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 +380 -0
- parse/train/1YLJDvSx6J4/1YLJDvSx6J4_layout.pdf +3 -0
- parse/train/1YLJDvSx6J4/1YLJDvSx6J4_origin.pdf +3 -0
- parse/train/1YLJDvSx6J4/1YLJDvSx6J4_span.pdf +3 -0
- parse/train/27acGyyI1BY/27acGyyI1BY_layout.pdf +3 -0
- parse/train/27acGyyI1BY/27acGyyI1BY_origin.pdf +3 -0
- parse/train/27acGyyI1BY/27acGyyI1BY_span.pdf +3 -0
- parse/train/3qMwV98zLIk/3qMwV98zLIk_layout.pdf +3 -0
- parse/train/3qMwV98zLIk/3qMwV98zLIk_origin.pdf +3 -0
- parse/train/3qMwV98zLIk/3qMwV98zLIk_span.pdf +3 -0
- parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_layout.pdf +3 -0
- parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_origin.pdf +3 -0
- parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_span.pdf +3 -0
- parse/train/6nbpPqUCIi7/6nbpPqUCIi7_layout.pdf +3 -0
- parse/train/6nbpPqUCIi7/6nbpPqUCIi7_origin.pdf +3 -0
- parse/train/6nbpPqUCIi7/6nbpPqUCIi7_span.pdf +3 -0
- parse/train/86NHK__yFDl/86NHK__yFDl_layout.pdf +3 -0
- parse/train/86NHK__yFDl/86NHK__yFDl_origin.pdf +3 -0
- parse/train/86NHK__yFDl/86NHK__yFDl_span.pdf +3 -0
- parse/train/8PA2nX9v_r2/8PA2nX9v_r2.md +269 -0
- parse/train/8PA2nX9v_r2/8PA2nX9v_r2_content_list.json +1196 -0
- parse/train/8PA2nX9v_r2/8PA2nX9v_r2_middle.json +0 -0
- parse/train/8PA2nX9v_r2/8PA2nX9v_r2_model.json +0 -0
- parse/train/9p2ekP904Rs/9p2ekP904Rs_layout.pdf +3 -0
- parse/train/9p2ekP904Rs/9p2ekP904Rs_origin.pdf +3 -0
- parse/train/9p2ekP904Rs/9p2ekP904Rs_span.pdf +3 -0
- parse/train/B18WgG-CZ/B18WgG-CZ_layout.pdf +3 -0
- parse/train/B18WgG-CZ/B18WgG-CZ_origin.pdf +3 -0
- parse/train/B18WgG-CZ/B18WgG-CZ_span.pdf +3 -0
- parse/train/B1Gi6LeRZ/B1Gi6LeRZ_layout.pdf +3 -0
- parse/train/B1Gi6LeRZ/B1Gi6LeRZ_origin.pdf +3 -0
- parse/train/B1Gi6LeRZ/B1Gi6LeRZ_span.pdf +3 -0
- parse/train/B1eY_pVYvB/B1eY_pVYvB_layout.pdf +3 -0
- parse/train/B1eY_pVYvB/B1eY_pVYvB_origin.pdf +3 -0
- parse/train/B1eY_pVYvB/B1eY_pVYvB_span.pdf +3 -0
- parse/train/B1gjs6EtDr/B1gjs6EtDr_layout.pdf +3 -0
- parse/train/B1gjs6EtDr/B1gjs6EtDr_origin.pdf +3 -0
- parse/train/B1gjs6EtDr/B1gjs6EtDr_span.pdf +3 -0
- parse/train/B1l3M64KwB/B1l3M64KwB_layout.pdf +3 -0
- parse/train/B1l3M64KwB/B1l3M64KwB_origin.pdf +3 -0
- parse/train/B1l3M64KwB/B1l3M64KwB_span.pdf +3 -0
- parse/train/B1lgUkBFwr/B1lgUkBFwr_layout.pdf +3 -0
- parse/train/B1lgUkBFwr/B1lgUkBFwr_origin.pdf +3 -0
- parse/train/B1xtFpVtvB/B1xtFpVtvB.md +256 -0
- parse/train/B1xtFpVtvB/B1xtFpVtvB_model.json +0 -0
- parse/train/BJ6anzb0Z/BJ6anzb0Z_layout.pdf +3 -0
- parse/train/BJ6anzb0Z/BJ6anzb0Z_origin.pdf +3 -0
- parse/train/BJ6anzb0Z/BJ6anzb0Z_span.pdf +3 -0
- parse/train/BJuWrGW0Z/BJuWrGW0Z_layout.pdf +3 -0
- parse/train/BJuWrGW0Z/BJuWrGW0Z_origin.pdf +3 -0
.gitattributes
CHANGED
|
@@ -8329,3 +8329,383 @@ parse/train/rJe2syrtvS/rJe2syrtvS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
|
| 8329 |
parse/train/rJe2syrtvS/rJe2syrtvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8330 |
parse/train/rJe2syrtvS/rJe2syrtvS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8331 |
parse/train/B1lgUkBFwr/B1lgUkBFwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8329 |
parse/train/rJe2syrtvS/rJe2syrtvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8330 |
parse/train/rJe2syrtvS/rJe2syrtvS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8331 |
parse/train/B1lgUkBFwr/B1lgUkBFwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8332 |
+
parse/train/B1lgUkBFwr/B1lgUkBFwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8333 |
+
parse/train/B1lgUkBFwr/B1lgUkBFwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8334 |
+
parse/train/BJ6anzb0Z/BJ6anzb0Z_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8335 |
+
parse/train/BJ6anzb0Z/BJ6anzb0Z_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8336 |
+
parse/train/BJ6anzb0Z/BJ6anzb0Z_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8337 |
+
parse/train/J8_GttYLFgr/J8_GttYLFgr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8338 |
+
parse/train/J8_GttYLFgr/J8_GttYLFgr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8339 |
+
parse/train/J8_GttYLFgr/J8_GttYLFgr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8340 |
+
parse/train/TtYSU29zgR/TtYSU29zgR_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8341 |
+
parse/train/TtYSU29zgR/TtYSU29zgR_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8342 |
+
parse/train/TtYSU29zgR/TtYSU29zgR_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8343 |
+
parse/train/rydeCEhs-/rydeCEhs-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8344 |
+
parse/train/rydeCEhs-/rydeCEhs-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8345 |
+
parse/train/rydeCEhs-/rydeCEhs-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8346 |
+
parse/train/jEYKjPE1xYN/jEYKjPE1xYN_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8347 |
+
parse/train/jEYKjPE1xYN/jEYKjPE1xYN_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8348 |
+
parse/train/jEYKjPE1xYN/jEYKjPE1xYN_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8349 |
+
parse/train/px0-N3_KjA/px0-N3_KjA_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8350 |
+
parse/train/px0-N3_KjA/px0-N3_KjA_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8351 |
+
parse/train/px0-N3_KjA/px0-N3_KjA_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8352 |
+
parse/train/SkgsACVKPH/SkgsACVKPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8353 |
+
parse/train/SkgsACVKPH/SkgsACVKPH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8354 |
+
parse/train/SkgsACVKPH/SkgsACVKPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8355 |
+
parse/train/Hyl9ahVFwH/Hyl9ahVFwH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8356 |
+
parse/train/Hyl9ahVFwH/Hyl9ahVFwH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8357 |
+
parse/train/Hyl9ahVFwH/Hyl9ahVFwH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8358 |
+
parse/train/r1lYRjC9F7/r1lYRjC9F7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8359 |
+
parse/train/r1lYRjC9F7/r1lYRjC9F7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8360 |
+
parse/train/r1lYRjC9F7/r1lYRjC9F7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8361 |
+
parse/train/HkGSniC9FQ/HkGSniC9FQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8362 |
+
parse/train/HkGSniC9FQ/HkGSniC9FQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8363 |
+
parse/train/HkGSniC9FQ/HkGSniC9FQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8364 |
+
parse/train/jB0Nlbwlybm/jB0Nlbwlybm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8365 |
+
parse/train/jB0Nlbwlybm/jB0Nlbwlybm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8366 |
+
parse/train/jB0Nlbwlybm/jB0Nlbwlybm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8367 |
+
parse/train/SkgGCkrKvH/SkgGCkrKvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8368 |
+
parse/train/SkgGCkrKvH/SkgGCkrKvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8369 |
+
parse/train/SkgGCkrKvH/SkgGCkrKvH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8370 |
+
parse/train/rkHVZWZAZ/rkHVZWZAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8371 |
+
parse/train/rkHVZWZAZ/rkHVZWZAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8372 |
+
parse/train/rkHVZWZAZ/rkHVZWZAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8373 |
+
parse/train/rJNwDjAqYX/rJNwDjAqYX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8374 |
+
parse/train/rJNwDjAqYX/rJNwDjAqYX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8375 |
+
parse/train/rJNwDjAqYX/rJNwDjAqYX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8376 |
+
parse/train/nHRGW_wETLQ/nHRGW_wETLQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8377 |
+
parse/train/nHRGW_wETLQ/nHRGW_wETLQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8378 |
+
parse/train/nHRGW_wETLQ/nHRGW_wETLQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8379 |
+
parse/train/B18WgG-CZ/B18WgG-CZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8380 |
+
parse/train/B18WgG-CZ/B18WgG-CZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8381 |
+
parse/train/B18WgG-CZ/B18WgG-CZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8382 |
+
parse/train/rkxn7nR5KX/rkxn7nR5KX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8383 |
+
parse/train/rkxn7nR5KX/rkxn7nR5KX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8384 |
+
parse/train/rkxn7nR5KX/rkxn7nR5KX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8385 |
+
parse/train/HJe6uANtwH/HJe6uANtwH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8386 |
+
parse/train/HJe6uANtwH/HJe6uANtwH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8387 |
+
parse/train/HJe6uANtwH/HJe6uANtwH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8388 |
+
parse/train/S1gTA5VggE/S1gTA5VggE_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8389 |
+
parse/train/S1gTA5VggE/S1gTA5VggE_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8390 |
+
parse/train/S1gTA5VggE/S1gTA5VggE_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8391 |
+
parse/train/H1z-PsR5KX/H1z-PsR5KX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8392 |
+
parse/train/H1z-PsR5KX/H1z-PsR5KX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8393 |
+
parse/train/H1z-PsR5KX/H1z-PsR5KX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8394 |
+
parse/train/vVjIW3sEc1s/vVjIW3sEc1s_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8395 |
+
parse/train/vVjIW3sEc1s/vVjIW3sEc1s_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8396 |
+
parse/train/vVjIW3sEc1s/vVjIW3sEc1s_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8397 |
+
parse/train/SJgNkpVFPr/SJgNkpVFPr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8398 |
+
parse/train/SJgNkpVFPr/SJgNkpVFPr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8399 |
+
parse/train/SJgNkpVFPr/SJgNkpVFPr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8400 |
+
parse/train/QDm4QXNOsuQVE/QDm4QXNOsuQVE_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8401 |
+
parse/train/QDm4QXNOsuQVE/QDm4QXNOsuQVE_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8402 |
+
parse/train/QDm4QXNOsuQVE/QDm4QXNOsuQVE_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8403 |
+
parse/train/ttnAE7vaATtaK/ttnAE7vaATtaK_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8404 |
+
parse/train/ttnAE7vaATtaK/ttnAE7vaATtaK_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8405 |
+
parse/train/ttnAE7vaATtaK/ttnAE7vaATtaK_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8406 |
+
parse/train/S1g6xeSKDS/S1g6xeSKDS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8407 |
+
parse/train/S1g6xeSKDS/S1g6xeSKDS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8408 |
+
parse/train/S1g6xeSKDS/S1g6xeSKDS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8409 |
+
parse/train/bDP7I2fYwLj/bDP7I2fYwLj_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8410 |
+
parse/train/bDP7I2fYwLj/bDP7I2fYwLj_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8411 |
+
parse/train/bDP7I2fYwLj/bDP7I2fYwLj_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8412 |
+
parse/train/oSrM_jG_Ng/oSrM_jG_Ng_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8413 |
+
parse/train/oSrM_jG_Ng/oSrM_jG_Ng_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8414 |
+
parse/train/oSrM_jG_Ng/oSrM_jG_Ng_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8415 |
+
parse/train/BkgStySKPB/BkgStySKPB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8416 |
+
parse/train/BkgStySKPB/BkgStySKPB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8417 |
+
parse/train/BkgStySKPB/BkgStySKPB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8418 |
+
parse/train/pBKOx_dxYAN/pBKOx_dxYAN_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8419 |
+
parse/train/pBKOx_dxYAN/pBKOx_dxYAN_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8420 |
+
parse/train/pBKOx_dxYAN/pBKOx_dxYAN_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8421 |
+
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8422 |
+
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8423 |
+
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8424 |
+
parse/train/oRMRIR4qPC1/oRMRIR4qPC1_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8425 |
+
parse/train/oRMRIR4qPC1/oRMRIR4qPC1_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8426 |
+
parse/train/oRMRIR4qPC1/oRMRIR4qPC1_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8427 |
+
parse/train/rkTS8lZAb/rkTS8lZAb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8428 |
+
parse/train/rkTS8lZAb/rkTS8lZAb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8429 |
+
parse/train/rkTS8lZAb/rkTS8lZAb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8430 |
+
parse/train/r1xapAEKwS/r1xapAEKwS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8431 |
+
parse/train/r1xapAEKwS/r1xapAEKwS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8432 |
+
parse/train/r1xapAEKwS/r1xapAEKwS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8433 |
+
parse/train/wg_kD_nyAF/wg_kD_nyAF_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8434 |
+
parse/train/wg_kD_nyAF/wg_kD_nyAF_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8435 |
+
parse/train/wg_kD_nyAF/wg_kD_nyAF_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8436 |
+
parse/train/UlSjqPEkI1V/UlSjqPEkI1V_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8437 |
+
parse/train/UlSjqPEkI1V/UlSjqPEkI1V_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8438 |
+
parse/train/UlSjqPEkI1V/UlSjqPEkI1V_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8439 |
+
parse/train/rcQdycl0zyk/rcQdycl0zyk_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8440 |
+
parse/train/rcQdycl0zyk/rcQdycl0zyk_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8441 |
+
parse/train/rcQdycl0zyk/rcQdycl0zyk_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8442 |
+
parse/train/BkxUvnEYDH/BkxUvnEYDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8443 |
+
parse/train/BkxUvnEYDH/BkxUvnEYDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8444 |
+
parse/train/BkxUvnEYDH/BkxUvnEYDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8445 |
+
parse/train/tFbuFKWX3MFC8/tFbuFKWX3MFC8_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8446 |
+
parse/train/tFbuFKWX3MFC8/tFbuFKWX3MFC8_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8447 |
+
parse/train/tFbuFKWX3MFC8/tFbuFKWX3MFC8_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8448 |
+
parse/train/gwnoVHIES05/gwnoVHIES05_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8449 |
+
parse/train/gwnoVHIES05/gwnoVHIES05_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8450 |
+
parse/train/gwnoVHIES05/gwnoVHIES05_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8451 |
+
parse/train/S1nQvfgA-/S1nQvfgA-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8452 |
+
parse/train/S1nQvfgA-/S1nQvfgA-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8453 |
+
parse/train/S1nQvfgA-/S1nQvfgA-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8454 |
+
parse/train/mQPBmvyAuk/mQPBmvyAuk_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8455 |
+
parse/train/mQPBmvyAuk/mQPBmvyAuk_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8456 |
+
parse/train/mQPBmvyAuk/mQPBmvyAuk_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8457 |
+
parse/train/H1lefTEKDS/H1lefTEKDS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8458 |
+
parse/train/H1lefTEKDS/H1lefTEKDS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8459 |
+
parse/train/H1lefTEKDS/H1lefTEKDS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8460 |
+
parse/train/S1HEBe_Jl/S1HEBe_Jl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8461 |
+
parse/train/S1HEBe_Jl/S1HEBe_Jl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8462 |
+
parse/train/S1HEBe_Jl/S1HEBe_Jl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8463 |
+
parse/train/r1Ue8Hcxg/r1Ue8Hcxg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8464 |
+
parse/train/r1Ue8Hcxg/r1Ue8Hcxg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8465 |
+
parse/train/r1Ue8Hcxg/r1Ue8Hcxg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8466 |
+
parse/train/MRQJmsNPp8E/MRQJmsNPp8E_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8467 |
+
parse/train/MRQJmsNPp8E/MRQJmsNPp8E_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8468 |
+
parse/train/MRQJmsNPp8E/MRQJmsNPp8E_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8469 |
+
parse/train/nIAxjsniDzg/nIAxjsniDzg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8470 |
+
parse/train/nIAxjsniDzg/nIAxjsniDzg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8471 |
+
parse/train/nIAxjsniDzg/nIAxjsniDzg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8472 |
+
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8473 |
+
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8474 |
+
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8475 |
+
parse/train/Z3yd722b5X5/Z3yd722b5X5_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8476 |
+
parse/train/Z3yd722b5X5/Z3yd722b5X5_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8477 |
+
parse/train/Z3yd722b5X5/Z3yd722b5X5_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8478 |
+
parse/train/Syx5V2CcFm/Syx5V2CcFm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8479 |
+
parse/train/Syx5V2CcFm/Syx5V2CcFm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8480 |
+
parse/train/Syx5V2CcFm/Syx5V2CcFm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8481 |
+
parse/train/bM4Iqfg8M2k/bM4Iqfg8M2k_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8482 |
+
parse/train/bM4Iqfg8M2k/bM4Iqfg8M2k_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8483 |
+
parse/train/bM4Iqfg8M2k/bM4Iqfg8M2k_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8484 |
+
parse/train/rkeYUsRqKQ/rkeYUsRqKQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8485 |
+
parse/train/rkeYUsRqKQ/rkeYUsRqKQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8486 |
+
parse/train/rkeYUsRqKQ/rkeYUsRqKQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8487 |
+
parse/train/S1XolQbRW/S1XolQbRW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8488 |
+
parse/train/S1XolQbRW/S1XolQbRW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8489 |
+
parse/train/S1XolQbRW/S1XolQbRW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8490 |
+
parse/train/SyhRVm-Rb/SyhRVm-Rb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8491 |
+
parse/train/SyhRVm-Rb/SyhRVm-Rb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8492 |
+
parse/train/SyhRVm-Rb/SyhRVm-Rb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8493 |
+
parse/train/r1xCMyBtPS/r1xCMyBtPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8494 |
+
parse/train/r1xCMyBtPS/r1xCMyBtPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8495 |
+
parse/train/r1xCMyBtPS/r1xCMyBtPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8496 |
+
parse/train/Ja-hVQrfeGZ/Ja-hVQrfeGZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8497 |
+
parse/train/Ja-hVQrfeGZ/Ja-hVQrfeGZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8498 |
+
parse/train/Ja-hVQrfeGZ/Ja-hVQrfeGZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8499 |
+
parse/train/GzHjhdpk-YH/GzHjhdpk-YH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8500 |
+
parse/train/GzHjhdpk-YH/GzHjhdpk-YH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8501 |
+
parse/train/GzHjhdpk-YH/GzHjhdpk-YH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8502 |
+
parse/train/9p2ekP904Rs/9p2ekP904Rs_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8503 |
+
parse/train/9p2ekP904Rs/9p2ekP904Rs_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8504 |
+
parse/train/9p2ekP904Rs/9p2ekP904Rs_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8505 |
+
parse/train/r1Usiwcex/r1Usiwcex_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8506 |
+
parse/train/r1Usiwcex/r1Usiwcex_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8507 |
+
parse/train/r1Usiwcex/r1Usiwcex_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8508 |
+
parse/train/rkg98yBFDr/rkg98yBFDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8509 |
+
parse/train/rkg98yBFDr/rkg98yBFDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8510 |
+
parse/train/rkg98yBFDr/rkg98yBFDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8511 |
+
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8512 |
+
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8513 |
+
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8514 |
+
parse/train/tc5qisoB-C/tc5qisoB-C_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8515 |
+
parse/train/tc5qisoB-C/tc5qisoB-C_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8516 |
+
parse/train/tc5qisoB-C/tc5qisoB-C_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8517 |
+
parse/train/HktRlUlAZ/HktRlUlAZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8518 |
+
parse/train/HktRlUlAZ/HktRlUlAZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8519 |
+
parse/train/HktRlUlAZ/HktRlUlAZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8520 |
+
parse/train/SygLehCqtm/SygLehCqtm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8521 |
+
parse/train/SygLehCqtm/SygLehCqtm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8522 |
+
parse/train/SygLehCqtm/SygLehCqtm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8523 |
+
parse/train/Rd138pWXMvG/Rd138pWXMvG_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8524 |
+
parse/train/Rd138pWXMvG/Rd138pWXMvG_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8525 |
+
parse/train/Rd138pWXMvG/Rd138pWXMvG_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8526 |
+
parse/train/HyGTuv9eg/HyGTuv9eg_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8527 |
+
parse/train/HyGTuv9eg/HyGTuv9eg_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8528 |
+
parse/train/HyGTuv9eg/HyGTuv9eg_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8529 |
+
parse/train/SygXPaEYvH/SygXPaEYvH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8530 |
+
parse/train/SygXPaEYvH/SygXPaEYvH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8531 |
+
parse/train/SygXPaEYvH/SygXPaEYvH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8532 |
+
parse/train/SJzwb2RcK7/SJzwb2RcK7_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8533 |
+
parse/train/SJzwb2RcK7/SJzwb2RcK7_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8534 |
+
