \n[16] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020. \n[17] Kaiming He, Ross Girshick, and Piotr Dollár. Rethinking imagenet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4918–4927, 2019. \n[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. \n[19] Xuehai He, Chunyuan Li, Pengchuan Zhang, Jianwei Yang, and Xin Eric Wang. Parameterefficient fine-tuning for vision transformers. arXiv preprint arXiv:2203.16329, 2022. \n[20] Qibin Hou, Zihang Jiang, Li Yuan, Ming-Ming Cheng, Shuicheng Yan, and Jiashi Feng. Vision permutator: A permutable mlp-like architecture for visual recognition. arXiv preprint arXiv:2106.12368, 2021. \n[21] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790–2799. PMLR, 2019. \n[22] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, To Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. \n[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. \n[24] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. \n[25] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501–1510, 2017. \n[26] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448–456. PMLR, 2015. \n[27] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2704–2713, 2018. \n[28] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. \n[29] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision (ECCV), 2022. \n[30] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR Workshop on Fine-Grained Visual Categorization (FGVC), volume 2. Citeseer, 2011. \n[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. \n[32] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. \n[33] Dongze Lian, Zehao Yu, Xing Sun, and Shenghua Gao. As-mlp: An axial shifted mlp architecture for vision. In International Conference on Learning Representations (ICLR), 2022. \n[34] Dongze Lian, Yin Zheng, Yintao Xu, Yanxiong Lu, Leyu Lin, Peilin Zhao, Junzhou Huang, and Shenghua Gao. Towards fast adaptation of neural architectures with meta learning. In International Conference on Learning Representations (ICLR), 2020. \n[35] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. In International Conference on Learning Representations (ICLR), 2019. \n[36] Huihui Liu, Yiding Yang, and Xinchao Wang. Overcoming catastrophic forgetting in graph neural networks. In AAAI Conference on Artificial Intelligence, 2021. \n[37] Songhua Liu, Jingwen Ye, Sucheng Ren, and Xinchao Wang. Dynast: Dynamic sparse transformer for exemplar-guided image generation. In Proceedings of the European Conference on Computer Vision, 2022. \n[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. \n[39] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. \n[40] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. \n[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019. \n[42] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. arXiv preprint arXiv:2106.04647, 2021. \n[43] Rabeeh Karimi Mahabadi, Sebastian Ruder, Mostafa Dehghani, and James Henderson. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. arXiv preprint arXiv:2106.04489, 2021. number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722–729. IEEE, 2008. \n[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. \n[46] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In International conference on machine learning, pages 4095– 4104. PMLR, 2018. \n[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748–8763. PMLR, 2021. \n[48] Sucheng Ren, Daquan Zhou, Shengfeng He, Jiashi Feng, and Xinchao Wang. Shunted selfattention via multi-scale token aggregation. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. \n[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2015. \n[50] Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 2016. \n[51] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning (ICML), pages 6105–6114, 2019. \n[52] Ilya Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, et al. MLP-Mixer: An all-mlp architecture for vision. arXiv preprint arXiv:2105.01601, 2021. \n[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. \n[54] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about vision transformers. arXiv preprint arXiv:2203.09795, 2022. \n[55] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 595–604, 2015. \n[56] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. \n[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. \n[58] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. \n[59] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. \n[60] Xingyi Yang, Jingwen Ye, and Xinchao Wang. Factorizing knowledge in neural networks. In Proceedings of the European Conference on Computer Vision, 2022. \n[61] Yiding Yang, Zunlei Feng, Mingli Song, and Xinchao Wang. Factorizable graph convolutional networks. In Conference on Neural Information Processing Systems, 2020. \n[62] Yiding Yang, Jiayan Qiu, Mingli Song, Dacheng Tao, and Xinchao Wang. Distilling knowledge from graph convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition, 2020. \n[63] Jingwen Ye, Yifang Fu, Jie Song, Xingyi Yang, Songhua Liu, Xin Jin, Mingli Song, and Xinchao Wang. Learning with recoverable forgetting. In Proceedings of the European Conference on Computer Vision, 2022. \n[64] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In IEEE Conference on Computer Vision and Pattern Recognition, 2022. \n[65] Sergey Zagoruyko and Nikos Komodakis. Diracnets: Training very deep neural networks without skip-connections. arXiv preprint arXiv:1706.00388, 2017. \n[66] Elad Ben Zaken, Shauli Ravfogel, and Yoav Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. arXiv preprint arXiv:2106.10199, 2021. \n[67] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. \n[68] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Neural prompt search. arXiv preprint arXiv:2206.04673, 2022. \n[69] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. DeepViT: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. \n[70] Daquan Zhou, Zhiding Yu, Enze Xie, Chaowei Xiao, Anima Anandkumar, Jiashi Feng, and Jose M Alvarez. Understanding the robustness in vision transformers. arXiv preprint arXiv:2204.12451, 2022. \n[71] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134, 2021. ",
|
| 1101 |
+
"bbox": [
|
| 1102 |
+
173,
|
| 1103 |
+
140,
|
| 1104 |
+
828,
|
| 1105 |
+
905
|
| 1106 |
+
],
|
| 1107 |
+
"page_idx": 10
|
| 1108 |
+
},
|
| 1109 |
+
{
|
| 1110 |
+
"type": "text",
|
| 1111 |
+
"text": "",
|
| 1112 |
+
"bbox": [
|
| 1113 |
+
171,
|
| 1114 |
+
90,
|
| 1115 |
+
828,
|
| 1116 |
+
898
|
| 1117 |
+
],
|
| 1118 |
+
"page_idx": 11
|
| 1119 |
+
},
|
| 1120 |
+
{
|
| 1121 |
+
"type": "text",
|
| 1122 |
+
"text": "",
|
| 1123 |
+
"bbox": [
|
| 1124 |
+
171,
|
| 1125 |
+
109,
|
| 1126 |
+
828,
|
| 1127 |
+
897
|
| 1128 |
+
],
|
| 1129 |
+
"page_idx": 12
|
| 1130 |
+
},
|
| 1131 |
+
{
|
| 1132 |
+
"type": "text",
|
| 1133 |
+
"text": "",
|
| 1134 |
+
"bbox": [
|
| 1135 |
+
173,
|
| 1136 |
+
90,
|
| 1137 |
+
826,
|
| 1138 |
+
320
|
| 1139 |
+
],
|
| 1140 |
+
"page_idx": 13
|
| 1141 |
+
},
|
| 1142 |
+
{
|
| 1143 |
+
"type": "text",
|
| 1144 |
+
"text": "Checklist ",
|
| 1145 |
+
"text_level": 1,
|
| 1146 |
+
"bbox": [
|
| 1147 |
+
174,
|
| 1148 |
+
89,
|
| 1149 |
+
254,
|
| 1150 |
+
106
|
| 1151 |
+
],
|
| 1152 |
+
"page_idx": 14
|
| 1153 |
+
},
|
| 1154 |
+
{
|
| 1155 |
+
"type": "text",
|
| 1156 |
+
"text": "1. For all authors... ",
|
| 1157 |
+
"bbox": [
|
| 1158 |
+
214,
|
| 1159 |
+
116,
|
| 1160 |
+
339,
|
| 1161 |
+
130
|
| 1162 |
+
],
|
| 1163 |
+
"page_idx": 14
|
| 1164 |
+
},
|
| 1165 |
+
{
|
| 1166 |
+
"type": "text",
|
| 1167 |
+