parse/train/SJzwb2RcK7/SJzwb2RcK7_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8535 |
+
parse/train/B1eY_pVYvB/B1eY_pVYvB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8536 |
+
parse/train/B1eY_pVYvB/B1eY_pVYvB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8537 |
+
parse/train/B1eY_pVYvB/B1eY_pVYvB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8538 |
+
parse/train/NJS8kp15zzH/NJS8kp15zzH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8539 |
+
parse/train/NJS8kp15zzH/NJS8kp15zzH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8540 |
+
parse/train/NJS8kp15zzH/NJS8kp15zzH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8541 |
+
parse/train/ryewE3R5YX/ryewE3R5YX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8542 |
+
parse/train/ryewE3R5YX/ryewE3R5YX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8543 |
+
parse/train/ryewE3R5YX/ryewE3R5YX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8544 |
+
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8545 |
+
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8546 |
+
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8547 |
+
parse/train/Hke0V1rKPS/Hke0V1rKPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8548 |
+
parse/train/Hke0V1rKPS/Hke0V1rKPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8549 |
+
parse/train/Hke0V1rKPS/Hke0V1rKPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8550 |
+
parse/train/rkgvXlrKwH/rkgvXlrKwH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8551 |
+
parse/train/rkgvXlrKwH/rkgvXlrKwH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8552 |
+
parse/train/rkgvXlrKwH/rkgvXlrKwH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8553 |
+
parse/train/BJuWrGW0Z/BJuWrGW0Z_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8554 |
+
parse/train/BJuWrGW0Z/BJuWrGW0Z_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8555 |
+
parse/train/BJuWrGW0Z/BJuWrGW0Z_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8556 |
+
parse/train/ETBc_MIMgoX/ETBc_MIMgoX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8557 |
+
parse/train/ETBc_MIMgoX/ETBc_MIMgoX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8558 |
+
parse/train/ETBc_MIMgoX/ETBc_MIMgoX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8559 |
+
parse/train/ryxWIgBFPS/ryxWIgBFPS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8560 |
+
parse/train/ryxWIgBFPS/ryxWIgBFPS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8561 |
+
parse/train/ryxWIgBFPS/ryxWIgBFPS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8562 |
+
parse/train/d3k38LTDCyO/d3k38LTDCyO_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8563 |
+
parse/train/d3k38LTDCyO/d3k38LTDCyO_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8564 |
+
parse/train/d3k38LTDCyO/d3k38LTDCyO_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8565 |
+
parse/train/BkxnKkrtvS/BkxnKkrtvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8566 |
+
parse/train/BkxnKkrtvS/BkxnKkrtvS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8567 |
+
parse/train/BkxnKkrtvS/BkxnKkrtvS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8568 |
+
parse/train/r1q7n9gAb/r1q7n9gAb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8569 |
+
parse/train/r1q7n9gAb/r1q7n9gAb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8570 |
+
parse/train/r1q7n9gAb/r1q7n9gAb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8571 |
+
parse/train/HJaDJZ-0W/HJaDJZ-0W_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8572 |
+
parse/train/HJaDJZ-0W/HJaDJZ-0W_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8573 |
+
parse/train/HJaDJZ-0W/HJaDJZ-0W_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8574 |
+
parse/train/rkgOLb-0W/rkgOLb-0W_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8575 |
+
parse/train/rkgOLb-0W/rkgOLb-0W_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8576 |
+
parse/train/rkgOLb-0W/rkgOLb-0W_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8577 |
+
parse/train/pbXQtKXwLS/pbXQtKXwLS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8578 |
+
parse/train/pbXQtKXwLS/pbXQtKXwLS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8579 |
+
parse/train/pbXQtKXwLS/pbXQtKXwLS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8580 |
+
parse/train/B1gjs6EtDr/B1gjs6EtDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8581 |
+
parse/train/B1gjs6EtDr/B1gjs6EtDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8582 |
+
parse/train/B1gjs6EtDr/B1gjs6EtDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8583 |
+
parse/train/B1l3M64KwB/B1l3M64KwB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8584 |
+
parse/train/B1l3M64KwB/B1l3M64KwB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8585 |
+
parse/train/B1l3M64KwB/B1l3M64KwB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8586 |
+
parse/train/fylclEqgvgd/fylclEqgvgd_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8587 |
+
parse/train/fylclEqgvgd/fylclEqgvgd_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8588 |
+
parse/train/fylclEqgvgd/fylclEqgvgd_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8589 |
+
parse/train/HJl0jiRqtX/HJl0jiRqtX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8590 |
+
parse/train/HJl0jiRqtX/HJl0jiRqtX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8591 |
+
parse/train/HJl0jiRqtX/HJl0jiRqtX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8592 |
+
parse/train/BJx8YnEFPH/BJx8YnEFPH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8593 |
+
parse/train/BJx8YnEFPH/BJx8YnEFPH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8594 |
+
parse/train/BJx8YnEFPH/BJx8YnEFPH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8595 |
+
parse/train/Byl5NREFDr/Byl5NREFDr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8596 |
+
parse/train/Byl5NREFDr/Byl5NREFDr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8597 |
+
parse/train/Byl5NREFDr/Byl5NREFDr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8598 |
+
parse/train/By41BjA9YQ/By41BjA9YQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8599 |
+
parse/train/By41BjA9YQ/By41BjA9YQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8600 |
+
parse/train/By41BjA9YQ/By41BjA9YQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8601 |
+
parse/train/HylVB3AqYm/HylVB3AqYm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8602 |
+
parse/train/HylVB3AqYm/HylVB3AqYm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8603 |
+
parse/train/HylVB3AqYm/HylVB3AqYm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8604 |
+
parse/train/rkYTTf-AZ/rkYTTf-AZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8605 |
+
parse/train/rkYTTf-AZ/rkYTTf-AZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8606 |
+
parse/train/rkYTTf-AZ/rkYTTf-AZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8607 |
+
parse/train/HyexAiA5Fm/HyexAiA5Fm_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8608 |
+
parse/train/HyexAiA5Fm/HyexAiA5Fm_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8609 |
+
parse/train/HyexAiA5Fm/HyexAiA5Fm_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8610 |
+
parse/train/uSQQH7Fj5U/uSQQH7Fj5U_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8611 |
+
parse/train/uSQQH7Fj5U/uSQQH7Fj5U_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8612 |
+
parse/train/uSQQH7Fj5U/uSQQH7Fj5U_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8613 |
+
parse/train/rkxQ-nA9FX/rkxQ-nA9FX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8614 |
+
parse/train/rkxQ-nA9FX/rkxQ-nA9FX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8615 |
+
parse/train/rkxQ-nA9FX/rkxQ-nA9FX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8616 |
+
parse/train/HygBZnRctX/HygBZnRctX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8617 |
+
parse/train/HygBZnRctX/HygBZnRctX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8618 |
+
parse/train/HygBZnRctX/HygBZnRctX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8619 |
+
parse/train/SJky6Ry0W/SJky6Ry0W_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8620 |
+
parse/train/SJky6Ry0W/SJky6Ry0W_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8621 |
+
parse/train/SJky6Ry0W/SJky6Ry0W_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8622 |
+
parse/train/ry831QWAb/ry831QWAb_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8623 |
+
parse/train/ry831QWAb/ry831QWAb_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8624 |
+
parse/train/ry831QWAb/ry831QWAb_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8625 |
+
parse/train/v5gjXpmR8J/v5gjXpmR8J_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8626 |
+
parse/train/v5gjXpmR8J/v5gjXpmR8J_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8627 |
+
parse/train/v5gjXpmR8J/v5gjXpmR8J_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8628 |
+
parse/train/S1xSSTNKDB/S1xSSTNKDB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8629 |
+
parse/train/S1xSSTNKDB/S1xSSTNKDB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8630 |
+
parse/train/S1xSSTNKDB/S1xSSTNKDB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8631 |
+
parse/train/n4IMHNb8_f/n4IMHNb8_f_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8632 |
+
parse/train/n4IMHNb8_f/n4IMHNb8_f_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8633 |
+
parse/train/n4IMHNb8_f/n4IMHNb8_f_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8634 |
+
parse/train/HJcSzz-CZ/HJcSzz-CZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8635 |
+
parse/train/HJcSzz-CZ/HJcSzz-CZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8636 |
+
parse/train/HJcSzz-CZ/HJcSzz-CZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8637 |
+
parse/train/Sy3fJXbA-/Sy3fJXbA-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8638 |
+
parse/train/Sy3fJXbA-/Sy3fJXbA-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8639 |
+
parse/train/Sy3fJXbA-/Sy3fJXbA-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8640 |
+
parse/train/yILzFBjR0Y/yILzFBjR0Y_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8641 |
+
parse/train/yILzFBjR0Y/yILzFBjR0Y_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8642 |
+
parse/train/yILzFBjR0Y/yILzFBjR0Y_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8643 |
+
parse/train/86NHK__yFDl/86NHK__yFDl_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8644 |
+
parse/train/86NHK__yFDl/86NHK__yFDl_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8645 |
+
parse/train/86NHK__yFDl/86NHK__yFDl_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8646 |
+
parse/train/eXlxB3aLOe/eXlxB3aLOe_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8647 |
+
parse/train/eXlxB3aLOe/eXlxB3aLOe_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8648 |
+
parse/train/eXlxB3aLOe/eXlxB3aLOe_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8649 |
+
parse/train/YwpZmcAehZ/YwpZmcAehZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8650 |
+
parse/train/YwpZmcAehZ/YwpZmcAehZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8651 |
+
parse/train/YwpZmcAehZ/YwpZmcAehZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8652 |
+
parse/train/3qMwV98zLIk/3qMwV98zLIk_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8653 |
+
parse/train/3qMwV98zLIk/3qMwV98zLIk_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8654 |
+
parse/train/3qMwV98zLIk/3qMwV98zLIk_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8655 |
+
parse/train/HJgJtT4tvB/HJgJtT4tvB_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8656 |
+
parse/train/HJgJtT4tvB/HJgJtT4tvB_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8657 |
+
parse/train/HJgJtT4tvB/HJgJtT4tvB_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8658 |
+
parse/train/H1e8wsCqYX/H1e8wsCqYX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8659 |
+
parse/train/H1e8wsCqYX/H1e8wsCqYX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8660 |
+
parse/train/H1e8wsCqYX/H1e8wsCqYX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8661 |
+
parse/train/F1vEjWK-lH_/F1vEjWK-lH__span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8662 |
+
parse/train/F1vEjWK-lH_/F1vEjWK-lH__layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8663 |
+
parse/train/F1vEjWK-lH_/F1vEjWK-lH__origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8664 |
+
parse/train/SyxS0T4tvS/SyxS0T4tvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8665 |
+
parse/train/SyxS0T4tvS/SyxS0T4tvS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8666 |
+
parse/train/SyxS0T4tvS/SyxS0T4tvS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8667 |
+
parse/train/a2Gr9gNFD-J/a2Gr9gNFD-J_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8668 |
+
parse/train/a2Gr9gNFD-J/a2Gr9gNFD-J_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8669 |
+
parse/train/a2Gr9gNFD-J/a2Gr9gNFD-J_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8670 |
+
parse/train/SkgewU5ll/SkgewU5ll_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8671 |
+
parse/train/SkgewU5ll/SkgewU5ll_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8672 |
+
parse/train/SkgewU5ll/SkgewU5ll_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8673 |
+
parse/train/l2UWXn5iBQI/l2UWXn5iBQI_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8674 |
+
parse/train/l2UWXn5iBQI/l2UWXn5iBQI_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8675 |
+
parse/train/l2UWXn5iBQI/l2UWXn5iBQI_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8676 |
+
parse/train/LyjH88yV7F/LyjH88yV7F_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8677 |
+
parse/train/LyjH88yV7F/LyjH88yV7F_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8678 |
+
parse/train/LyjH88yV7F/LyjH88yV7F_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8679 |
+
parse/train/H1WgVz-AZ/H1WgVz-AZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8680 |
+
parse/train/H1WgVz-AZ/H1WgVz-AZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8681 |
+
parse/train/H1WgVz-AZ/H1WgVz-AZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8682 |
+
parse/train/JXAyJeYqUkZ/JXAyJeYqUkZ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8683 |
+
parse/train/JXAyJeYqUkZ/JXAyJeYqUkZ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8684 |
+
parse/train/JXAyJeYqUkZ/JXAyJeYqUkZ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8685 |
+
parse/train/HkxJpnVtPr/HkxJpnVtPr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8686 |
+
parse/train/HkxJpnVtPr/HkxJpnVtPr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8687 |
+
parse/train/HkxJpnVtPr/HkxJpnVtPr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8688 |
+
parse/train/S1gUsoR9YX/S1gUsoR9YX_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8689 |
+
parse/train/S1gUsoR9YX/S1gUsoR9YX_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8690 |
+
parse/train/S1gUsoR9YX/S1gUsoR9YX_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8691 |
+
parse/train/SJQO7UJCW/SJQO7UJCW_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8692 |
+
parse/train/SJQO7UJCW/SJQO7UJCW_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8693 |
+
parse/train/SJQO7UJCW/SJQO7UJCW_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8694 |
+
parse/train/SJe5P6EYvS/SJe5P6EYvS_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8695 |
+
parse/train/SJe5P6EYvS/SJe5P6EYvS_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8696 |
+
parse/train/SJe5P6EYvS/SJe5P6EYvS_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8697 |
+
parse/train/27acGyyI1BY/27acGyyI1BY_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8698 |
+
parse/train/27acGyyI1BY/27acGyyI1BY_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8699 |
+
parse/train/27acGyyI1BY/27acGyyI1BY_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8700 |
+
parse/train/H1loF2NFwr/H1loF2NFwr_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8701 |
+
parse/train/H1loF2NFwr/H1loF2NFwr_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8702 |
+
parse/train/H1loF2NFwr/H1loF2NFwr_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8703 |
+
parse/train/rkl03ySYDH/rkl03ySYDH_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8704 |
+
parse/train/rkl03ySYDH/rkl03ySYDH_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8705 |
+
parse/train/rkl03ySYDH/rkl03ySYDH_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8706 |
+
parse/train/BydjJte0-/BydjJte0-_span.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8707 |
+
parse/train/BydjJte0-/BydjJte0-_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8708 |
+
parse/train/BydjJte0-/BydjJte0-_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8709 |
+
parse/train/GWRkOYr4jxQ/GWRkOYr4jxQ_layout.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8710 |
+
parse/train/GWRkOYr4jxQ/GWRkOYr4jxQ_origin.pdf filter=lfs diff=lfs merge=lfs -text
|
| 8711 |
+
parse/train/GWRkOYr4jxQ/GWRkOYr4jxQ_span.pdf filter=lfs diff=lfs merge=lfs -text
|
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fe98ec492152afc5a58bea0c6a674b928b5af490346bce3c7b227f9b40b9aa9e
|
| 3 |
+
size 966221
|
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0073f058972c0816a2a91647b72c35a095525c5fc8b0c2dd24689dc9cdfb289e
|
| 3 |
+
size 788584
|
parse/train/1YLJDvSx6J4/1YLJDvSx6J4_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8242d5458ab77eb72f64a9d603ceb4f54efbe4918b74841072315b204de35db5
|
| 3 |
+
size 972725
|
parse/train/27acGyyI1BY/27acGyyI1BY_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:493630f26aeab14832941d296625b367b5040a6a727251db350f4e10818d11f7
|
| 3 |
+
size 2478305
|
parse/train/27acGyyI1BY/27acGyyI1BY_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b41bbaffc7006d27995f62d07870b3441956f50b06f94c2f05b3436638d725a7
|
| 3 |
+
size 2250785
|
parse/train/27acGyyI1BY/27acGyyI1BY_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:76cfaaccab7d8504aa874c81c004ac8d5daca3c4a7d3e8a21f2a4b89ac3988fa
|
| 3 |
+
size 2490330
|
parse/train/3qMwV98zLIk/3qMwV98zLIk_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f8f0384f63311838b1f8057ce150a0c78dcaf775cd06b4d6ff86ce2de9bb3a61
|
| 3 |
+
size 742190
|
parse/train/3qMwV98zLIk/3qMwV98zLIk_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:83d902b095c37e0876b853ec7a34c92ff68edbc2d96d7135c7502fcb6245ed66
|
| 3 |
+
size 614488
|
parse/train/3qMwV98zLIk/3qMwV98zLIk_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:081a28abc25cf0c01238041e11952e4a78c528f0f8afbb509ebad5d6647824f3
|
| 3 |
+
size 750816
|
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2c3f0b199b2b2011b361d528bd3b55285fe467249880ee1a82a75d7c8cdff3e3
|
| 3 |
+
size 1985436
|
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aec0d43b4de91a4f19378a1e26c915721eb64e63852b0b35132238daf0ebb9b7
|
| 3 |
+
size 1803427
|
parse/train/6EWOVxvJ5FI/6EWOVxvJ5FI_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:85eba4d4cd1b539c4a340c3d44ff8ee2efebfb3c69388f1c51249472a3b43dd4
|
| 3 |
+
size 2008883
|
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6900941d6d9c112bd78dbe92eeb0d8432dcbd1e1ede4e0d44aadedf87554d60a
|
| 3 |
+
size 1398026
|
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c483e4be665a1c53607924c40c5ed6e3db5b538ad7fe8958fb3318bd1401af0b
|
| 3 |
+
size 814459
|
parse/train/6nbpPqUCIi7/6nbpPqUCIi7_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6459fb2099cd7007c5fddc80e45a1bf74cd9ab5dcdf65f8ca514443e8ae8b204
|
| 3 |
+
size 1409404
|
parse/train/86NHK__yFDl/86NHK__yFDl_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:69e586ac4d135bb618ca484307f904c2dfa70828604c302e6d9222ca2751517f
|
| 3 |
+
size 1871622
|
parse/train/86NHK__yFDl/86NHK__yFDl_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c6a23e1ab3f98f6c46b2ea6aa892c47cbd6015438cf94e5068b1ace1a0dc63c0
|
| 3 |
+
size 1699827
|
parse/train/86NHK__yFDl/86NHK__yFDl_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:610f01bfd9e7afef9eae063f43dac35f2480a7d238b4481f876ce70a15bcfc31
|
| 3 |
+
size 1873835
|
parse/train/8PA2nX9v_r2/8PA2nX9v_r2.md
ADDED
|
@@ -0,0 +1,269 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Aligning Pretraining for Detection via Object-Level Contrastive Learning
|
| 2 |
+
|
| 3 |
+
Fangyun Wei∗ Yue Gao∗ Zhirong Wu Han Hu Stephen Lin
|
| 4 |
+
|
| 5 |
+
Microsoft Research Asia {fawe, yuegao, wuzhiron, hanhu, stevelin}@microsoft.com
|
| 6 |
+
|
| 7 |
+
# Abstract
|
| 8 |
+
|
| 9 |
+
Image-level contrastive representation learning has proven to be highly effective as a generic model for transfer learning. Such generality for transfer learning, however, sacrifices specificity if we are interested in a certain downstream task. We argue that this could be sub-optimal and thus advocate a design principle which encourages alignment between the self-supervised pretext task and the downstream task. In this paper, we follow this principle with a pretraining method specifically designed for the task of object detection. We attain alignment in the following three aspects: 1) object-level representations are introduced via selective search bounding boxes as object proposals; 2) the pretraining network architecture incorporates the same dedicated modules used in the detection pipeline (e.g. FPN); 3) the pretraining is equipped with object detection properties such as object-level translation invariance and scale invariance. Our method, called Selective Object COntrastive learning (SoCo), achieves state-of-the-art results for transfer performance on COCO detection using a Mask R-CNN framework. Code is available at https://github.com/hologerry/SoCo.
|
| 10 |
+
|
| 11 |
+
# 1 Introduction
|
| 12 |
+
|
| 13 |
+
Pretraining and finetuning has been the dominant paradigm of training deep neural networks in computer vision. Downstream tasks usually leverage pretrained weights learned on large labeled datasets such as ImageNet [1] for initialization. As a result, supervised ImageNet pretraining has been prevalent throughout the field. Recently, self-supervised pretraining [2, 3, 4, 5, 6, 7, 8, 9] has achieved considerable progress and alleviated the dependency on labeled data. These methods aim to learn generic visual representations for various downstream tasks by means of image-level pretext tasks, such as instance discrimination. Some recent works [10, 11, 12, 13] observe that the image-level representations are sub-optimal for dense prediction tasks such as object detection and semantic segmentation. A potential reason is that image-level pretraining may overfit to holistic representations and fail to learn properties that are important outside of image classification.
|
| 14 |
+
|
| 15 |
+
The goal of this work is to develop self-supervised pretraining that is aligned to object detection. In object detection, bounding boxes are widely adopted as the representation for objects. Translation and scale invariance for object detection are reflected by the location and size of the bounding boxes. An obvious representation gap exists between image-level pretraining and the object-level bounding boxes of object detection.
|
| 16 |
+
|
| 17 |
+
Motivated by this, we present an object-level self-supervised pretraining framework, called Selective Object COntrastive learning (SoCo), specifically for the downstream task of object detection. To introduce object-level representations into pretraining, SoCo utilizes off-the-shelf selective search [14] to generate object proposals. Different from prior image-level contrastive learning methods which treat the whole image as an instance, SoCo treats each object proposal in the image as an independent instance. This enables us to design a new pretext task for learning object-level visual representations with properties that are compatible with object detection. Specifically, SoCo constructs object-level views where the scales and locations of the same object instance are augmented. Contrastive learning follows to maximize the similarity of the object across augmented views.
|
| 18 |
+
|
| 19 |
+
The introduction of the object-level representation also allows us to further bridge the gap in network architecture between pretraining and finetuning. Object detection often involves dedicated modules, e.g., feature pyramid network (FPN) [15], and special-purpose sub-networks, e.g., R-CNN head [16, 17]. In contrast to image-level contrastive learning methods where only feature backbones are pretrained and transferred, SoCo performs pretraining over all the network modules used in detectors. As a result, all layers of the detectors can be well-initialized.
|
| 20 |
+
|
| 21 |
+
Experimentally, the proposed SoCo achieves state-of-the-art transfer performance from ImageNet to COCO. Concretely, by using Mask R-CNN with an R50-FPN backbone, it obtains $4 3 . 2 \ : \mathrm { A P ^ { \tilde { b } b } } / 3 8 . 4$ $\mathbf { A P } ^ { \mathrm { m k } }$ on COCO with a $1 \times$ schedule, which are $+ 4 . 3 \ \mathrm { A P } ^ { \mathrm { b b } }$ $/ + 3 . 0 \mathrm { \ A P ^ { m k } }$ better than the supervised pretraining baseline, and $4 4 . 3 \ : \mathrm { A P ^ { b b } } / 3 9 . 6 \ : \mathrm { A P ^ { m k } }$ on COCO with a $2 \times$ schedule, which are $+ \dot { 3 } . 0 \mathrm { A P ^ { b b } }$ $\bar { / } + 2 . 3 \ \mathrm { A P ^ { \mathrm { i n k } } }$ better than the supervised pretraining baseline. When transferred to Mask R-CNN with an R50-C4 backbone, it achieves $4 0 . 9 \ : \mathrm { \dot { A } P ^ { b b } } / 3 5 . 3 \ : \mathrm { A P ^ { m k } }$ and $4 2 . 0 \mathrm { A P ^ { b b } } / 3 6 . 3 \mathrm { A P ^ { m k } }$ on the $1 \times$ and $2 \times$ schedule, respectively.
|
| 22 |
+
|
| 23 |
+
# 2 Related Work
|
| 24 |
+
|
| 25 |
+
Unsupervised feature learning has a long history in training deep neural networks. Auto-encoders [18] and Deep Boltzmann Machines [19] learn layers of representations by reconstructing image pixels. Unsupervised pretraining is often used as a method for initialization, which is followed by supervised training on the same dataset such as handwritten digits for classification.
|
| 26 |
+
|
| 27 |
+
Recent advances in unsupervised learning, especially self-supervised learning, tend to formulate the problem in a transfer learning scenario, where pretraining and finetuning are trained on different datasets with different purposes. Pretext tasks such as colorization [20], context prediction [21], inpainting [22] and rotation prediction [23] force the network to learn semantic information in order to solve the pretext tasks. Contrastive methods based on the instance discrimination pretext task [2] learn to map augmented views of the same instance to similar embeddings. Such approaches [24, 4, 5, 9, 25, 26] have shown strong transfer ability for a number of downstream tasks, sometimes even outperforming the supervised counterpart by large margins [4].
|
| 28 |
+
|
| 29 |
+
With new technical innovations such as learning clustering assignments [7], large-scale contrastive learning has been successfully applied on non-curated data sets [27, 28]. While the linear evaluation result on ImageNet classification has improved significantly [3], the progress on transfer performance for dense prediction tasks has been limited. Due to this, a growing number of works investigate pretraining specifically for object detection and semantic segmentation. The idea is to shift imagelevel representations to pixel-level or region-level representations. VADer [29], PixPro [10] and DenseCL [11] propose to learn pixel-level representations by matching point features of the same physical location under different views. InsLoc [12] learns to match region-level features from composited imagery. DetCon [13] additionally uses the bottom-up segmentation of MCG [30] for supervising intra-segment pixel variations. UP-DETR [31] proposes a pretext task named random query patch detection to pretrain DETR detector. Self-EMD [32] explores self-supervised representation learning for object detection without using Imagenet images.
|
| 30 |
+
|
| 31 |
+
Our work is also motivated and inspired by object detection methods which advocate architectures and training schemes invariant to translation and scale transformations [15, 33]. We find that incorporating them in the pretraining stage also benefits object detection, which has largely been overlooked in previous self-supervised learning works.
|
| 32 |
+
|
| 33 |
+
# 3 Method
|
| 34 |
+
|
| 35 |
+
We introduce a new contrastive learning method which maximizes the similarity of object-level features representing different augmentations of the same object. Meanwhile, we introduce architectural alignment and important properties of object detection into pretraining when designing the objectlevel pretext task. We use an example where transfer learning is performed on Mask-RCNN with a R50-FPN backbone to illustrate the core design principles. To further demonstrate the extensibility and flexibility of our method, we also apply SoCo on a R50-C4 structure. In this section, we first present an overview of the proposed SoCo in Section 3.1. Then we describe the process of object proposal generation and view construction in Section 3.2. Finally, the object-level contrastive learning and our design principles are introduced in Section 3.3.
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
Figure 1: Overview of SoCo. SoCo utilizes selective search to generate a set of object proposals for each raw image. $K$ proposals are randomly selected in each training step. We construct three views $\left\{ V _ { 1 } , V _ { 2 } , V _ { 3 } \right\}$ where the scales and locations of the same object are different. We adopt a backbone with FPN to encode image-level features and RoIAlign to extract object-level features. Object proposals are assigned to different pyramid levels according to their image areas. Contrastive learning is performed at the object level to learn translation-invariant and scale-invariant representations. The target network is updated by an exponential moving average of the online network.
|
| 39 |
+
|
| 40 |
+
# 3.1 Overview
|
| 41 |
+
|
| 42 |
+
Figure 1 displays an overview of SoCo. SoCo aims to align pretraining to object detection in two aspects: 1) network architecture alignment between pretraining and object detection; 2) introducing central properties of detection. Concretely, besides pretraining a backbone as done in existing selfsupervised contrastive learning methods, SoCo also pretrains all the network modules used in an object detector, such as FPN and the head in the Mask R-CNN framework. As a result, all layers of the detector can be well-initialized. Furthermore, SoCo strives to learn object-level representations which are not only meaningful, but also invariant to translation and scale. To achieve this, it encourages diversity of scales and locations of objects by constructing multiple augmented views and applying a scale-aware assignment strategy for different levels of a feature pyramid. Finally, object-level contrastive learning is applied to maximize the feature similarity of the same object across augmented views.
|
| 43 |
+
|
| 44 |
+
# 3.2 Data Preprocessing
|
| 45 |
+
|
| 46 |
+
Object Proposal Generation. Inspired by R-CNN [34] and Fast R-CNN [35], we use selective search [14], an unsupervised object proposal generation algorithm which takes into account color similarity, texture similarity, size of region and fit between regions, to generate a set of object proposals for each of the raw images. We represent each object proposal as a bounding box $b \overset { \cdot } { = } \{ x , y , w , h \}$ , where $( x , y )$ denotes the coordinates of the bounding box center, and $w$ and $h$ are the corresponding width and height, respectively. We keep only the proposals that satisfy the following requirements:√ √ 1) $1 / 3 \leq w / h \leq 3 ;$ 2) $0 . 3 \leq \dot { \sqrt { w h } } / \sqrt { W H } \leq 0 . 8$ , where $W$ and $H$ denote width and height of the input image. The object proposal generation step is performed offline. In each training iteration, we randomly select $K$ proposals for each input image.
|
| 47 |
+
|
| 48 |
+
View Construction. Three views, namely $V _ { 1 }$ , $V _ { 2 }$ and $V _ { 3 }$ , are used in SoCo. The input image is resized to $2 2 4 \times 2 2 4$ to obtain $V _ { 1 }$ . Then we apply a random crop with a scale range of [0.5, 1.0] on $V _ { 1 }$ in generating $V _ { 2 }$ . $V _ { 2 }$ is then resized to the same size as $V _ { 1 }$ and object proposals outside of $V _ { 2 }$ are dropped. Next, we downsample $V _ { 2 }$ to a fixed size (e.g. $1 1 2 \times 1 1 2 ,$ ) to produce $V _ { 3 }$ . In all of these cases, the bounding boxes transformed according to the cropping and resizing of the RGB images (see Figure 1, Data Preprocessing). Finally, each view is randomly and independently augmented. We adopt the augmentation pipeline of BYOL [3] but discard the random crop augmentation since spatial transformation is already applied on all three views. Notice that the scale and location of the same object proposal are different across the augmented views, which enables the model to learn translation-invariant and scale-invariant object-level representations.
|
| 49 |
+
|
| 50 |
+
Box Jitter. To further encourage variance of scales and locations of object proposals across views, we adopt a box jitter strategy on the generated proposals as an object-level data augmentation. Specifically, given an object proposal $\boldsymbol { b } = \{ x , y , w , h \}$ , we randomly generate a jittered box $\hat { b } =$ $\{ \hat { x } , \hat { y } , \hat { w } , \hat { h } \}$ as follows: 1) $\hat { x } = x + r \cdot w$ ; 2) ${ \hat { y } } = y + r \cdot h ;$ 3) $\boldsymbol { \hat { w } } = \boldsymbol { w } + \boldsymbol { r } \cdot \boldsymbol { w } ;$ 4) $\hat { h } = h + r \cdot h$ , where $r \in [ - 0 . 1 , 0 . 1 ]$ . The box jitter is randomly applied on each proposal with a probability of 0.5.
|
| 51 |
+
|
| 52 |
+
# 3.3 Object-Level Contrastive Learning
|
| 53 |
+
|
| 54 |
+
The goal of SoCo is to align pretraining to object detection. Here, we use the representative framework Mask R-CNN [17] with feature pyramid network (FPN) [15] to demonstrate our key design principles. The alignment mainly involves aligning the pretraining architecture with that of object detection and integrating important object detection properties such as object-level translation invariance and scale invariance into the pretraining.
|
| 55 |
+
|
| 56 |
+
Aligning Pretraining Architecture to Object Detection. Following Mask R-CNN, we use a backbone with FPN as the image-level feature extractor $f ^ { I }$ . We denote the output of FPN as $\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \}$ with a stride of $\{ 4 , 8 , 1 6 , 3 2 \}$ . Here, we do not use $P _ { 6 }$ due to its low resolution. With the bounding box representation $b$ , RoIAlign [17] is applied to extract the foreground feature from the corresponding scale level. For further architectural alignment, we additionally introduce an R-CNN head $f ^ { H }$ into pretraining. The object-level feature representation $h$ of bounding box $b$ is extracted from an image view $V$ as:
|
| 57 |
+
|
| 58 |
+
$$
|
| 59 |
+
h = f ^ { H } ( \mathrm { R o I A l i g n } ( f ^ { I } ( V ) , b ) ) .
|
| 60 |
+
$$
|
| 61 |
+
|
| 62 |
+
SoCo uses two neural networks to learn, namely an online network and a target network. The online network and the target network share the same architecture but with different sets of weights. Concretely, the target network weights $f _ { \xi } ^ { I } , f _ { \xi } ^ { H }$ are the exponential moving average (EMA) with the momentum coefficient $\tau$ of the online parameters $f _ { \theta } ^ { I }$ and $f _ { \theta } ^ { H }$ . Denote the set of object proposals $\{ b _ { i } \}$ considered in the image. Let $h _ { i }$ be the object-level representation of proposal $b _ { i }$ in view $V _ { 1 }$ , and $h _ { i } ^ { \prime } , h _ { i } ^ { \prime \prime }$ be the representation of $b _ { i }$ in view $V _ { 2 } , V _ { 3 }$ . They are extracted using the online network and the target network respectively,
|
| 63 |
+
|
| 64 |
+
$$
|
| 65 |
+
h _ { i } = f _ { \theta } ^ { H } \big ( \mathrm { R o I A l i g n } \big ( f _ { \theta } ^ { I } ( V _ { 1 } ) , b _ { i } \big ) \big ) ,
|
| 66 |
+
$$
|
| 67 |
+
|
| 68 |
+
$$
|
| 69 |
+
h _ { i } ^ { \prime } = f _ { \xi } ^ { H } ( \mathrm { R o I A l i g n } ( f _ { \xi } ^ { I } ( V _ { 2 } ) , b _ { i } ) ) , \quad h _ { i } ^ { \prime \prime } = f _ { \xi } ^ { H } ( \mathrm { R o I A l i g n } ( f _ { \xi } ^ { I } ( V _ { 3 } ) , b _ { i } ) ) .
|
| 70 |
+
$$
|
| 71 |
+
|
| 72 |
+
We follow BYOL [3] for learning contrastive representations. The online network is appended with a projector $g _ { \theta }$ , and a predictor $q _ { \theta }$ for obtaining latent embeddings. Both $g _ { \theta }$ and $q _ { \theta }$ are two-layer MLPs. The target network is only appended with the projector $g _ { \xi }$ for avoiding trivial solutions. We use $v _ { i }$ , $\boldsymbol { v } _ { i } ^ { \prime }$ and $v _ { i } ^ { \prime \prime }$ to denote the latent embeddings of object-level representations $\{ h _ { i } , h _ { i } ^ { \prime } , h _ { i } ^ { \prime \prime } \}$ , respectively:
|
| 73 |
+
|
| 74 |
+
$$
|
| 75 |
+
v _ { i } = q _ { \theta } ( g _ { \theta } ( h _ { i } ) ) , \quad v _ { i } ^ { \prime } = g _ { \xi } ( h _ { i } ^ { \prime } ) , \quad v _ { i } ^ { \prime \prime } = g _ { \xi } ( h _ { i } ^ { \prime \prime } ) .
|
| 76 |
+
$$
|
| 77 |
+
|
| 78 |
+
The contrastive loss for the $i$ -th object proposal is defined as:
|
| 79 |
+
|
| 80 |
+
$$
|
| 81 |
+
\mathcal { L } _ { i } = - 2 \cdot \frac { \langle v _ { i } , v _ { i } ^ { \prime } \rangle } { \left. v _ { i } \right. _ { 2 } \cdot \left. v _ { i } ^ { \prime } \right. _ { 2 } } - 2 \cdot \frac { \langle v _ { i } , v _ { i } ^ { \prime \prime } \rangle } { \left. v _ { i } \right. _ { 2 } \cdot \left. v _ { i } ^ { \prime \prime } \right. _ { 2 } } .
|
| 82 |
+
$$
|
| 83 |
+
|
| 84 |
+
Then we can formulate the overall loss function for each image as:
|
| 85 |
+
|
| 86 |
+
$$
|
| 87 |
+
\mathcal { L } = \frac { 1 } { K } \sum _ { i = 1 } ^ { K } \mathcal { L } _ { i } ,
|
| 88 |
+
$$
|
| 89 |
+
|
| 90 |
+
where $K$ is the number of object proposals. We symmetrize the loss $\mathcal { L }$ in Eq. 6 by separately feeding $V _ { 1 }$ to the target network and $\{ V _ { 2 } , V _ { 3 } \}$ to the online network to compute $\widetilde { \mathcal { L } }$ . At each training iteration, we perform a stochastic optimization step to minimize $\mathcal { L } ^ { \mathrm { S o C o } } = \mathcal { L } + \widetilde { \mathcal { L } }$ .
|
| 91 |
+
|
| 92 |
+
Scale-Aware Assignment. Mask R-CNN with FPN uses the IoU between anchors and ground-truth boxes to determine positive samples. It defines anchors to have areas of $\{ 3 2 ^ { 2 } , 6 4 ^ { 2 } , \overline { { { 1 2 8 ^ { 2 } } } } , 2 5 6 ^ { 2 } \}$ pixels on $\{ P _ { 2 } , P _ { 3 } , \bar { P } _ { 4 } , P _ { 5 } \}$ , respectively, which means ground-truth boxes of scale within a range are assigned to a specific pyramid level. Inspired by this, we propose a scale-aware assignment strategy, which greatly encourages the pretraining model to learn object-level scaleinvariant representations. Concretely, we assign object proposals of area within a range of $\left\{ 0 - 4 8 ^ { 2 } , 4 \dot { 9 } ^ { 2 } - 9 6 ^ { 2 } , 9 7 ^ { 2 } - 1 9 2 ^ { 2 } , 1 9 \dot { 3 } ^ { 2 } - 2 2 4 ^ { 2 } \right\}$ pixels to $\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \}$ , respectively. Notice that the maximum proposal size is $2 2 4 \times 2 2 4$ since all of the views are resized to a fixed resolution smaller than 224. The advantage is that the same object proposals at different scales are encouraged to learn consistent representations through contrastive learning. As a result, SoCo learns object-level scale-invariant visual representations, which is important for object detection.
|
| 93 |
+
|
| 94 |
+
Introducing Properties of Detection to Pretraining. Here, we discuss how SoCo promotes important properties of object detection in the pretraining. Object detection uses tight bounding boxes to represent objects. To introduce object-level representations, SoCo generates object proposals by selective search. Translation invariance and scale invariance at the object level are regarded as the most important properties for object detection, i.e., feature representations of objects belonging to same category should be insensitive to scale and location. Recall that $V _ { 2 }$ is a randomly cropped patch of $V _ { 1 }$ . Random cropping introduces box shift and thus contrastive learning between $V _ { 1 }$ and $V _ { 2 }$ encourages the pretraining model to learn location-invariant representations. $V _ { 3 }$ is generated by downsampling $V _ { 2 }$ , which results in a scale augmentation of object proposals. With our scale-aware assignment strategy, the contrastive loss between $V _ { 1 }$ and $V _ { 3 }$ guides the pretraining towards learning scale-invariant visual representations.
|
| 95 |
+
|
| 96 |
+
Extending to Other Object Detectors. SoCo can be easily extended to align to other detectors besides Mask R-CNN with FPN. Here, we apply SoCo to Mask R-CNN with a C4 structure, which is a popular non-FPN detection framework. The modification is three-fold: 1) for all object proposals, RoIAlign is performed on $C _ { 4 }$ ; 2) the R-CNN head is replaced by the entire 5-th residual block; 3) view $V _ { 3 }$ is discarded and the remaining $V _ { 1 }$ and $V _ { 2 }$ are kept for object-level contrastive learning. Experiments and comparisons with state-of-the-art methods in Section 4.3 demonstrate the extensibility and flexibility of SoCo.
|
| 97 |
+
|
| 98 |
+
# 4 Experiments
|
| 99 |
+
|
| 100 |
+
# 4.1 Pretraining Settings
|
| 101 |
+
|
| 102 |
+
Architecture. Through the introduction of object proposals, the architectural discrepancy is reduced between pretraining and downstream detection finetuning. Mask R-CNN [17] is a commonly adopted framework to evaluate transfer performance. To demonstrate the extensibility and flexibility of SoCo, we provide details of SoCo alignment for the detection architectures R50-FPN and R50- C4. SoCo-R50-FPN: ResNet-50 [36] with FPN [15] is used as the image-level feature encoder. RoIAlign [17] is then used to extract RoI features on feature maps $\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \}$ with a stride of $\{ 4 , 8 , 1 6 , 3 2 \}$ . According to the image areas of object proposals, each RoI feature is then transformed to an object-level representation by the head network as in Mask R-CNN. SoCo-R50-C4: on the standard ResNet-50 architecture, we insert the RoI operation on the output of the 4-th residual block. The entire 5-th residual block is treated as the head network to encode object-level features. Both the projection network and prediction network are 2-layer MLPs which consist of a linear layer with output size 4096 followed by batch normalization [37], rectified linear units (ReLU) [38], and a final linear layer with output dimension 256.
|
| 103 |
+
|
| 104 |
+
Dataset. We adopt the widely used ImageNet [1] which consists of ${ \sim } 1 . 2 8$ million images for self-supervised pretraining.
|
| 105 |
+
|
| 106 |
+
Data Augmentation. Once all views are constructed, we employ the data augmentation pipeline of BYOL [3]. Specifically, we apply random horizontal flip, color distortion, Gaussian blur, grayscaling, and the solarization operation. We remove the random crop augmentation since spatial transformations have already been applied on all views.