"text": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? [Yes] See abstract, introduction and experiments. \n(b) Did you describe the limitations of your work? [Yes] See appendix. \n(c) Did you discuss any potential negative societal impacts of your work? [Yes] See appendix. \n(d) Have you read the ethics review guidelines and ensured that your paper conforms to them? [Yes] ",
|
| 1168 |
+
"bbox": [
|
| 1169 |
+
238,
|
| 1170 |
+
135,
|
| 1171 |
+
825,
|
| 1172 |
+
239
|
| 1173 |
+
],
|
| 1174 |
+
"page_idx": 14
|
| 1175 |
+
},
|
| 1176 |
+
{
|
| 1177 |
+
"type": "text",
|
| 1178 |
+
"text": "2. If you are including theoretical results... ",
|
| 1179 |
+
"bbox": [
|
| 1180 |
+
214,
|
| 1181 |
+
243,
|
| 1182 |
+
493,
|
| 1183 |
+
258
|
| 1184 |
+
],
|
| 1185 |
+
"page_idx": 14
|
| 1186 |
+
},
|
| 1187 |
+
{
|
| 1188 |
+
"type": "text",
|
| 1189 |
+
"text": "(a) Did you state the full set of assumptions of all theoretical results? [N/A] (b) Did you include complete proofs of all theoretical results? [N/A] ",
|
| 1190 |
+
"bbox": [
|
| 1191 |
+
238,
|
| 1192 |
+
262,
|
| 1193 |
+
738,
|
| 1194 |
+
294
|
| 1195 |
+
],
|
| 1196 |
+
"page_idx": 14
|
| 1197 |
+
},
|
| 1198 |
+
{
|
| 1199 |
+
"type": "text",
|
| 1200 |
+
"text": "3. If you ran experiments... ",
|
| 1201 |
+
"bbox": [
|
| 1202 |
+
212,
|
| 1203 |
+
299,
|
| 1204 |
+
393,
|
| 1205 |
+
313
|
| 1206 |
+
],
|
| 1207 |
+
"page_idx": 14
|
| 1208 |
+
},
|
| 1209 |
+
{
|
| 1210 |
+
"type": "text",
|
| 1211 |
+
"text": "(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See abstract and experiments. \n(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)? [Yes] See experiments. \n(c) Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)? [N/A] A part of the experiments is tested several times. \n(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See experiments. ",
|
| 1212 |
+
"bbox": [
|
| 1213 |
+
238,
|
| 1214 |
+
316,
|
| 1215 |
+
825,
|
| 1216 |
+
449
|
| 1217 |
+
],
|
| 1218 |
+
"page_idx": 14
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"type": "text",
|
| 1222 |
+
"text": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets... ",
|
| 1223 |
+
"bbox": [
|
| 1224 |
+
218,
|
| 1225 |
+
453,
|
| 1226 |
+
823,
|
| 1227 |
+
468
|
| 1228 |
+
],
|
| 1229 |
+
"page_idx": 14
|
| 1230 |
+
},
|
| 1231 |
+
{
|
| 1232 |
+
"type": "text",
|
| 1233 |
+
"text": "(a) If your work uses existing assets, did you cite the creators? [Yes] \n(b) Did you mention the license of the assets? [Yes] \n(c) Did you include any new assets either in the supplemental material or as a URL? [Yes] \n(d) Did you discuss whether and how consent was obtained from people whose data you’re using/curating? [N/A] \n(e) Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content? [Yes] See appendix. ",
|
| 1234 |
+
"bbox": [
|
| 1235 |
+
238,
|
| 1236 |
+
472,
|
| 1237 |
+
823,
|
| 1238 |
+
579
|
| 1239 |
+
],
|
| 1240 |
+
"page_idx": 14
|
| 1241 |
+
},
|
| 1242 |
+
{
|
| 1243 |
+
"type": "text",
|
| 1244 |
+
"text": "5. If you used crowdsourcing or conducted research with human subjects... ",
|
| 1245 |
+
"bbox": [
|
| 1246 |
+
214,
|
| 1247 |
+
583,
|
| 1248 |
+
705,
|
| 1249 |
+
598
|
| 1250 |
+
],
|
| 1251 |
+
"page_idx": 14
|
| 1252 |
+
},
|
| 1253 |
+
{
|
| 1254 |
+
"type": "text",
|
| 1255 |
+
"text": "(a) Did you include the full text of instructions given to participants and screenshots, if applicable? [N/A] \n(b) Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable? [N/A] \n(c) Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation? [N/A] ",
|
| 1256 |
+
"bbox": [
|
| 1257 |
+
238,
|
| 1258 |
+
602,
|
| 1259 |
+
825,
|
| 1260 |
+
691
|
| 1261 |
+
],
|
| 1262 |
+
"page_idx": 14
|
| 1263 |
+
}
|
| 1264 |
+
]
|
parse/dev/XtyeppctGgc/XtyeppctGgc_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|