|
| 107 |
+
|
| 108 |
+
Optimization. We use a 100-epoch training schedule in all the ablation studies and report the results of 100-epochs and 400-epochs in the comparisons with state-of-the-art methods. We use the LARS optimizer [39] with a cosine decay learning rate schedule [40] and a warm-up period of 10 epochs.
|
| 109 |
+
|
| 110 |
+
Table 1: Comparison with state-of-the-art methods on COCO by using Mask R-CNN with R50-FPN.
|
| 111 |
+
|
| 112 |
+
<table><tr><td rowspan="2">Methods</td><td rowspan="2">Epoch</td><td colspan="5">1× Schedule</td><td rowspan="2"></td><td colspan="6">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>AP</td><td>Apbb</td><td>AP AP</td><td>Apmk</td><td>AP</td><td>AP</td></tr><tr><td>Scratch</td><td>-</td><td>31.0</td><td>49.5</td><td>33.2</td><td>28.5</td><td>46.8</td><td>30.4</td><td>38.4</td><td>57.5</td><td>42.0</td><td>34.7</td><td>54.8</td><td>37.2</td></tr><tr><td>Supervised</td><td>90</td><td>38.9</td><td>59.6</td><td>42.7</td><td>35.4</td><td>56.5</td><td>38.1</td><td>41.3</td><td>61.3</td><td>45.0</td><td>37.3</td><td>58.3</td><td>40.3</td></tr><tr><td>MoCo[4]</td><td>200</td><td>38.5</td><td>58.9</td><td>42.0</td><td>35.1</td><td>55.9</td><td>37.7</td><td>40.8</td><td>61.6</td><td>44.7</td><td>36.9</td><td>58.4</td><td>39.7</td></tr><tr><td>MoCo v2[5]</td><td>200</td><td>40.4</td><td>60.2</td><td>44.2</td><td>36.4</td><td>57.2</td><td>38.9</td><td>41.7</td><td>61.6</td><td>45.6</td><td>37.6</td><td>58.7</td><td>40.5</td></tr><tr><td>InfoMin [6]</td><td>200</td><td>40.6</td><td>60.6</td><td>44.6</td><td>36.7</td><td>57.7</td><td>39.4</td><td>42.5</td><td>62.7</td><td>46.8</td><td>38.4</td><td>59.7</td><td>41.4</td></tr><tr><td>BYOL[3]</td><td>300</td><td>40.4</td><td>61.6</td><td>44.1</td><td>37.2</td><td>58.8</td><td>39.8</td><td>42.3</td><td>62.6</td><td>46.2</td><td>38.3</td><td>59.6</td><td>41.1</td></tr><tr><td>SwAV[7]</td><td>400</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>42.3</td><td>62.8</td><td>46.3</td><td>38.2</td><td>60.0</td><td>41.0</td></tr><tr><td>ReSim-FPNT [45]</td><td>200</td><td>39.8</td><td>60.2</td><td>43.5</td><td>36.0</td><td>57.1</td><td>38.6</td><td>41.4</td><td>61.9</td><td>45.4</td><td>37.5</td><td>59.1</td><td>40.3</td></tr><tr><td>PixPro[10]</td><td>400</td><td>41.4</td><td>61.6</td><td>45.4</td><td>1</td><td>-</td><td>1</td><td>-</td><td>-</td><td>-</td><td>-</td><td>1</td><td>-</td></tr><tr><td>InsLoc [12]</td><td>400</td><td>42.0</td><td>62.3</td><td>45.8</td><td>37.6</td><td>59.0</td><td>40.5</td><td>43.3</td><td>63.6</td><td>47.3</td><td>38.8</td><td>60.9</td><td>41.7</td></tr><tr><td>DenseCL[11]</td><td>200</td><td>40.3</td><td>59.9</td><td>44.3</td><td>36.4</td><td>57.0</td><td>39.2</td><td>41.2</td><td>61.9</td><td>45.1</td><td>37.3</td><td>58.9</td><td>40.1</td></tr><tr><td>DetCons [13]</td><td>1000</td><td>41.8</td><td></td><td>1</td><td>37.4</td><td>-</td><td>：</td><td>42.9</td><td>1</td><td>-</td><td>38.1</td><td></td><td>：</td></tr><tr><td>DetConB [13]</td><td>1000</td><td>42.7</td><td></td><td>-</td><td>38.2</td><td>1</td><td>：</td><td>43.4</td><td>1</td><td>-</td><td>38.7</td><td>-</td><td>：</td></tr><tr><td>SoCo</td><td>100</td><td>42.3</td><td>62.5</td><td>46.5</td><td>37.6</td><td>59.1</td><td>40.5</td><td>43.2</td><td>63.3</td><td>47.3</td><td>38.8</td><td>60.6</td><td>41.9</td></tr><tr><td>SoCo</td><td>400</td><td>43.0</td><td>63.3</td><td>47.1</td><td>38.2</td><td>60.2</td><td>41.0</td><td>44.0</td><td>64.0</td><td>48.4</td><td>39.0</td><td>61.3</td><td>41.7</td></tr><tr><td>SoCo*</td><td>400</td><td>43.2</td><td>63.5</td><td>47.4</td><td>38.4</td><td>60.2</td><td>41.4</td><td>44.3</td><td>64.6</td><td>48.9</td><td>39.6</td><td>61.8</td><td>42.5</td></tr></table>
|
| 113 |
+
|
| 114 |
+
The base learning rate $l r _ { \mathrm { b a s e } }$ is set to 1.0 and is scaled linearly [41] with the batch size $\mathit { l r } = { l r } _ { \mathrm { b a s e } } \times$ BatchSize/256). The weight decay is set to $1 . 0 \times \mathrm { e } ^ { - 5 }$ . The total batch size is set to 2048 over 16 Nvidia V100 GPUs. For the update of the target network, following [3], the momentum coefficient $\tau$ starts from 0.99 and is increased to 1 during training. Synchronized batch normalization is enabled.
|
| 115 |
+
|
| 116 |
+
# 4.2 Transfer Learning Settings
|
| 117 |
+
|
| 118 |
+
COCO [42] and Pascal VOC [43] datasets are used for transfer learning. Detectron2 [44] is used as the code base.
|
| 119 |
+
|
| 120 |
+
COCO Object Detection and Instance Segmentation. We use the COCO train2017 set which contains ${ \sim } 1 1 8 \mathrm { k }$ images with bounding box and instance segmentation annotations in 80 object categories. Transfer performance is evaluated on the COCO val2017 set. We adopt the Mask R-CNN detector [17] withfor object detection, and 5, andand 50-C4 backbones. We reportfor instance segmentation. $\mathsf { A P } ^ { \mathrm { b b } }$ ,e $\mathrm { A \hat { P } _ { 5 0 } ^ { b b } }$ andhe r $\mathsf { A P } _ { 7 5 } ^ { \mathsf { b b } }$ $\mathbf { A P } ^ { \mathrm { m k } }$ $\mathbf { A P } _ { 5 0 } ^ { \mathrm { m k } }$ $\mathbf { A P } _ { 7 5 } ^ { \mathrm { m k } }$
|
| 121 |
+
under the COCO $1 \times$ and $2 \times$ schedules in comparisons with state-of-the-art methods. All ablation studies are conducted on the COCO $1 \times$ schedule.
|
| 122 |
+
|
| 123 |
+
Pascal VOC Object Detection. We use the Pascal VOC trainval $0 7 + 1 2$ set which contains ${ \sim } 1 6 . 5 \mathrm { k }$ images with bounding box annotations in 20 object categories as the training set. Transfer performance is evaluated on the Pascal VOC test2007 set. We report the results of $\mathsf { A P } ^ { \mathrm { b b } }$ , $\mathsf { A P } _ { 5 0 } ^ { \mathrm { b b } }$ and $\mathsf { A P } _ { 7 5 } ^ { \mathrm { b b } }$ for object detection.
|
| 124 |
+
|
| 125 |
+
Optimization. All the pretrained weights except for projection and prediction are loaded into the object detection network for transfer. Following [5], synchronized batch normalization is used across all layers including the newly initialized batch normalization layers in finetuning. For both the COCO and Pascal VOC datasets, we finetune with stochastic gradient descent and a batch size of 16 split across 8 GPUs. For COCO transfer, we use a weight decay of $2 . 5 \times \mathrm { e } ^ { - 5 }$ , and a base learning rate of 0.02 that increases linearly for the first 1000 iterations and drops twice by a factor of 10, after $\frac { 2 } { 3 }$ and $\frac { 8 } { 9 }$ of the total training time. For Pascal VOC transfer, the weight decay is $1 . 0 \times \mathrm { e } ^ { - 4 }$ , and the base learning rate is set to 0.02 and divided by 10 at $\textstyle { \frac { 3 } { 4 } }$ and $\textstyle { \frac { 1 1 } { 1 2 } }$ of the total training time.
|
| 126 |
+
|
| 127 |
+
# 4.3 Comparison with State-of-the-Art Methods
|
| 128 |
+
|
| 129 |
+
Mask R-CNN with R50-FPN on COCO. Table 1 shows the transfer results for Mask R-CNN with R50-FPN backbone. We compare SoCo with the state-of-the-art unsupervised pretraining methods on the COCO $1 \times$ and $2 \times$ schedules. We report our results under 100 epochs and 400 epochs of pretraining. On the 100-epoch training schedule, SoCo achieves $4 2 . 3 \mathrm { \ A p ^ { b b } / 3 7 . 6 \ A P ^ { m k } }$ and $4 3 . 2 \ : \dot { \mathrm { A P ^ { b b } } } / \ : 3 8 . { \bar { 8 } } \ : \mathrm { A P ^ { m k } }$ on the $1 \times$ and $2 \times$ schedules, respectively. Pretrained for 400 epochs, SoCo outperforms all previous pretraining methods designed for either image classification or object detection, achieving $4 \dot { 3 } . 0 \ : \mathrm { A P ^ { b b } } \ : \dot { / } 3 8 . 2 \ : \mathrm { A P ^ { m i } }$ for the $1 \times$ schedule and $4 4 . 0 \mathrm { A P ^ { \bar { b } b } / 3 9 . 0 \mathrm { A P ^ { m k } } }$ for the $2 \times$ schedule. Moreover, we propose an enhanced version named $S o C o ^ { * }$ , which constructs an additional view $V _ { 4 }$ for pretraining. Similar to the construction step of $V _ { 2 }$ , $V _ { 4 }$ is a randomly cropped patch of $V _ { 1 }$ . We resize $V _ { 4 }$ to $1 9 2 \times 1 9 2$ and perform the same forward process as for $V _ { 2 }$ and $V _ { 3 }$ . Compared with SoCo, $\mathrm { S o C o ^ { * } }$ further boosts the performance and obtains an improvement of $+ 0 . 2 \mathrm { \ A P ^ { b b } \ } \bar { / } + 0 . 2$ $\mathbf { A P } ^ { \mathrm { m k } }$ on the $1 \times$ schedule, achieving $\dot { 4 } 3 . 2 \ : \mathrm { A P ^ { b b } } / 3 8 . 4 \ : \mathrm { A P ^ { m k } }$ , and $+ 0 . { \dot { 3 } } \operatorname { A P } ^ { \mathrm { b b } } / + 0 . 6 \operatorname { A P } ^ { \mathrm { m k } }$ on the $2 \times$ schedule, achieving $4 4 . 3 \ : \mathrm { A P ^ { b b } } / 3 9 . \mathsf { \breve { 6 } A P ^ { m k } }$ , respectively.
|
| 130 |
+
|
| 131 |
+
Table 2: Comparison with state-of-the-art methods on COCO using Mask R-CNN with R50-C4.
|
| 132 |
+
|
| 133 |
+
<table><tr><td rowspan="2">Methods</td><td rowspan="2">Epoch</td><td colspan="6">1× Schedule</td><td colspan="6">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td><td>Apbb</td><td>AP8</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td></tr><tr><td>Scratch</td><td>-</td><td>26.4</td><td>44.0</td><td>27.8</td><td>29.3</td><td>46.9</td><td>30.8</td><td>35.6</td><td>54.6</td><td>38.2</td><td>31.4</td><td>51.5</td><td>33.5</td></tr><tr><td>Supervised</td><td>90</td><td>38.2</td><td>58.2</td><td>41.2</td><td>33.3</td><td>54.7</td><td>35.2</td><td>40.0</td><td>59.9</td><td>43.1</td><td>34.7</td><td>56.5</td><td>36.9</td></tr><tr><td>MoCo [4]</td><td>200</td><td>38.5</td><td>58.3</td><td>41.6</td><td>33.6</td><td>54.8</td><td>35.6</td><td>40.7</td><td>60.5</td><td>44.1</td><td>35.4</td><td>57.3</td><td>37.6 37.1</td></tr><tr><td>SimCLR [9] MoCo v2 [5]</td><td>200 800</td><td>- 39.3</td><td>-</td><td>■</td><td>-</td><td>-</td><td>1</td><td>39.6 41.2</td><td>59.1 60.9</td><td>42.9 44.6</td><td>34.6 35.8</td><td>55.9 57.7</td><td>38.2</td></tr><tr><td>InfoMin 6]</td><td></td><td></td><td>58.9</td><td>42.5</td><td>34.3</td><td>55.7</td><td>36.5</td><td></td><td>61.2</td><td>45.0</td><td></td><td>57.9</td><td>38.3</td></tr><tr><td>BYOL[3]</td><td>200 300</td><td>39.0</td><td>58.5</td><td>42.0</td><td>34.1</td><td>55.2</td><td>36.3</td><td>41.3</td><td></td><td></td><td>36.0</td><td>56.8</td><td>37.3</td></tr><tr><td>SwAV[7]</td><td>400</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>40.3</td><td>60.5</td><td>43.9</td><td>35.1</td><td></td><td>36.6</td></tr><tr><td>SimSiam [8]</td><td>200</td><td>- 39.2</td><td>- 59.3</td><td>-</td><td>1</td><td>1</td><td>-</td><td>39.6</td><td>60.1</td><td>42.9</td><td>34.7</td><td>56.6</td><td></td></tr><tr><td>PixPro[10]</td><td>400</td><td>40.5</td><td>59.8</td><td>42.1</td><td>34.4</td><td>56.0</td><td>36.7</td><td>-</td><td>：</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>InsLoc [12]</td><td>400</td><td>39.8</td><td>59.6</td><td>44.0 42.9</td><td>-</td><td>-</td><td>-</td><td>- 41.8</td><td>- 61.6</td><td>-</td><td>-</td><td>- 58.2</td><td>- 38.8</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td>34.7</td><td>56.3</td><td>36.9</td><td></td><td></td><td>45.4</td><td>36.3</td><td></td><td></td></tr><tr><td>SoCo</td><td>100</td><td>40.4</td><td>60.4</td><td>43.7</td><td>34.9</td><td>56.8</td><td>37.0</td><td>41.1</td><td>61.0</td><td>44.4</td><td>35.6</td><td>57.5</td><td>38.0</td></tr><tr><td>SoCo</td><td>400</td><td>40.9</td><td>60.9</td><td>44.3</td><td>35.3</td><td>57.5</td><td>37.3</td><td>42.0</td><td>61.8</td><td>45.6</td><td>36.3</td><td>58.5</td><td>38.8</td></tr></table>
|
| 134 |
+
|
| 135 |
+
Table 3: Comparison with state-of-the-art methods on Pascal VOC. Faster R-CNN with R50-C4 is adopted. Table is split to two sub-tables.
|
| 136 |
+
|
| 137 |
+
<table><tr><td colspan="5">(a) Sub-table 1.</td></tr><tr><td>Methods</td><td>Epoch</td><td>APbb</td><td>AP</td><td>AP</td></tr><tr><td>Scratch Supervised</td><td>- 90</td><td>33.8 53.5</td><td>60.2 81.3</td><td>33.1 58.8</td></tr><tr><td>ReSim-C4 [45] PixPro [10] InsLoc [12] DenseCL[11]</td><td>200 400 400 200</td><td>58.7 60.2 58.4 58.7</td><td>83.1 83.8 83.0 82.8</td><td>66.3 67.7 65.3 65.2</td></tr><tr><td>SoCo SoCo</td><td>100 400</td><td>59.1 59.7</td><td>83.4 83.8</td><td>65.6 66.8</td></tr></table>
|
| 138 |
+
|
| 139 |
+
(b) Sub-table 2.
|
| 140 |
+
|
| 141 |
+
<table><tr><td>Methods</td><td>Epoch</td><td>Apbb</td><td>AP</td><td>AP</td></tr><tr><td>MoCo[4]</td><td>200</td><td>55.9</td><td>81.5</td><td>62.6</td></tr><tr><td>SimCLR [9]</td><td>1000</td><td>56.3</td><td>81.9</td><td>62.5</td></tr><tr><td>MoCo v2[5]</td><td>800</td><td>57.6</td><td>82.7</td><td>64.4</td></tr><tr><td>InfoMin [6]</td><td>200</td><td>57.6</td><td>82.7</td><td>64.6</td></tr><tr><td>BYOL[3]</td><td>300</td><td>51.9</td><td>81.0</td><td>56.5</td></tr><tr><td>SwAV[7]</td><td>400</td><td>45.1</td><td>77.4</td><td>46.5</td></tr><tr><td>SimSiam[8]</td><td>200</td><td>57.0</td><td>82.4</td><td>63.7</td></tr><tr><td>ReSim-FPN[45]</td><td>200</td><td>59.2</td><td>82.9</td><td>65.9</td></tr></table>
|
| 142 |
+
|
| 143 |
+
Mask R-CNN with R50-C4 on COCO. To demonstrate the extensibility and flexibility of SoCo, we also evaluate transfer performance using Mask R-CNN with R50-C4 backbone on the COCO benchmark. We report the results of SoCo under 100 epochs and 400 epochs. Table 2 compares the proposed method to previous state-of-the-art methods. Without bells and whistles, SoCo obtains state-of-the-art performance, achieving $4 0 . 9 ~ \mathrm { A P ^ { b b } } / 3 5 . 3 ~ \mathrm { A P ^ { m k } }$ and $4 2 . 0 ~ \mathrm { A P ^ { b b } } / 3 6 . 3 ~ \mathrm { A P ^ { m k } }$ on the COCO $1 \times$ and $2 \times$ schedules, respectively.
|
| 144 |
+
|
| 145 |
+
Faster R-CNN with R50-C4 on Pascal VOC. We also evaluate the transfer ability of SoCo on the Pascal VOC benchmark. Following [5], we adopt Faster R-CNN with R50-C4 backbone for transfer learning. Table 3 shows the comparison. SoCo obtains an improvement of $+ 6 . 2 \ \mathrm { A P } ^ { \mathrm { b b } }$ against the supervised pretraining baseline, achieving $5 9 . 7 \mathrm { A P } ^ { \mathrm { b b } }$ for VOC object detection.
|
| 146 |
+
|
| 147 |
+
# 4.4 Ablation Study
|
| 148 |
+
|
| 149 |
+
To further understand the advantages of SoCo, we conduct a series of ablation studies that examine the effectiveness of object-level contrastive learning, the effects of alignment between pretraining and detection, and different hyper-parameters. For all ablation studies, we use Mask R-CNN with R50-FPN backbone for transfer learning and adopt a 100-epoch SoCo pretraining schedule. Transfer performance is evaluated under the COCO $1 \times$ schedule. For the ablation of each hyper-parameter or component, we fix all other hyper-parameters to the following default settings: view $V _ { 3 }$ with $1 1 2 \times 1 1 2$ resolution, batch size of 2048, momentum coefficient $\tau = 0 . 9 9$ , proposal number $K = 4$ box jitter and scale-aware assignment are used, FPN and R-CNN head are pretrained and transferred, and selective search is used as the proposal generator.
|
| 150 |
+
|
| 151 |
+
Table 4: Ablation study on the effectiveness of aligning pretraining to object detection.
|
| 152 |
+
|
| 153 |
+
<table><tr><td>Whole Image</td><td>Selective Search</td><td>FPN</td><td>Head</td><td>Scale-aware Assignment</td><td>Box Jitter</td><td>Multi View</td><td>Apbb</td><td>Apmk</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>38.1</td><td>34.4</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>40.6 (+2.5)</td><td>36.8 (+2.4)</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>40.2 (+2.1)</td><td>36.2 (+1.8)</td></tr><tr><td>√</td><td></td><td></td><td>广</td><td></td><td></td><td></td><td>41.2 (+3.1)</td><td>37.0 (+2.6)</td></tr><tr><td>√</td><td></td><td></td><td></td><td>1</td><td></td><td></td><td>41.6 (+3.5)</td><td>37.3 (+2.9)</td></tr><tr><td>√</td><td>>>>>>></td><td>>>>>></td><td>1</td><td></td><td></td><td>4</td><td>41.7 (+3.6)</td><td>37.5 (+3.1)</td></tr><tr><td>√</td><td></td><td></td><td>厂</td><td>厂</td><td></td><td></td><td>42.3 (+4.2)</td><td>37.6 (+3.2)</td></tr></table>
|
| 154 |
+
|
| 155 |
+
Table 5: Ablation studies on hyper-parameters for the proposed SoCo method. y on image size of view $V _ { 3 }$ . (c) Study on proposal generation and proposal number $K$
|
| 156 |
+
|
| 157 |
+
<table><tr><td>Image Size</td><td>Apbb</td><td>Apmk</td></tr><tr><td>96</td><td>42.1</td><td>37.7</td></tr><tr><td>112</td><td>42.3</td><td>37.6</td></tr><tr><td>128</td><td>42.1</td><td>37.7</td></tr><tr><td>160</td><td>42.0</td><td>37.6</td></tr><tr><td>192</td><td>42.2</td><td>37.8</td></tr></table>
|
| 158 |
+
|
| 159 |
+
(b) Study on batch size.
|
| 160 |
+
|
| 161 |
+
<table><tr><td>Batch Size</td><td>APbb</td><td>Apmk</td></tr><tr><td>512</td><td>41.7</td><td>37.6</td></tr><tr><td>1024</td><td>41.9</td><td>37.6</td></tr><tr><td>2048</td><td>42.3</td><td>37.6</td></tr><tr><td>4096</td><td>41.4</td><td>37.3</td></tr></table>
|
| 162 |
+
|
| 163 |
+
(d) Study on momentum coefficient $\tau$
|
| 164 |
+
|
| 165 |
+
<table><tr><td>Selective Search</td><td>Random</td><td>K</td><td>APbb</td><td>Apmk</td></tr><tr><td><</td><td></td><td>1</td><td>41.6</td><td>37.3</td></tr><tr><td></td><td></td><td>486</td><td>42.3 41.6</td><td>37.6</td></tr><tr><td>√</td><td></td><td></td><td></td><td>37.4</td></tr><tr><td>√</td><td></td><td></td><td>41.2</td><td>37.0</td></tr><tr><td></td><td>>>></td><td>148</td><td>41.4</td><td>36.9</td></tr><tr><td></td><td></td><td></td><td>NaN NaN</td><td>NaN NaN</td></tr></table>
|
| 166 |
+
|
| 167 |
+
<table><tr><td>T</td><td>APbb</td><td>Apmk</td></tr><tr><td>0.98</td><td>35.0</td><td>31.7</td></tr><tr><td>0.99</td><td>42.3</td><td>37.6</td></tr><tr><td>0.993</td><td>41.8</td><td>37.6</td></tr></table>
|
| 168 |
+
|
| 169 |
+
Effectiveness of Aligning Pretraining to Object Detection. We ablate each component of SoCo step by step to demonstrate the the effectiveness of aligning pretraining to object detection. Table 4 reports the studies. The baseline treats the whole image as an instance, without considering any detection properties or architecture alignment. It obtains $3 8 . 1 \mathrm { \ A P ^ { b b } }$ and $3 4 . 4 \mathrm { A P ^ { m k } }$ . Selective search introduces object-level representations. By leveraging generated object proposals and conducting contrastive learning at the object level, our method obtains an improvement of $+ 2 . 5 \mathrm { \ A P ^ { b b } }$ and $+ 2 . 4$ $\mathbf { A P } ^ { \mathrm { m k } }$ . Next, we further minimize the architectural discrepancy between pretraining and object detection pipeline by introducing FPN and an R-CNN head into the pretraining architecture. We observe that only introducing FPN into pretraining slightly hurt the performance. In contrast, including both FPN and the R-CNN head improves the transfer performance to $4 1 . 2 ~ \mathrm { A P ^ { b b } } ~ / ~ 3 7 . 0$ $\mathbf { A P } ^ { \mathrm { m k } }$ , which verifies the effectiveness of architectural alignment between self-supervised pretraining and downstream tasks. On top of architectural alignment, we further leverage scale-aware assignment which encourages scale-invariant object-level visual representations, and an improvement of $+ 3 . 5$ $\mathsf { A P } ^ { \mathrm { b b } }$ / $+ 2 . 9 \mathrm { \ A P ^ { \mathrm { \bar { m k } } } }$ against the baseline is found. The box jitter strategy slightly elevates performance. Finally, multiple views $( V _ { 3 } )$ further directs SoCo towards learning scale-invariant and translationinvariant representations, and our method achieves $4 2 . 3 \mathrm { \ A P ^ { b b } }$ and $\mathsf { \bar { 3 } 7 . 6 \mathbf { A P } ^ { m k } }$ .
|
| 170 |
+
|
| 171 |
+
Ablation Study on Hyper-Parameters. Table 5 examines sensitivity to the hyper-parameters of SoCo.
|
| 172 |
+
|
| 173 |
+
Table 5(a) ablates the resolution of view $V _ { 3 }$ . SoCo is found to be insensitive to the image size of $V _ { 3 }$ We use $1 1 2 \times 1 1 2$ as the default resolution due to its slightly better transfer performance and low training computation cost.
|
| 174 |
+
|
| 175 |
+
Table 6: Transfer Learning on LVIS dataset using Mask R-CNN with R50-FPN.
|
| 176 |
+
|
| 177 |
+
<table><tr><td rowspan="2">Method</td><td rowspan="2">Epoch</td><td colspan="5">1× Schedule</td><td rowspan="2"></td><td colspan="6">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>APmk</td><td>AP</td><td>AP7</td><td>Apbb AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>AP</td></tr><tr><td>Supervised</td><td>90</td><td>20.4</td><td>32.9</td><td>21.7</td><td>19.4</td><td>30.6</td><td>20.5</td><td>23.4</td><td>36.9</td><td>24.9</td><td>22.3</td><td>34.7</td><td>23.5</td></tr><tr><td>SoCo*</td><td>400</td><td>26.3</td><td>41.2</td><td>27.8</td><td>25.0</td><td>38.5</td><td>26.8</td><td>28.3</td><td>43.5</td><td>30.7</td><td>26.9</td><td>41.1</td><td>28.7</td></tr></table>
|
| 178 |
+
|
| 179 |
+
Table 7: Transfer learning on RetinaNet and FCOS.
|
| 180 |
+
|
| 181 |
+
<table><tr><td>Method</td><td>Epoch</td><td>Apbb</td><td>AP8</td><td>AP</td></tr><tr><td>Supervised</td><td>90</td><td>36.3</td><td>55.3</td><td>38.6</td></tr><tr><td>SoCo* 一</td><td>400</td><td>38.3</td><td>57.2</td><td>41.2</td></tr></table>
|
| 182 |
+
|
| 183 |
+
(b) Transfer to FCOS.
|
| 184 |
+
|
| 185 |
+
<table><tr><td>Method</td><td>Epoch</td><td>Apbb</td><td>AP8</td><td>AP9</td></tr><tr><td>Supervised</td><td>90</td><td>36.6</td><td>56.0</td><td>38.8</td></tr><tr><td>SoCo*</td><td>400</td><td>37.4</td><td>56.3</td><td>39.9</td></tr></table>
|
| 186 |
+
|
| 187 |
+
Table 5(b) ablates the training batch size of SoCo. It can be seen that larger or smaller batch sizes hurt the performance. We use a batch size of 2048 by default.
|
| 188 |
+
|
| 189 |
+
Table 5(c) ablates the effectiveness of object proposal generation strategies. To demonstrate the superiority of meaningful proposals generated by selective search, we randomly generate $K$ bounding boxes in each of the training images to replace the selective search proposals (namely Random in Table 5(c)). SoCo can still yield satisfactory results using a single random bounding box to learn object-level representations, which is not surprising, since the RoIAlign operation performed on the randomly generated bounding boxes also introduces object-level representation learning. This fact indirectly demonstrates that downstream object detectors can benefit from self-supervised pretraining that involves object-level representations. However, the result is still slightly worse than the case where only one object proposal generated by selective search is used in the pretraining. For the cases where we randomly generate 4 and 8 bounding boxes, the noisy and meaningless proposals cause the pretraining to diverge. From the table, we also find that applying more object proposals $K = 8$ and $K = 1 6$ ) generated by selective search can be harmful. One potential reason is that most of the images in the ImageNet dataset only contain a few objects, so a larger $K$ may create duplicated and redundant proposals.
|
| 190 |
+
|
| 191 |
+
Table 5(d) ablates the momentum coefficient $\tau$ of the exponential moving average, and $\tau = 0 . 9 9$ yields the best performance.
|
| 192 |
+
|
| 193 |
+
# 4.5 More Experiments
|
| 194 |
+
|
| 195 |
+
Transfer Learning on LVIS Dataset. In addition to COCO and PASCAL VOC object detection, we also consider the challenging LVIS v1 dataset [46] to demonstrate the effectiveness and generality of our approach. The LVIS v1 detection benchmark contains 1203 categories in a long-tailed distribution with few training samples. It is considered to be more challenging than the COCO benchmark. We use Mask R-CNN with R50-FPN backbone and follow the standard LVIS $1 \times$ and $2 \times$ training schedules. We do not use any training or post-processing tricks. Table 6 shows the results. We can see that our method achieves $+ 5 . 9 \mathrm { \ A \bar { P } ^ { b b } / + \bar { S } . 6 \mathrm { \ A P ^ { m k } } }$ and $+ 4 . 9 \mathrm { \ A P ^ { b b } / + 4 . 6 \ A P ^ { m k } }$ improvements under the LVIS $1 \times$ and $2 \times$ training schedules, respectively.
|
| 196 |
+
|
| 197 |
+
Transfer Learning on Different Object Detectors. In addition to two-stage detectors, we further conduct experiments on RetinaNet [47] and FCOS [48], which are representative works for singlestage detectors and anchor-free detectors. We transfer the weights of backbone and FPN layers learned by $\mathrm { S o C o ^ { * } }$ to RetinaNet and FCOS architectures and follow the standard COCO $1 \times$ schedule. We use MMDetection [49] as code base and default optimization configs are adopted. Table 7 shows the comparison.
|
| 198 |
+
|
| 199 |
+
Evaluation on Mini COCO. Transfer to the full COCO dataset may be of limited significance due to the extensive supervision available from its large-scale annotated training data. To further demonstrate the generalization ability of SoCo, we conduct an experiment on a mini version of the COCO dataset, named Mini COCO. Concretely, we randomly select $5 \%$ or $10 \%$ of the training data from COCO train2017 to form Mini COCO benchmarks. Mask-RCNN with R50-FPN backbone is adopted to evaluate the transfer performance on the COCO $1 \times$ schedule. COCO val2017 is still used as the evaluation set. The other settings remain unchanged. Table 8 summarizes the results. Compared with the supervised pretraining baseline which achieves $1 9 . 4 \ : \mathrm { A P ^ { b b } } / \ : 1 8 . 3 \ : \mathrm { A P ^ { m k } }$ and $2 4 . 7 \mathrm { A P } ^ { \mathrm { b b } }$ / 22.9 $\mathbf { A P } ^ { \mathrm { m k } }$ on the $5 \%$ and $10 \%$ Mini COCO benchmarks, SoCo obtains large improvements of $+ 6 . 6$ $\mathsf { A P } ^ { \mathrm { b b } }$ $^ { \mathrm { b } } / { + 4 . 6 } \mathrm { A P ^ { \mathrm { m k } } }$ and $+ 5 . 7$ $\mathsf { A P } ^ { \mathrm { b b } }$ / $+ 3 . 9 \mathrm { A P } ^ { \mathrm { m k } }$ , respectively. $\mathrm { S o C o ^ { * } }$ further boosts performance with improvements of $+ 7 . 4 \mathrm { \ A P ^ { b b } }$ $' + 5 . 5 \ : \mathrm { A P } ^ { \mathrm { m k } }$ and $+ 6 . 4 \mathrm { \ A P ^ { b b } }$ $\mathrm { ^ { \prime } { + } 4 . 7 \ A P ^ { m k } }$ over the supervised pretraining baseline.
|
| 200 |
+
|
| 201 |
+
Table 8: Results on Mini COCO $\mathbf { 1 } \times$ schedule. Mask R-CNN with R50-FPN backbone is adopted.
|
| 202 |
+
|
| 203 |
+
<table><tr><td rowspan="2">Methods</td><td rowspan="2">Epoch</td><td colspan="5">Mini COCO (5%)</td><td rowspan="2"></td><td colspan="5">Mini C0CO (10%)</td></tr><tr><td>Apbb AP8</td><td>AP9</td><td>APmk</td><td>APk</td><td>APm</td><td>Apbb</td><td>AP8</td><td>AP</td><td>Apmk</td><td>AP APP</td></tr><tr><td>Supervised</td><td>90</td><td>19.4</td><td>36.6</td><td>18.6</td><td>18.3</td><td>33.5</td><td>17.9</td><td>24.7</td><td>43.1</td><td>25.3</td><td>22.9</td><td>40.0</td><td>23.4</td></tr><tr><td>SoCo</td><td>100</td><td>24.6</td><td>42.2</td><td>25.6</td><td>22.1</td><td>38.8</td><td>22.4</td><td>29.2</td><td>47.7</td><td>31.1</td><td>26.1</td><td>44.4</td><td>27.0</td></tr><tr><td>SoCo</td><td>400</td><td>26.0</td><td>43.2</td><td>27.5</td><td>22.9</td><td>40.0</td><td>23.4</td><td>30.4</td><td>48.6</td><td>32.5</td><td>26.8</td><td>45.2</td><td>27.9</td></tr><tr><td>SoCo*</td><td>400</td><td>26.8</td><td>45.0</td><td>28.3</td><td>23.8</td><td>41.4</td><td>24.2</td><td>31.1</td><td>49.9</td><td>33.4</td><td>27.6</td><td>46.4</td><td>28.9</td></tr></table>
|
| 204 |
+
|
| 205 |
+
Table 9: Pretraining on non-object-centric dataset.
|
| 206 |
+
|
| 207 |
+
<table><tr><td>Pretraining dataset</td><td>Epoch</td><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td></tr><tr><td>ImageNet</td><td>100</td><td>42.3</td><td>62.5</td><td>46.5</td><td>37.6</td><td>59.1</td><td>40.5</td></tr><tr><td>ImageNet-subset</td><td>100</td><td>36.9</td><td>56.2</td><td>40.1</td><td>33.2</td><td>53.0</td><td>35.6</td></tr><tr><td>COCO train set + unlabeled set</td><td>100</td><td>37.3</td><td>56.5</td><td>40.6</td><td>33.5</td><td>53.4</td><td>36.0</td></tr><tr><td>COCO train set + unlabeled set</td><td>530</td><td>40.6</td><td>61.1</td><td>44.4</td><td>36.4</td><td>58.1</td><td>38.7</td></tr></table>
|
| 208 |
+
|
| 209 |
+
Pretraining on Non-object-centric Dataset. We use COCO training set and unlabeled set as pretraining data. Since ImageNet is ${ \sim } 5 . 3$ times larger than COCO, we pretrain our method on COCO for 100 epochs and prolonged 530 epochs respectively. Furthermore, we also generate a dataset named ImageNet-Subset which contains the same number of images of COCO to verify the impact of different pretraining datasets. Table 9 shows the results. Compared with the ImageNet pretrained model, the COCO pretrained model under 530 epochs is lower by 1.7 AP, possibly because the scale of COCO is smaller than ImageNet. Compared with the ImageNet-Subset pretrained model, the COCO pretrained model under 100 epochs is $+ 0 . 4$ AP higher, which demonstrates the importance of domain alignment between the pretraining dataset and detection dataset.
|
| 210 |
+
|
| 211 |
+
Longer Finetuning. We transfer $\mathrm { { \bf S o C o ^ { * } } }$ to Mask R-CNN with R50-FPN under the COCO $4 \times$ schedule. Table 10 shows the results. Our model continues to improve the performance with the $4 \times$ finetuning schedule, surpassing the supervised baseline by $+ 2 . 6 \ : \mathrm { \bar { A P } ^ { b b } } / + 1 . \bar { 9 } \ : \mathrm { A P ^ { m k } }$ .
|
| 212 |
+
|
| 213 |
+
Table 10: Finetuning on COCO $4 \times$ schedule using Mask R-CNN with R50-FPN.
|
| 214 |
+
|
| 215 |
+
<table><tr><td>Method</td><td>Epoch</td><td>APbb</td><td>APb 50</td><td>AP</td><td>Apmk</td><td>AP 50</td><td>APP</td></tr><tr><td>Supervised</td><td>90</td><td>41.9</td><td>61.5</td><td>45.4</td><td>37.7</td><td>58.8</td><td>40.5</td></tr><tr><td>SoCo*</td><td>400</td><td>44.5</td><td>64.2</td><td>48.8</td><td>39.6</td><td>61.5</td><td>42.4</td></tr></table>
|
| 216 |
+
|
| 217 |
+
# 5 Conclusion
|
| 218 |
+
|
| 219 |
+
In this paper, we propose a novel object-level self-supervised pretraining method named Selective Object COntrastive learning (SoCo), which aims to align pretraining to object detection. Different from prior image-level contrastive learning methods which treat the whole image as an instance, SoCo treats each object proposal generated by the selective search algorithm as an independent instance, enabling SoCo to learn object-level visual representations. Further alignment is obtained in two ways. One is through network alignment between pretraining and downstream object detection, such that all layers of the detectors can be well-initialized. The other is by accounting for important properties of object detection such as object-level translation invariance and scale invariance. SoCo achieves state-of-the-art transfer performance on COCO detection using a Mask R-CNN detector. Experiments on two-stage and single-stage detectors demonstrate the generality and extensibility of SoCo.
|
| 220 |
+
|
| 221 |
+
# References
|
| 222 |
+
|
| 223 |
+
[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A largescale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009.
|
| 224 |
+
[2] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018.
|
| 225 |
+
[3] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. [4] 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.
|
| 226 |
+
[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.
|
| 227 |
+
[6] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning. arXiv preprint arXiv:2005.10243, 2020. [7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. arXiv preprint arXiv:2006.09882, 2020.
|
| 228 |
+
[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. arXiv preprint arXiv:2011.10566, 2020.
|
| 229 |
+
[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020.
|
| 230 |
+
[10] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020.
|
| 231 |
+
[11] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020.
|
| 232 |
+
[12] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for selfsupervised detection pretraining. arXiv preprint arXiv:2102.08318, 2021.
|
| 233 |
+
[13] Olivier J Hénaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. arXiv preprint arXiv:2103.10957, 2021.
|
| 234 |
+
[14] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104(2):154–171, 2013.
|
| 235 |
+
[15] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017.
|
| 236 |
+
[16] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497, 2015.
|
| 237 |
+
[17] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017.
|
| 238 |
+
[18] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010.
|
| 239 |
+
[19] Ruslan Salakhutdinov and Geoffrey Hinton. Deep boltzmann machines. In Artificial intelligence and statistics, pages 448–455. PMLR, 2009.
|
| 240 |
+
[20] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6874–6883, 2017.
|
| 241 |
+
[21] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015.
|
| 242 |
+
[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536–2544, 2016.
|
| 243 |
+
[23] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018.
|
| 244 |
+
[24] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182–4192. PMLR, 2020.
|
| 245 |
+
[25] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Selfsupervised learning via redundancy reduction. arXiv preprint arXiv:2103.03230, 2021.
|
| 246 |
+
[26] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021.
|
| 247 |
+
[27] Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Unsupervised pretraining of image features on non-curated data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2959–2968, 2019.
|
| 248 |
+
[28] Priya Goyal, Mathilde Caron, Benjamin Lefaudeux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021.
|
| 249 |
+
[29] Pedro O Pinheiro, Amjad Almahairi, Ryan Y Benmaleck, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. arXiv preprint arXiv:2011.05499, 2020.
|
| 250 |
+
[30] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328–335, 2014.
|
| 251 |
+
[31] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601–1610, 2021.
|
| 252 |
+
[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020.
|
| 253 |
+
[33] Bharat Singh and Larry S. Davis. An analysis of scale invariance in object detection - snip, 2018.
|
| 254 |
+
[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014.
|
| 255 |
+
[35] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440–1448, 2015.
|
| 256 |
+
[36] 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, pages 770–778, 2016.
|
| 257 |
+
[37] 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.
|
| 258 |
+
[38] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010.
|
| 259 |
+
[39] Yang You, Igor Gitman, and Boris Ginsburg. Scaling sgd batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 6:12, 2017.
|
| 260 |
+
[40] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016.
|
| 261 |
+
[41] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017.
|
| 262 |
+
[42] 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. In European conference on computer vision, pages 740–755. Springer, 2014.
|
| 263 |
+
[43] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338, 2010.
|
| 264 |
+
[44] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019.
|
| 265 |
+
[45] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. arXiv preprint arXiv:2103.12902, 2021.
|
| 266 |
+
[46] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356–5364, 2019.
|
| 267 |
+
[47] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017.
|
| 268 |
+
[48] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627–9636, 2019.
|
| 269 |
+
[49] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019.
|
parse/train/8PA2nX9v_r2/8PA2nX9v_r2_content_list.json
ADDED
|
@@ -0,0 +1,1196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"type": "text",
|
| 4 |
+
"text": "Aligning Pretraining for Detection via Object-Level Contrastive Learning ",
|
| 5 |
+
"text_level": 1,
|
| 6 |
+
"bbox": [
|
| 7 |
+
267,
|
| 8 |
+
122,
|
| 9 |
+
727,
|
| 10 |
+
172
|
| 11 |
+
],
|
| 12 |
+
"page_idx": 0
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"type": "text",
|
| 16 |
+
"text": "Fangyun Wei∗ Yue Gao∗ Zhirong Wu Han Hu Stephen Lin ",
|
| 17 |
+
"bbox": [
|
| 18 |
+
205,
|
| 19 |
+
224,
|
| 20 |
+
785,
|
| 21 |
+
241
|
| 22 |
+
],
|
| 23 |
+
"page_idx": 0
|
| 24 |
+
},
|
| 25 |
+
{
|
| 26 |
+
"type": "text",
|
| 27 |
+
"text": "Microsoft Research Asia {fawe, yuegao, wuzhiron, hanhu, stevelin}@microsoft.com ",
|
| 28 |
+
"bbox": [
|
| 29 |
+
264,
|
| 30 |
+
247,
|
| 31 |
+
736,
|
| 32 |
+
275
|
| 33 |
+
],
|
| 34 |
+
"page_idx": 0
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"type": "text",
|
| 38 |
+
"text": "Abstract ",
|
| 39 |
+
"text_level": 1,
|
| 40 |
+
"bbox": [
|
| 41 |
+
462,
|
| 42 |
+
310,
|
| 43 |
+
535,
|
| 44 |
+
327
|
| 45 |
+
],
|
| 46 |
+
"page_idx": 0
|
| 47 |
+
},
|
| 48 |
+
{
|
| 49 |
+
"type": "text",
|
| 50 |
+
"text": "Image-level contrastive representation learning has proven to be highly effective as a generic model for transfer learning. Such generality for transfer learning, however, sacrifices specificity if we are interested in a certain downstream task. We argue that this could be sub-optimal and thus advocate a design principle which encourages alignment between the self-supervised pretext task and the downstream task. In this paper, we follow this principle with a pretraining method specifically designed for the task of object detection. We attain alignment in the following three aspects: 1) object-level representations are introduced via selective search bounding boxes as object proposals; 2) the pretraining network architecture incorporates the same dedicated modules used in the detection pipeline (e.g. FPN); 3) the pretraining is equipped with object detection properties such as object-level translation invariance and scale invariance. Our method, called Selective Object COntrastive learning (SoCo), achieves state-of-the-art results for transfer performance on COCO detection using a Mask R-CNN framework. Code is available at https://github.com/hologerry/SoCo. ",
|
| 51 |
+
"bbox": [
|
| 52 |
+
233,
|
| 53 |
+
342,
|
| 54 |
+
766,
|
| 55 |
+
549
|
| 56 |
+
],
|
| 57 |
+
"page_idx": 0
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"type": "text",
|
| 61 |
+
"text": "1 Introduction ",
|
| 62 |
+
"text_level": 1,
|
| 63 |
+
"bbox": [
|
| 64 |
+
174,
|
| 65 |
+
574,
|
| 66 |
+
310,
|
| 67 |
+
590
|
| 68 |
+
],
|
| 69 |
+
"page_idx": 0
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"type": "text",
|
| 73 |
+
"text": "Pretraining and finetuning has been the dominant paradigm of training deep neural networks in computer vision. Downstream tasks usually leverage pretrained weights learned on large labeled datasets such as ImageNet [1] for initialization. As a result, supervised ImageNet pretraining has been prevalent throughout the field. Recently, self-supervised pretraining [2, 3, 4, 5, 6, 7, 8, 9] has achieved considerable progress and alleviated the dependency on labeled data. These methods aim to learn generic visual representations for various downstream tasks by means of image-level pretext tasks, such as instance discrimination. Some recent works [10, 11, 12, 13] observe that the image-level representations are sub-optimal for dense prediction tasks such as object detection and semantic segmentation. A potential reason is that image-level pretraining may overfit to holistic representations and fail to learn properties that are important outside of image classification. ",
|
| 74 |
+
"bbox": [
|
| 75 |
+
174,
|
| 76 |
+
602,
|
| 77 |
+
825,
|
| 78 |
+
741
|
| 79 |
+
],
|
| 80 |
+
"page_idx": 0
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"type": "text",
|
| 84 |
+
"text": "The goal of this work is to develop self-supervised pretraining that is aligned to object detection. In object detection, bounding boxes are widely adopted as the representation for objects. Translation and scale invariance for object detection are reflected by the location and size of the bounding boxes. An obvious representation gap exists between image-level pretraining and the object-level bounding boxes of object detection. ",
|
| 85 |
+
"bbox": [
|
| 86 |
+
174,
|
| 87 |
+
747,
|
| 88 |
+
825,
|
| 89 |
+
815
|
| 90 |
+
],
|
| 91 |
+
"page_idx": 0
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"type": "text",
|
| 95 |
+
"text": "Motivated by this, we present an object-level self-supervised pretraining framework, called Selective Object COntrastive learning (SoCo), specifically for the downstream task of object detection. To introduce object-level representations into pretraining, SoCo utilizes off-the-shelf selective search [14] to generate object proposals. Different from prior image-level contrastive learning methods which treat the whole image as an instance, SoCo treats each object proposal in the image as an independent instance. This enables us to design a new pretext task for learning object-level visual representations with properties that are compatible with object detection. Specifically, SoCo constructs object-level views where the scales and locations of the same object instance are augmented. Contrastive learning follows to maximize the similarity of the object across augmented views. ",
|
| 96 |
+
"bbox": [
|
| 97 |
+
174,
|
| 98 |
+
821,
|
| 99 |
+
823,
|
| 100 |
+
878
|
| 101 |
+
],
|
| 102 |
+
"page_idx": 0
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"type": "text",
|
| 106 |
+
"text": "",
|
| 107 |
+
"bbox": [
|
| 108 |
+
174,
|
| 109 |
+
92,
|
| 110 |
+
825,
|
| 111 |
+
160
|
| 112 |
+
],
|
| 113 |
+
"page_idx": 1
|
| 114 |
+
},
|
| 115 |
+
{
|
| 116 |
+
"type": "text",
|
| 117 |
+
"text": "The introduction of the object-level representation also allows us to further bridge the gap in network architecture between pretraining and finetuning. Object detection often involves dedicated modules, e.g., feature pyramid network (FPN) [15], and special-purpose sub-networks, e.g., R-CNN head [16, 17]. In contrast to image-level contrastive learning methods where only feature backbones are pretrained and transferred, SoCo performs pretraining over all the network modules used in detectors. As a result, all layers of the detectors can be well-initialized. ",
|
| 118 |
+
"bbox": [
|
| 119 |
+
173,
|
| 120 |
+
166,
|
| 121 |
+
825,
|
| 122 |
+
250
|
| 123 |
+
],
|
| 124 |
+
"page_idx": 1
|
| 125 |
+
},
|
| 126 |
+
{
|
| 127 |
+
"type": "text",
|
| 128 |
+
"text": "Experimentally, the proposed SoCo achieves state-of-the-art transfer performance from ImageNet to COCO. Concretely, by using Mask R-CNN with an R50-FPN backbone, it obtains $4 3 . 2 \\ : \\mathrm { A P ^ { \\tilde { b } b } } / 3 8 . 4$ $\\mathbf { A P } ^ { \\mathrm { m k } }$ on COCO with a $1 \\times$ schedule, which are $+ 4 . 3 \\ \\mathrm { A P } ^ { \\mathrm { b b } }$ $/ + 3 . 0 \\mathrm { \\ A P ^ { m k } }$ better than the supervised pretraining baseline, and $4 4 . 3 \\ : \\mathrm { A P ^ { b b } } / 3 9 . 6 \\ : \\mathrm { A P ^ { m k } }$ on COCO with a $2 \\times$ schedule, which are $+ \\dot { 3 } . 0 \\mathrm { A P ^ { b b } }$ $\\bar { / } + 2 . 3 \\ \\mathrm { A P ^ { \\mathrm { i n k } } }$ better than the supervised pretraining baseline. When transferred to Mask R-CNN with an R50-C4 backbone, it achieves $4 0 . 9 \\ : \\mathrm { \\dot { A } P ^ { b b } } / 3 5 . 3 \\ : \\mathrm { A P ^ { m k } }$ and $4 2 . 0 \\mathrm { A P ^ { b b } } / 3 6 . 3 \\mathrm { A P ^ { m k } }$ on the $1 \\times$ and $2 \\times$ schedule, respectively. ",
|
| 129 |
+
"bbox": [
|
| 130 |
+
173,
|
| 131 |
+
256,
|
| 132 |
+
825,
|
| 133 |
+
353
|
| 134 |
+
],
|
| 135 |
+
"page_idx": 1
|
| 136 |
+
},
|
| 137 |
+
{
|
| 138 |
+
"type": "text",
|
| 139 |
+
"text": "2 Related Work ",
|
| 140 |
+
"text_level": 1,
|
| 141 |
+
"bbox": [
|
| 142 |
+
174,
|
| 143 |
+
367,
|
| 144 |
+
321,
|
| 145 |
+
383
|
| 146 |
+
],
|
| 147 |
+
"page_idx": 1
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"type": "text",
|
| 151 |
+
"text": "Unsupervised feature learning has a long history in training deep neural networks. Auto-encoders [18] and Deep Boltzmann Machines [19] learn layers of representations by reconstructing image pixels. Unsupervised pretraining is often used as a method for initialization, which is followed by supervised training on the same dataset such as handwritten digits for classification. ",
|
| 152 |
+
"bbox": [
|
| 153 |
+
174,
|
| 154 |
+
392,
|
| 155 |
+
825,
|
| 156 |
+
448
|
| 157 |
+
],
|
| 158 |
+
"page_idx": 1
|
| 159 |
+
},
|
| 160 |
+
{
|
| 161 |
+
"type": "text",
|
| 162 |
+
"text": "Recent advances in unsupervised learning, especially self-supervised learning, tend to formulate the problem in a transfer learning scenario, where pretraining and finetuning are trained on different datasets with different purposes. Pretext tasks such as colorization [20], context prediction [21], inpainting [22] and rotation prediction [23] force the network to learn semantic information in order to solve the pretext tasks. Contrastive methods based on the instance discrimination pretext task [2] learn to map augmented views of the same instance to similar embeddings. Such approaches [24, 4, 5, 9, 25, 26] have shown strong transfer ability for a number of downstream tasks, sometimes even outperforming the supervised counterpart by large margins [4]. ",
|
| 163 |
+
"bbox": [
|
| 164 |
+
174,
|
| 165 |
+
454,
|
| 166 |
+
825,
|
| 167 |
+
565
|
| 168 |
+
],
|
| 169 |
+
"page_idx": 1
|
| 170 |
+
},
|
| 171 |
+
{
|
| 172 |
+
"type": "text",
|
| 173 |
+
"text": "With new technical innovations such as learning clustering assignments [7], large-scale contrastive learning has been successfully applied on non-curated data sets [27, 28]. While the linear evaluation result on ImageNet classification has improved significantly [3], the progress on transfer performance for dense prediction tasks has been limited. Due to this, a growing number of works investigate pretraining specifically for object detection and semantic segmentation. The idea is to shift imagelevel representations to pixel-level or region-level representations. VADer [29], PixPro [10] and DenseCL [11] propose to learn pixel-level representations by matching point features of the same physical location under different views. InsLoc [12] learns to match region-level features from composited imagery. DetCon [13] additionally uses the bottom-up segmentation of MCG [30] for supervising intra-segment pixel variations. UP-DETR [31] proposes a pretext task named random query patch detection to pretrain DETR detector. Self-EMD [32] explores self-supervised representation learning for object detection without using Imagenet images. ",
|
| 174 |
+
"bbox": [
|
| 175 |
+
174,
|
| 176 |
+
571,
|
| 177 |
+
825,
|
| 178 |
+
738
|
| 179 |
+
],
|
| 180 |
+
"page_idx": 1
|
| 181 |
+
},
|
| 182 |
+
{
|
| 183 |
+
"type": "text",
|
| 184 |
+
"text": "Our work is also motivated and inspired by object detection methods which advocate architectures and training schemes invariant to translation and scale transformations [15, 33]. We find that incorporating them in the pretraining stage also benefits object detection, which has largely been overlooked in previous self-supervised learning works. ",
|
| 185 |
+
"bbox": [
|
| 186 |
+
174,
|
| 187 |
+
743,
|
| 188 |
+
825,
|
| 189 |
+
799
|
| 190 |
+
],
|
| 191 |
+
"page_idx": 1
|
| 192 |
+
},
|
| 193 |
+
{
|
| 194 |
+
"type": "text",
|
| 195 |
+
"text": "3 Method ",
|
| 196 |
+
"text_level": 1,
|
| 197 |
+
"bbox": [
|
| 198 |
+
174,
|
| 199 |
+
816,
|
| 200 |
+
271,
|
| 201 |
+
833
|
| 202 |
+
],
|
| 203 |
+
"page_idx": 1
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"type": "text",
|
| 207 |
+
"text": "We introduce a new contrastive learning method which maximizes the similarity of object-level features representing different augmentations of the same object. Meanwhile, we introduce architectural alignment and important properties of object detection into pretraining when designing the objectlevel pretext task. We use an example where transfer learning is performed on Mask-RCNN with a R50-FPN backbone to illustrate the core design principles. To further demonstrate the extensibility and flexibility of our method, we also apply SoCo on a R50-C4 structure. In this section, we first present an overview of the proposed SoCo in Section 3.1. Then we describe the process of object proposal generation and view construction in Section 3.2. Finally, the object-level contrastive learning and our design principles are introduced in Section 3.3. ",
|
| 208 |
+
"bbox": [
|
| 209 |
+
174,
|
| 210 |
+
842,
|
| 211 |
+
825,
|
| 212 |
+
911
|
| 213 |
+
],
|
| 214 |
+
"page_idx": 1
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"type": "image",
|
| 218 |
+
"img_path": "images/efaee5b665f9402b9b8037f66209a95b7629f8b34d3536f5f212eadf82a062d0.jpg",
|
| 219 |
+
"image_caption": [
|
| 220 |
+
"Figure 1: Overview of SoCo. SoCo utilizes selective search to generate a set of object proposals for each raw image. $K$ proposals are randomly selected in each training step. We construct three views $\\left\\{ V _ { 1 } , V _ { 2 } , V _ { 3 } \\right\\}$ where the scales and locations of the same object are different. We adopt a backbone with FPN to encode image-level features and RoIAlign to extract object-level features. Object proposals are assigned to different pyramid levels according to their image areas. Contrastive learning is performed at the object level to learn translation-invariant and scale-invariant representations. The target network is updated by an exponential moving average of the online network. "
|
| 221 |
+
],
|
| 222 |
+
"image_footnote": [],
|
| 223 |
+
"bbox": [
|
| 224 |
+
183,
|
| 225 |
+
92,
|
| 226 |
+
815,
|
| 227 |
+
273
|
| 228 |
+
],
|
| 229 |
+
"page_idx": 2
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"type": "text",
|
| 233 |
+
"text": "",
|
| 234 |
+
"bbox": [
|
| 235 |
+
174,
|
| 236 |
+
411,
|
| 237 |
+
825,
|
| 238 |
+
467
|
| 239 |
+
],
|
| 240 |
+
"page_idx": 2
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"type": "text",
|
| 244 |
+
"text": "3.1 Overview ",
|
| 245 |
+
"text_level": 1,
|
| 246 |
+
"bbox": [
|
| 247 |
+
174,
|
| 248 |
+
486,
|
| 249 |
+
279,
|
| 250 |
+
500
|
| 251 |
+
],
|
| 252 |
+
"page_idx": 2
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"type": "text",
|
| 256 |
+
"text": "Figure 1 displays an overview of SoCo. SoCo aims to align pretraining to object detection in two aspects: 1) network architecture alignment between pretraining and object detection; 2) introducing central properties of detection. Concretely, besides pretraining a backbone as done in existing selfsupervised contrastive learning methods, SoCo also pretrains all the network modules used in an object detector, such as FPN and the head in the Mask R-CNN framework. As a result, all layers of the detector can be well-initialized. Furthermore, SoCo strives to learn object-level representations which are not only meaningful, but also invariant to translation and scale. To achieve this, it encourages diversity of scales and locations of objects by constructing multiple augmented views and applying a scale-aware assignment strategy for different levels of a feature pyramid. Finally, object-level contrastive learning is applied to maximize the feature similarity of the same object across augmented views. ",
|
| 257 |
+
"bbox": [
|
| 258 |
+
173,
|
| 259 |
+
511,
|
| 260 |
+
825,
|
| 261 |
+
662
|
| 262 |
+
],
|
| 263 |
+
"page_idx": 2
|
| 264 |
+
},
|
| 265 |
+
{
|
| 266 |
+
"type": "text",
|
| 267 |
+
"text": "3.2 Data Preprocessing ",
|
| 268 |
+
"text_level": 1,
|
| 269 |
+
"bbox": [
|
| 270 |
+
174,
|
| 271 |
+
683,
|
| 272 |
+
348,
|
| 273 |
+
698
|
| 274 |
+
],
|
| 275 |
+
"page_idx": 2
|
| 276 |
+
},
|
| 277 |
+
{
|
| 278 |
+
"type": "text",
|
| 279 |
+
"text": "Object Proposal Generation. Inspired by R-CNN [34] and Fast R-CNN [35], we use selective search [14], an unsupervised object proposal generation algorithm which takes into account color similarity, texture similarity, size of region and fit between regions, to generate a set of object proposals for each of the raw images. We represent each object proposal as a bounding box $b \\overset { \\cdot } { = } \\{ x , y , w , h \\}$ , where $( x , y )$ denotes the coordinates of the bounding box center, and $w$ and $h$ are the corresponding width and height, respectively. We keep only the proposals that satisfy the following requirements:√ √ 1) $1 / 3 \\leq w / h \\leq 3 ;$ 2) $0 . 3 \\leq \\dot { \\sqrt { w h } } / \\sqrt { W H } \\leq 0 . 8$ , where $W$ and $H$ denote width and height of the input image. The object proposal generation step is performed offline. In each training iteration, we randomly select $K$ proposals for each input image. ",
|
| 280 |
+
"bbox": [
|
| 281 |
+
173,
|
| 282 |
+
708,
|
| 283 |
+
825,
|
| 284 |
+
835
|
| 285 |
+
],
|
| 286 |
+
"page_idx": 2
|
| 287 |
+
},
|
| 288 |
+
{
|
| 289 |
+
"type": "text",
|
| 290 |
+
"text": "View Construction. Three views, namely $V _ { 1 }$ , $V _ { 2 }$ and $V _ { 3 }$ , are used in SoCo. The input image is resized to $2 2 4 \\times 2 2 4$ to obtain $V _ { 1 }$ . Then we apply a random crop with a scale range of [0.5, 1.0] on $V _ { 1 }$ in generating $V _ { 2 }$ . $V _ { 2 }$ is then resized to the same size as $V _ { 1 }$ and object proposals outside of $V _ { 2 }$ are dropped. Next, we downsample $V _ { 2 }$ to a fixed size (e.g. $1 1 2 \\times 1 1 2 ,$ ) to produce $V _ { 3 }$ . In all of these cases, the bounding boxes transformed according to the cropping and resizing of the RGB images (see Figure 1, Data Preprocessing). Finally, each view is randomly and independently augmented. We adopt the augmentation pipeline of BYOL [3] but discard the random crop augmentation since spatial transformation is already applied on all three views. Notice that the scale and location of the same object proposal are different across the augmented views, which enables the model to learn translation-invariant and scale-invariant object-level representations. ",
|
| 291 |
+
"bbox": [
|
| 292 |
+
174,
|
| 293 |
+
842,
|
| 294 |
+
823,
|
| 295 |
+
911
|
| 296 |
+
],
|
| 297 |
+
"page_idx": 2
|
| 298 |
+
},
|
| 299 |
+
{
|
| 300 |
+
"type": "text",
|
| 301 |
+
"text": "",
|
| 302 |
+
"bbox": [
|
| 303 |
+
173,
|
| 304 |
+
90,
|
| 305 |
+
825,
|
| 306 |
+
161
|
| 307 |
+
],
|
| 308 |
+
"page_idx": 3
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"type": "text",
|
| 312 |
+
"text": "Box Jitter. To further encourage variance of scales and locations of object proposals across views, we adopt a box jitter strategy on the generated proposals as an object-level data augmentation. Specifically, given an object proposal $\\boldsymbol { b } = \\{ x , y , w , h \\}$ , we randomly generate a jittered box $\\hat { b } =$ $\\{ \\hat { x } , \\hat { y } , \\hat { w } , \\hat { h } \\}$ as follows: 1) $\\hat { x } = x + r \\cdot w$ ; 2) ${ \\hat { y } } = y + r \\cdot h ;$ 3) $\\boldsymbol { \\hat { w } } = \\boldsymbol { w } + \\boldsymbol { r } \\cdot \\boldsymbol { w } ;$ 4) $\\hat { h } = h + r \\cdot h$ , where $r \\in [ - 0 . 1 , 0 . 1 ]$ . The box jitter is randomly applied on each proposal with a probability of 0.5. ",
|
| 313 |
+
"bbox": [
|
| 314 |
+
173,
|
| 315 |
+
166,
|
| 316 |
+
826,
|
| 317 |
+
243
|
| 318 |
+
],
|
| 319 |
+
"page_idx": 3
|
| 320 |
+
},
|
| 321 |
+
{
|
| 322 |
+
"type": "text",
|
| 323 |
+
"text": "3.3 Object-Level Contrastive Learning ",
|
| 324 |
+
"text_level": 1,
|
| 325 |
+
"bbox": [
|
| 326 |
+
176,
|
| 327 |
+
256,
|
| 328 |
+
457,
|
| 329 |
+
272
|
| 330 |
+
],
|
| 331 |
+
"page_idx": 3
|
| 332 |
+
},
|
| 333 |
+
{
|
| 334 |
+
"type": "text",
|
| 335 |
+
"text": "The goal of SoCo is to align pretraining to object detection. Here, we use the representative framework Mask R-CNN [17] with feature pyramid network (FPN) [15] to demonstrate our key design principles. The alignment mainly involves aligning the pretraining architecture with that of object detection and integrating important object detection properties such as object-level translation invariance and scale invariance into the pretraining. ",
|
| 336 |
+
"bbox": [
|
| 337 |
+
173,
|
| 338 |
+
281,
|
| 339 |
+
825,
|
| 340 |
+
353
|
| 341 |
+
],
|
| 342 |
+
"page_idx": 3
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"type": "text",
|
| 346 |
+
"text": "Aligning Pretraining Architecture to Object Detection. Following Mask R-CNN, we use a backbone with FPN as the image-level feature extractor $f ^ { I }$ . We denote the output of FPN as $\\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \\}$ with a stride of $\\{ 4 , 8 , 1 6 , 3 2 \\}$ . Here, we do not use $P _ { 6 }$ due to its low resolution. With the bounding box representation $b$ , RoIAlign [17] is applied to extract the foreground feature from the corresponding scale level. For further architectural alignment, we additionally introduce an R-CNN head $f ^ { H }$ into pretraining. The object-level feature representation $h$ of bounding box $b$ is extracted from an image view $V$ as: ",
|
| 347 |
+
"bbox": [
|
| 348 |
+
173,
|
| 349 |
+
358,
|
| 350 |
+
825,
|
| 351 |
+
455
|
| 352 |
+
],
|
| 353 |
+
"page_idx": 3
|
| 354 |
+
},
|
| 355 |
+
{
|
| 356 |
+
"type": "equation",
|
| 357 |
+
"img_path": "images/f89f439cb0cb0c8b584fa9ae97629ead84fd676837476610335e5c3ee3a888a7.jpg",
|
| 358 |
+
"text": "$$\nh = f ^ { H } ( \\mathrm { R o I A l i g n } ( f ^ { I } ( V ) , b ) ) .\n$$",
|
| 359 |
+
"text_format": "latex",
|
| 360 |
+
"bbox": [
|
| 361 |
+
395,
|
| 362 |
+
459,
|
| 363 |
+
602,
|
| 364 |
+
478
|
| 365 |
+
],
|
| 366 |
+
"page_idx": 3
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"type": "text",
|
| 370 |
+
"text": "SoCo uses two neural networks to learn, namely an online network and a target network. The online network and the target network share the same architecture but with different sets of weights. Concretely, the target network weights $f _ { \\xi } ^ { I } , f _ { \\xi } ^ { H }$ are the exponential moving average (EMA) with the momentum coefficient $\\tau$ of the online parameters $f _ { \\theta } ^ { I }$ and $f _ { \\theta } ^ { H }$ . Denote the set of object proposals $\\{ b _ { i } \\}$ considered in the image. Let $h _ { i }$ be the object-level representation of proposal $b _ { i }$ in view $V _ { 1 }$ , and $h _ { i } ^ { \\prime } , h _ { i } ^ { \\prime \\prime }$ be the representation of $b _ { i }$ in view $V _ { 2 } , V _ { 3 }$ . They are extracted using the online network and the target network respectively, ",
|
| 371 |
+
"bbox": [
|
| 372 |
+
173,
|
| 373 |
+
488,
|
| 374 |
+
825,
|
| 375 |
+
590
|
| 376 |
+
],
|
| 377 |
+
"page_idx": 3
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"type": "equation",
|
| 381 |
+
"img_path": "images/c688553915821526c28fc2e2fefc38fbb71711b15b63a576895bf2c93e78016c.jpg",
|
| 382 |
+
"text": "$$\nh _ { i } = f _ { \\theta } ^ { H } \\big ( \\mathrm { R o I A l i g n } \\big ( f _ { \\theta } ^ { I } ( V _ { 1 } ) , b _ { i } \\big ) \\big ) ,\n$$",
|
| 383 |
+
"text_format": "latex",
|
| 384 |
+
"bbox": [
|
| 385 |
+
387,
|
| 386 |
+
593,
|
| 387 |
+
607,
|
| 388 |
+
613
|
| 389 |
+
],
|
| 390 |
+
"page_idx": 3
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"type": "equation",
|
| 394 |
+
"img_path": "images/c4ef01395a0fffe3d6f6763ceb5958bffe97fd344afd6dcb7dfb4da4e3ca8a25.jpg",
|
| 395 |
+
"text": "$$\nh _ { i } ^ { \\prime } = f _ { \\xi } ^ { H } ( \\mathrm { R o I A l i g n } ( f _ { \\xi } ^ { I } ( V _ { 2 } ) , b _ { i } ) ) , \\quad h _ { i } ^ { \\prime \\prime } = f _ { \\xi } ^ { H } ( \\mathrm { R o I A l i g n } ( f _ { \\xi } ^ { I } ( V _ { 3 } ) , b _ { i } ) ) .\n$$",
|
| 396 |
+
"text_format": "latex",
|
| 397 |
+
"bbox": [
|
| 398 |
+
269,
|
| 399 |
+
627,
|
| 400 |
+
727,
|
| 401 |
+
647
|
| 402 |
+
],
|
| 403 |
+
"page_idx": 3
|
| 404 |
+
},
|
| 405 |
+
{
|
| 406 |
+
"type": "text",
|
| 407 |
+
"text": "We follow BYOL [3] for learning contrastive representations. The online network is appended with a projector $g _ { \\theta }$ , and a predictor $q _ { \\theta }$ for obtaining latent embeddings. Both $g _ { \\theta }$ and $q _ { \\theta }$ are two-layer MLPs. The target network is only appended with the projector $g _ { \\xi }$ for avoiding trivial solutions. We use $v _ { i }$ , $\\boldsymbol { v } _ { i } ^ { \\prime }$ and $v _ { i } ^ { \\prime \\prime }$ to denote the latent embeddings of object-level representations $\\{ h _ { i } , h _ { i } ^ { \\prime } , h _ { i } ^ { \\prime \\prime } \\}$ , respectively: ",
|
| 408 |
+
"bbox": [
|
| 409 |
+
173,
|
| 410 |
+
654,
|
| 411 |
+
826,
|
| 412 |
+
710
|
| 413 |
+
],
|
| 414 |
+
"page_idx": 3
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"type": "equation",
|
| 418 |
+
"img_path": "images/caab9fcc3be4167cce8738b81ac7f5a8c5ffa9c2cb0eb203541ac8c8d08cf0e5.jpg",
|
| 419 |
+
"text": "$$\nv _ { i } = q _ { \\theta } ( g _ { \\theta } ( h _ { i } ) ) , \\quad v _ { i } ^ { \\prime } = g _ { \\xi } ( h _ { i } ^ { \\prime } ) , \\quad v _ { i } ^ { \\prime \\prime } = g _ { \\xi } ( h _ { i } ^ { \\prime \\prime } ) .\n$$",
|
| 420 |
+
"text_format": "latex",
|
| 421 |
+
"bbox": [
|
| 422 |
+
338,
|
| 423 |
+
713,
|
| 424 |
+
660,
|
| 425 |
+
732
|
| 426 |
+
],
|
| 427 |
+
"page_idx": 3
|
| 428 |
+
},
|
| 429 |
+
{
|
| 430 |
+
"type": "text",
|
| 431 |
+
"text": "The contrastive loss for the $i$ -th object proposal is defined as: ",
|
| 432 |
+
"bbox": [
|
| 433 |
+
173,
|
| 434 |
+
742,
|
| 435 |
+
575,
|
| 436 |
+
757
|
| 437 |
+
],
|
| 438 |
+
"page_idx": 3
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"type": "equation",
|
| 442 |
+
"img_path": "images/83f8a071040b7b71fbdddbc749640d6768c6abbd1230bbea363bc38bb232b24f.jpg",
|
| 443 |
+
"text": "$$\n\\mathcal { L } _ { i } = - 2 \\cdot \\frac { \\langle v _ { i } , v _ { i } ^ { \\prime } \\rangle } { \\left. v _ { i } \\right. _ { 2 } \\cdot \\left. v _ { i } ^ { \\prime } \\right. _ { 2 } } - 2 \\cdot \\frac { \\langle v _ { i } , v _ { i } ^ { \\prime \\prime } \\rangle } { \\left. v _ { i } \\right. _ { 2 } \\cdot \\left. v _ { i } ^ { \\prime \\prime } \\right. _ { 2 } } .\n$$",
|
| 444 |
+
"text_format": "latex",
|
| 445 |
+
"bbox": [
|
| 446 |
+
348,
|
| 447 |
+
761,
|
| 448 |
+
650,
|
| 449 |
+
796
|
| 450 |
+
],
|
| 451 |
+
"page_idx": 3
|
| 452 |
+
},
|
| 453 |
+
{
|
| 454 |
+
"type": "text",
|
| 455 |
+
"text": "Then we can formulate the overall loss function for each image as: ",
|
| 456 |
+
"bbox": [
|
| 457 |
+
174,
|
| 458 |
+
799,
|
| 459 |
+
611,
|
| 460 |
+
814
|
| 461 |
+
],
|
| 462 |
+
"page_idx": 3
|
| 463 |
+
},
|
| 464 |
+
{
|
| 465 |
+
"type": "equation",
|
| 466 |
+
"img_path": "images/77fabf55e998da79083346da6be0c49b2f411147f634ab3633116d6062a30442.jpg",
|
| 467 |
+
"text": "$$\n\\mathcal { L } = \\frac { 1 } { K } \\sum _ { i = 1 } ^ { K } \\mathcal { L } _ { i } ,\n$$",
|
| 468 |
+
"text_format": "latex",
|
| 469 |
+
"bbox": [
|
| 470 |
+
446,
|
| 471 |
+
818,
|
| 472 |
+
550,
|
| 473 |
+
861
|
| 474 |
+
],
|
| 475 |
+
"page_idx": 3
|
| 476 |
+
},
|
| 477 |
+
{
|
| 478 |
+
"type": "text",
|
| 479 |
+
"text": "where $K$ is the number of object proposals. We symmetrize the loss $\\mathcal { L }$ in Eq. 6 by separately feeding $V _ { 1 }$ to the target network and $\\{ V _ { 2 } , V _ { 3 } \\}$ to the online network to compute $\\widetilde { \\mathcal { L } }$ . At each training iteration, we perform a stochastic optimization step to minimize $\\mathcal { L } ^ { \\mathrm { S o C o } } = \\mathcal { L } + \\widetilde { \\mathcal { L } }$ . ",
|
| 480 |
+
"bbox": [
|
| 481 |
+
174,
|
| 482 |
+
863,
|
| 483 |
+
825,
|
| 484 |
+
911
|
| 485 |
+
],
|
| 486 |
+
"page_idx": 3
|
| 487 |
+
},
|
| 488 |
+
{
|
| 489 |
+
"type": "text",
|
| 490 |
+
"text": "Scale-Aware Assignment. Mask R-CNN with FPN uses the IoU between anchors and ground-truth boxes to determine positive samples. It defines anchors to have areas of $\\{ 3 2 ^ { 2 } , 6 4 ^ { 2 } , \\overline { { { 1 2 8 ^ { 2 } } } } , 2 5 6 ^ { 2 } \\}$ pixels on $\\{ P _ { 2 } , P _ { 3 } , \\bar { P } _ { 4 } , P _ { 5 } \\}$ , respectively, which means ground-truth boxes of scale within a range are assigned to a specific pyramid level. Inspired by this, we propose a scale-aware assignment strategy, which greatly encourages the pretraining model to learn object-level scaleinvariant representations. Concretely, we assign object proposals of area within a range of $\\left\\{ 0 - 4 8 ^ { 2 } , 4 \\dot { 9 } ^ { 2 } - 9 6 ^ { 2 } , 9 7 ^ { 2 } - 1 9 2 ^ { 2 } , 1 9 \\dot { 3 } ^ { 2 } - 2 2 4 ^ { 2 } \\right\\}$ pixels to $\\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \\}$ , respectively. Notice that the maximum proposal size is $2 2 4 \\times 2 2 4$ since all of the views are resized to a fixed resolution smaller than 224. The advantage is that the same object proposals at different scales are encouraged to learn consistent representations through contrastive learning. As a result, SoCo learns object-level scale-invariant visual representations, which is important for object detection. ",
|
| 491 |
+
"bbox": [
|
| 492 |
+
174,
|
| 493 |
+
90,
|
| 494 |
+
825,
|
| 495 |
+
243
|
| 496 |
+
],
|
| 497 |
+
"page_idx": 4
|
| 498 |
+
},
|
| 499 |
+
{
|
| 500 |
+
"type": "text",
|
| 501 |
+
"text": "Introducing Properties of Detection to Pretraining. Here, we discuss how SoCo promotes important properties of object detection in the pretraining. Object detection uses tight bounding boxes to represent objects. To introduce object-level representations, SoCo generates object proposals by selective search. Translation invariance and scale invariance at the object level are regarded as the most important properties for object detection, i.e., feature representations of objects belonging to same category should be insensitive to scale and location. Recall that $V _ { 2 }$ is a randomly cropped patch of $V _ { 1 }$ . Random cropping introduces box shift and thus contrastive learning between $V _ { 1 }$ and $V _ { 2 }$ encourages the pretraining model to learn location-invariant representations. $V _ { 3 }$ is generated by downsampling $V _ { 2 }$ , which results in a scale augmentation of object proposals. With our scale-aware assignment strategy, the contrastive loss between $V _ { 1 }$ and $V _ { 3 }$ guides the pretraining towards learning scale-invariant visual representations. ",
|
| 502 |
+
"bbox": [
|
| 503 |
+
174,
|
| 504 |
+
250,
|
| 505 |
+
825,
|
| 506 |
+
401
|
| 507 |
+
],
|
| 508 |
+
"page_idx": 4
|
| 509 |
+
},
|
| 510 |
+
{
|
| 511 |
+
"type": "text",
|
| 512 |
+
"text": "Extending to Other Object Detectors. SoCo can be easily extended to align to other detectors besides Mask R-CNN with FPN. Here, we apply SoCo to Mask R-CNN with a C4 structure, which is a popular non-FPN detection framework. The modification is three-fold: 1) for all object proposals, RoIAlign is performed on $C _ { 4 }$ ; 2) the R-CNN head is replaced by the entire 5-th residual block; 3) view $V _ { 3 }$ is discarded and the remaining $V _ { 1 }$ and $V _ { 2 }$ are kept for object-level contrastive learning. Experiments and comparisons with state-of-the-art methods in Section 4.3 demonstrate the extensibility and flexibility of SoCo. ",
|
| 513 |
+
"bbox": [
|
| 514 |
+
174,
|
| 515 |
+
409,
|
| 516 |
+
825,
|
| 517 |
+
505
|
| 518 |
+
],
|
| 519 |
+
"page_idx": 4
|
| 520 |
+
},
|
| 521 |
+
{
|
| 522 |
+
"type": "text",
|
| 523 |
+
"text": "4 Experiments ",
|
| 524 |
+
"text_level": 1,
|
| 525 |
+
"bbox": [
|
| 526 |
+
174,
|
| 527 |
+
527,
|
| 528 |
+
312,
|
| 529 |
+
545
|
| 530 |
+
],
|
| 531 |
+
"page_idx": 4
|
| 532 |
+
},
|
| 533 |
+
{
|
| 534 |
+
"type": "text",
|
| 535 |
+
"text": "4.1 Pretraining Settings ",
|
| 536 |
+
"text_level": 1,
|
| 537 |
+
"bbox": [
|
| 538 |
+
176,
|
| 539 |
+
560,
|
| 540 |
+
352,
|
| 541 |
+
575
|
| 542 |
+
],
|
| 543 |
+
"page_idx": 4
|
| 544 |
+
},
|
| 545 |
+
{
|
| 546 |
+
"type": "text",
|
| 547 |
+
"text": "Architecture. Through the introduction of object proposals, the architectural discrepancy is reduced between pretraining and downstream detection finetuning. Mask R-CNN [17] is a commonly adopted framework to evaluate transfer performance. To demonstrate the extensibility and flexibility of SoCo, we provide details of SoCo alignment for the detection architectures R50-FPN and R50- C4. SoCo-R50-FPN: ResNet-50 [36] with FPN [15] is used as the image-level feature encoder. RoIAlign [17] is then used to extract RoI features on feature maps $\\{ P _ { 2 } , P _ { 3 } , P _ { 4 } , P _ { 5 } \\}$ with a stride of $\\{ 4 , 8 , 1 6 , 3 2 \\}$ . According to the image areas of object proposals, each RoI feature is then transformed to an object-level representation by the head network as in Mask R-CNN. SoCo-R50-C4: on the standard ResNet-50 architecture, we insert the RoI operation on the output of the 4-th residual block. The entire 5-th residual block is treated as the head network to encode object-level features. Both the projection network and prediction network are 2-layer MLPs which consist of a linear layer with output size 4096 followed by batch normalization [37], rectified linear units (ReLU) [38], and a final linear layer with output dimension 256. ",
|
| 548 |
+
"bbox": [
|
| 549 |
+
173,
|
| 550 |
+
587,
|
| 551 |
+
825,
|
| 552 |
+
766
|
| 553 |
+
],
|
| 554 |
+
"page_idx": 4
|
| 555 |
+
},
|
| 556 |
+
{
|
| 557 |
+
"type": "text",
|
| 558 |
+
"text": "Dataset. We adopt the widely used ImageNet [1] which consists of ${ \\sim } 1 . 2 8$ million images for self-supervised pretraining. ",
|
| 559 |
+
"bbox": [
|
| 560 |
+
174,
|
| 561 |
+
772,
|
| 562 |
+
821,
|
| 563 |
+
801
|
| 564 |
+
],
|
| 565 |
+
"page_idx": 4
|
| 566 |
+
},
|
| 567 |
+
{
|
| 568 |
+
"type": "text",
|
| 569 |
+
"text": "Data Augmentation. Once all views are constructed, we employ the data augmentation pipeline of BYOL [3]. Specifically, we apply random horizontal flip, color distortion, Gaussian blur, grayscaling, and the solarization operation. We remove the random crop augmentation since spatial transformations have already been applied on all views. ",
|
| 570 |
+
"bbox": [
|
| 571 |
+
174,
|
| 572 |
+
808,
|
| 573 |
+
825,
|
| 574 |
+
863
|
| 575 |
+
],
|
| 576 |
+
"page_idx": 4
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"type": "text",
|
| 580 |
+
"text": "Optimization. We use a 100-epoch training schedule in all the ablation studies and report the results of 100-epochs and 400-epochs in the comparisons with state-of-the-art methods. We use the LARS optimizer [39] with a cosine decay learning rate schedule [40] and a warm-up period of 10 epochs. ",
|
| 581 |
+
"bbox": [
|
| 582 |
+
176,
|
| 583 |
+
869,
|
| 584 |
+
825,
|
| 585 |
+
911
|
| 586 |
+
],
|
| 587 |
+
"page_idx": 4
|
| 588 |
+
},
|
| 589 |
+
{
|
| 590 |
+
"type": "table",
|
| 591 |
+
"img_path": "images/99353c7882d37d1e72337631eccb64ac26aa9dd0781aea4049c29be504f063cd.jpg",
|
| 592 |
+
"table_caption": [
|
| 593 |
+
"Table 1: Comparison with state-of-the-art methods on COCO by using Mask R-CNN with R50-FPN. "
|
| 594 |
+
],
|
| 595 |
+
"table_footnote": [],
|
| 596 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Epoch</td><td colspan=\"5\">1× Schedule</td><td rowspan=\"2\"></td><td colspan=\"6\">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>AP</td><td>Apbb</td><td>AP AP</td><td>Apmk</td><td>AP</td><td>AP</td></tr><tr><td>Scratch</td><td>-</td><td>31.0</td><td>49.5</td><td>33.2</td><td>28.5</td><td>46.8</td><td>30.4</td><td>38.4</td><td>57.5</td><td>42.0</td><td>34.7</td><td>54.8</td><td>37.2</td></tr><tr><td>Supervised</td><td>90</td><td>38.9</td><td>59.6</td><td>42.7</td><td>35.4</td><td>56.5</td><td>38.1</td><td>41.3</td><td>61.3</td><td>45.0</td><td>37.3</td><td>58.3</td><td>40.3</td></tr><tr><td>MoCo[4]</td><td>200</td><td>38.5</td><td>58.9</td><td>42.0</td><td>35.1</td><td>55.9</td><td>37.7</td><td>40.8</td><td>61.6</td><td>44.7</td><td>36.9</td><td>58.4</td><td>39.7</td></tr><tr><td>MoCo v2[5]</td><td>200</td><td>40.4</td><td>60.2</td><td>44.2</td><td>36.4</td><td>57.2</td><td>38.9</td><td>41.7</td><td>61.6</td><td>45.6</td><td>37.6</td><td>58.7</td><td>40.5</td></tr><tr><td>InfoMin [6]</td><td>200</td><td>40.6</td><td>60.6</td><td>44.6</td><td>36.7</td><td>57.7</td><td>39.4</td><td>42.5</td><td>62.7</td><td>46.8</td><td>38.4</td><td>59.7</td><td>41.4</td></tr><tr><td>BYOL[3]</td><td>300</td><td>40.4</td><td>61.6</td><td>44.1</td><td>37.2</td><td>58.8</td><td>39.8</td><td>42.3</td><td>62.6</td><td>46.2</td><td>38.3</td><td>59.6</td><td>41.1</td></tr><tr><td>SwAV[7]</td><td>400</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>42.3</td><td>62.8</td><td>46.3</td><td>38.2</td><td>60.0</td><td>41.0</td></tr><tr><td>ReSim-FPNT [45]</td><td>200</td><td>39.8</td><td>60.2</td><td>43.5</td><td>36.0</td><td>57.1</td><td>38.6</td><td>41.4</td><td>61.9</td><td>45.4</td><td>37.5</td><td>59.1</td><td>40.3</td></tr><tr><td>PixPro[10]</td><td>400</td><td>41.4</td><td>61.6</td><td>45.4</td><td>1</td><td>-</td><td>1</td><td>-</td><td>-</td><td>-</td><td>-</td><td>1</td><td>-</td></tr><tr><td>InsLoc [12]</td><td>400</td><td>42.0</td><td>62.3</td><td>45.8</td><td>37.6</td><td>59.0</td><td>40.5</td><td>43.3</td><td>63.6</td><td>47.3</td><td>38.8</td><td>60.9</td><td>41.7</td></tr><tr><td>DenseCL[11]</td><td>200</td><td>40.3</td><td>59.9</td><td>44.3</td><td>36.4</td><td>57.0</td><td>39.2</td><td>41.2</td><td>61.9</td><td>45.1</td><td>37.3</td><td>58.9</td><td>40.1</td></tr><tr><td>DetCons [13]</td><td>1000</td><td>41.8</td><td></td><td>1</td><td>37.4</td><td>-</td><td>：</td><td>42.9</td><td>1</td><td>-</td><td>38.1</td><td></td><td>：</td></tr><tr><td>DetConB [13]</td><td>1000</td><td>42.7</td><td></td><td>-</td><td>38.2</td><td>1</td><td>：</td><td>43.4</td><td>1</td><td>-</td><td>38.7</td><td>-</td><td>：</td></tr><tr><td>SoCo</td><td>100</td><td>42.3</td><td>62.5</td><td>46.5</td><td>37.6</td><td>59.1</td><td>40.5</td><td>43.2</td><td>63.3</td><td>47.3</td><td>38.8</td><td>60.6</td><td>41.9</td></tr><tr><td>SoCo</td><td>400</td><td>43.0</td><td>63.3</td><td>47.1</td><td>38.2</td><td>60.2</td><td>41.0</td><td>44.0</td><td>64.0</td><td>48.4</td><td>39.0</td><td>61.3</td><td>41.7</td></tr><tr><td>SoCo*</td><td>400</td><td>43.2</td><td>63.5</td><td>47.4</td><td>38.4</td><td>60.2</td><td>41.4</td><td>44.3</td><td>64.6</td><td>48.9</td><td>39.6</td><td>61.8</td><td>42.5</td></tr></table>",
|
| 597 |
+
"bbox": [
|
| 598 |
+
173,
|
| 599 |
+
111,
|
| 600 |
+
825,
|
| 601 |
+
324
|
| 602 |
+
],
|
| 603 |
+
"page_idx": 5
|
| 604 |
+
},
|
| 605 |
+
{
|
| 606 |
+
"type": "text",
|
| 607 |
+
"text": "The base learning rate $l r _ { \\mathrm { b a s e } }$ is set to 1.0 and is scaled linearly [41] with the batch size $\\mathit { l r } = { l r } _ { \\mathrm { b a s e } } \\times$ BatchSize/256). The weight decay is set to $1 . 0 \\times \\mathrm { e } ^ { - 5 }$ . The total batch size is set to 2048 over 16 Nvidia V100 GPUs. For the update of the target network, following [3], the momentum coefficient $\\tau$ starts from 0.99 and is increased to 1 during training. Synchronized batch normalization is enabled. ",
|
| 608 |
+
"bbox": [
|
| 609 |
+
174,
|
| 610 |
+
352,
|
| 611 |
+
825,
|
| 612 |
+
409
|
| 613 |
+
],
|
| 614 |
+
"page_idx": 5
|
| 615 |
+
},
|
| 616 |
+
{
|
| 617 |
+
"type": "text",
|
| 618 |
+
"text": "4.2 Transfer Learning Settings ",
|
| 619 |
+
"text_level": 1,
|
| 620 |
+
"bbox": [
|
| 621 |
+
174,
|
| 622 |
+
426,
|
| 623 |
+
401,
|
| 624 |
+
440
|
| 625 |
+
],
|
| 626 |
+
"page_idx": 5
|
| 627 |
+
},
|
| 628 |
+
{
|
| 629 |
+
"type": "text",
|
| 630 |
+
"text": "COCO [42] and Pascal VOC [43] datasets are used for transfer learning. Detectron2 [44] is used as the code base. ",
|
| 631 |
+
"bbox": [
|
| 632 |
+
173,
|
| 633 |
+
452,
|
| 634 |
+
823,
|
| 635 |
+
481
|
| 636 |
+
],
|
| 637 |
+
"page_idx": 5
|
| 638 |
+
},
|
| 639 |
+
{
|
| 640 |
+
"type": "text",
|
| 641 |
+
"text": "COCO Object Detection and Instance Segmentation. We use the COCO train2017 set which contains ${ \\sim } 1 1 8 \\mathrm { k }$ images with bounding box and instance segmentation annotations in 80 object categories. Transfer performance is evaluated on the COCO val2017 set. We adopt the Mask R-CNN detector [17] withfor object detection, and 5, andand 50-C4 backbones. We reportfor instance segmentation. $\\mathsf { A P } ^ { \\mathrm { b b } }$ ,e $\\mathrm { A \\hat { P } _ { 5 0 } ^ { b b } }$ andhe r $\\mathsf { A P } _ { 7 5 } ^ { \\mathsf { b b } }$ $\\mathbf { A P } ^ { \\mathrm { m k } }$ $\\mathbf { A P } _ { 5 0 } ^ { \\mathrm { m k } }$ $\\mathbf { A P } _ { 7 5 } ^ { \\mathrm { m k } }$ \nunder the COCO $1 \\times$ and $2 \\times$ schedules in comparisons with state-of-the-art methods. All ablation studies are conducted on the COCO $1 \\times$ schedule. ",
|
| 642 |
+
"bbox": [
|
| 643 |
+
174,
|
| 644 |
+
487,
|
| 645 |
+
825,
|
| 646 |
+
585
|
| 647 |
+
],
|
| 648 |
+
"page_idx": 5
|
| 649 |
+
},
|
| 650 |
+
{
|
| 651 |
+
"type": "text",
|
| 652 |
+
"text": "Pascal VOC Object Detection. We use the Pascal VOC trainval $0 7 + 1 2$ set which contains ${ \\sim } 1 6 . 5 \\mathrm { k }$ images with bounding box annotations in 20 object categories as the training set. Transfer performance is evaluated on the Pascal VOC test2007 set. We report the results of $\\mathsf { A P } ^ { \\mathrm { b b } }$ , $\\mathsf { A P } _ { 5 0 } ^ { \\mathrm { b b } }$ and $\\mathsf { A P } _ { 7 5 } ^ { \\mathrm { b b } }$ for object detection. ",
|
| 653 |
+
"bbox": [
|
| 654 |
+
174,
|
| 655 |
+
592,
|
| 656 |
+
825,
|
| 657 |
+
647
|
| 658 |
+
],
|
| 659 |
+
"page_idx": 5
|
| 660 |
+
},
|
| 661 |
+
{
|
| 662 |
+
"type": "text",
|
| 663 |
+
"text": "Optimization. All the pretrained weights except for projection and prediction are loaded into the object detection network for transfer. Following [5], synchronized batch normalization is used across all layers including the newly initialized batch normalization layers in finetuning. For both the COCO and Pascal VOC datasets, we finetune with stochastic gradient descent and a batch size of 16 split across 8 GPUs. For COCO transfer, we use a weight decay of $2 . 5 \\times \\mathrm { e } ^ { - 5 }$ , and a base learning rate of 0.02 that increases linearly for the first 1000 iterations and drops twice by a factor of 10, after $\\frac { 2 } { 3 }$ and $\\frac { 8 } { 9 }$ of the total training time. For Pascal VOC transfer, the weight decay is $1 . 0 \\times \\mathrm { e } ^ { - 4 }$ , and the base learning rate is set to 0.02 and divided by 10 at $\\textstyle { \\frac { 3 } { 4 } }$ and $\\textstyle { \\frac { 1 1 } { 1 2 } }$ of the total training time. ",
|
| 664 |
+
"bbox": [
|
| 665 |
+
173,
|
| 666 |
+
654,
|
| 667 |
+
825,
|
| 668 |
+
771
|
| 669 |
+
],
|
| 670 |
+
"page_idx": 5
|
| 671 |
+
},
|
| 672 |
+
{
|
| 673 |
+
"type": "text",
|
| 674 |
+
"text": "4.3 Comparison with State-of-the-Art Methods ",
|
| 675 |
+
"text_level": 1,
|
| 676 |
+
"bbox": [
|
| 677 |
+
174,
|
| 678 |
+
787,
|
| 679 |
+
514,
|
| 680 |
+
803
|
| 681 |
+
],
|
| 682 |
+
"page_idx": 5
|
| 683 |
+
},
|
| 684 |
+
{
|
| 685 |
+
"type": "text",
|
| 686 |
+
"text": "Mask R-CNN with R50-FPN on COCO. Table 1 shows the transfer results for Mask R-CNN with R50-FPN backbone. We compare SoCo with the state-of-the-art unsupervised pretraining methods on the COCO $1 \\times$ and $2 \\times$ schedules. We report our results under 100 epochs and 400 epochs of pretraining. On the 100-epoch training schedule, SoCo achieves $4 2 . 3 \\mathrm { \\ A p ^ { b b } / 3 7 . 6 \\ A P ^ { m k } }$ and $4 3 . 2 \\ : \\dot { \\mathrm { A P ^ { b b } } } / \\ : 3 8 . { \\bar { 8 } } \\ : \\mathrm { A P ^ { m k } }$ on the $1 \\times$ and $2 \\times$ schedules, respectively. Pretrained for 400 epochs, SoCo outperforms all previous pretraining methods designed for either image classification or object detection, achieving $4 \\dot { 3 } . 0 \\ : \\mathrm { A P ^ { b b } } \\ : \\dot { / } 3 8 . 2 \\ : \\mathrm { A P ^ { m i } }$ for the $1 \\times$ schedule and $4 4 . 0 \\mathrm { A P ^ { \\bar { b } b } / 3 9 . 0 \\mathrm { A P ^ { m k } } }$ for the $2 \\times$ schedule. Moreover, we propose an enhanced version named $S o C o ^ { * }$ , which constructs an additional view $V _ { 4 }$ for pretraining. Similar to the construction step of $V _ { 2 }$ , $V _ { 4 }$ is a randomly cropped patch of $V _ { 1 }$ . We resize $V _ { 4 }$ to $1 9 2 \\times 1 9 2$ and perform the same forward process as for $V _ { 2 }$ and $V _ { 3 }$ . Compared with SoCo, $\\mathrm { S o C o ^ { * } }$ further boosts the performance and obtains an improvement of $+ 0 . 2 \\mathrm { \\ A P ^ { b b } \\ } \\bar { / } + 0 . 2$ $\\mathbf { A P } ^ { \\mathrm { m k } }$ on the $1 \\times$ schedule, achieving $\\dot { 4 } 3 . 2 \\ : \\mathrm { A P ^ { b b } } / 3 8 . 4 \\ : \\mathrm { A P ^ { m k } }$ , and $+ 0 . { \\dot { 3 } } \\operatorname { A P } ^ { \\mathrm { b b } } / + 0 . 6 \\operatorname { A P } ^ { \\mathrm { m k } }$ on the $2 \\times$ schedule, achieving $4 4 . 3 \\ : \\mathrm { A P ^ { b b } } / 3 9 . \\mathsf { \\breve { 6 } A P ^ { m k } }$ , respectively. ",
|
| 687 |
+
"bbox": [
|
| 688 |
+
174,
|
| 689 |
+
814,
|
| 690 |
+
825,
|
| 691 |
+
911
|
| 692 |
+
],
|
| 693 |
+
"page_idx": 5
|
| 694 |
+
},
|
| 695 |
+
{
|
| 696 |
+
"type": "table",
|
| 697 |
+
"img_path": "images/f3c79168bc1219600c1425129d40d2ad19c5df24d0a45f30191dcb6477a211ef.jpg",
|
| 698 |
+
"table_caption": [
|
| 699 |
+
"Table 2: Comparison with state-of-the-art methods on COCO using Mask R-CNN with R50-C4. "
|
| 700 |
+
],
|
| 701 |
+
"table_footnote": [],
|
| 702 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Epoch</td><td colspan=\"6\">1× Schedule</td><td colspan=\"6\">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td><td>Apbb</td><td>AP8</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td></tr><tr><td>Scratch</td><td>-</td><td>26.4</td><td>44.0</td><td>27.8</td><td>29.3</td><td>46.9</td><td>30.8</td><td>35.6</td><td>54.6</td><td>38.2</td><td>31.4</td><td>51.5</td><td>33.5</td></tr><tr><td>Supervised</td><td>90</td><td>38.2</td><td>58.2</td><td>41.2</td><td>33.3</td><td>54.7</td><td>35.2</td><td>40.0</td><td>59.9</td><td>43.1</td><td>34.7</td><td>56.5</td><td>36.9</td></tr><tr><td>MoCo [4]</td><td>200</td><td>38.5</td><td>58.3</td><td>41.6</td><td>33.6</td><td>54.8</td><td>35.6</td><td>40.7</td><td>60.5</td><td>44.1</td><td>35.4</td><td>57.3</td><td>37.6 37.1</td></tr><tr><td>SimCLR [9] MoCo v2 [5]</td><td>200 800</td><td>- 39.3</td><td>-</td><td>■</td><td>-</td><td>-</td><td>1</td><td>39.6 41.2</td><td>59.1 60.9</td><td>42.9 44.6</td><td>34.6 35.8</td><td>55.9 57.7</td><td>38.2</td></tr><tr><td>InfoMin 6]</td><td></td><td></td><td>58.9</td><td>42.5</td><td>34.3</td><td>55.7</td><td>36.5</td><td></td><td>61.2</td><td>45.0</td><td></td><td>57.9</td><td>38.3</td></tr><tr><td>BYOL[3]</td><td>200 300</td><td>39.0</td><td>58.5</td><td>42.0</td><td>34.1</td><td>55.2</td><td>36.3</td><td>41.3</td><td></td><td></td><td>36.0</td><td>56.8</td><td>37.3</td></tr><tr><td>SwAV[7]</td><td>400</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>40.3</td><td>60.5</td><td>43.9</td><td>35.1</td><td></td><td>36.6</td></tr><tr><td>SimSiam [8]</td><td>200</td><td>- 39.2</td><td>- 59.3</td><td>-</td><td>1</td><td>1</td><td>-</td><td>39.6</td><td>60.1</td><td>42.9</td><td>34.7</td><td>56.6</td><td></td></tr><tr><td>PixPro[10]</td><td>400</td><td>40.5</td><td>59.8</td><td>42.1</td><td>34.4</td><td>56.0</td><td>36.7</td><td>-</td><td>：</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>InsLoc [12]</td><td>400</td><td>39.8</td><td>59.6</td><td>44.0 42.9</td><td>-</td><td>-</td><td>-</td><td>- 41.8</td><td>- 61.6</td><td>-</td><td>-</td><td>- 58.2</td><td>- 38.8</td></tr><tr><td></td><td></td><td></td><td></td><td></td><td>34.7</td><td>56.3</td><td>36.9</td><td></td><td></td><td>45.4</td><td>36.3</td><td></td><td></td></tr><tr><td>SoCo</td><td>100</td><td>40.4</td><td>60.4</td><td>43.7</td><td>34.9</td><td>56.8</td><td>37.0</td><td>41.1</td><td>61.0</td><td>44.4</td><td>35.6</td><td>57.5</td><td>38.0</td></tr><tr><td>SoCo</td><td>400</td><td>40.9</td><td>60.9</td><td>44.3</td><td>35.3</td><td>57.5</td><td>37.3</td><td>42.0</td><td>61.8</td><td>45.6</td><td>36.3</td><td>58.5</td><td>38.8</td></tr></table>",
|
| 703 |
+
"bbox": [
|
| 704 |
+
173,
|
| 705 |
+
111,
|
| 706 |
+
825,
|
| 707 |
+
299
|
| 708 |
+
],
|
| 709 |
+
"page_idx": 6
|
| 710 |
+
},
|
| 711 |
+
{
|
| 712 |
+
"type": "table",
|
| 713 |
+
"img_path": "images/ea2a893cc0304e2d6654cdfc1c0617449055ca5445f110c17ab0613d307accac.jpg",
|
| 714 |
+
"table_caption": [
|
| 715 |
+
"Table 3: Comparison with state-of-the-art methods on Pascal VOC. Faster R-CNN with R50-C4 is adopted. Table is split to two sub-tables. "
|
| 716 |
+
],
|
| 717 |
+
"table_footnote": [],
|
| 718 |
+
"table_body": "<table><tr><td colspan=\"5\">(a) Sub-table 1.</td></tr><tr><td>Methods</td><td>Epoch</td><td>APbb</td><td>AP</td><td>AP</td></tr><tr><td>Scratch Supervised</td><td>- 90</td><td>33.8 53.5</td><td>60.2 81.3</td><td>33.1 58.8</td></tr><tr><td>ReSim-C4 [45] PixPro [10] InsLoc [12] DenseCL[11]</td><td>200 400 400 200</td><td>58.7 60.2 58.4 58.7</td><td>83.1 83.8 83.0 82.8</td><td>66.3 67.7 65.3 65.2</td></tr><tr><td>SoCo SoCo</td><td>100 400</td><td>59.1 59.7</td><td>83.4 83.8</td><td>65.6 66.8</td></tr></table>",
|
| 719 |
+
"bbox": [
|
| 720 |
+
173,
|
| 721 |
+
347,
|
| 722 |
+
485,
|
| 723 |
+
502
|
| 724 |
+
],
|
| 725 |
+
"page_idx": 6
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"type": "table",
|
| 729 |
+
"img_path": "images/936d9d1ecb32da921c24e0a110561bcb83ff3d6172134e8e70d462836164b548.jpg",
|
| 730 |
+
"table_caption": [
|
| 731 |
+
"(b) Sub-table 2. "
|
| 732 |
+
],
|
| 733 |
+
"table_footnote": [],
|
| 734 |
+
"table_body": "<table><tr><td>Methods</td><td>Epoch</td><td>Apbb</td><td>AP</td><td>AP</td></tr><tr><td>MoCo[4]</td><td>200</td><td>55.9</td><td>81.5</td><td>62.6</td></tr><tr><td>SimCLR [9]</td><td>1000</td><td>56.3</td><td>81.9</td><td>62.5</td></tr><tr><td>MoCo v2[5]</td><td>800</td><td>57.6</td><td>82.7</td><td>64.4</td></tr><tr><td>InfoMin [6]</td><td>200</td><td>57.6</td><td>82.7</td><td>64.6</td></tr><tr><td>BYOL[3]</td><td>300</td><td>51.9</td><td>81.0</td><td>56.5</td></tr><tr><td>SwAV[7]</td><td>400</td><td>45.1</td><td>77.4</td><td>46.5</td></tr><tr><td>SimSiam[8]</td><td>200</td><td>57.0</td><td>82.4</td><td>63.7</td></tr><tr><td>ReSim-FPN[45]</td><td>200</td><td>59.2</td><td>82.9</td><td>65.9</td></tr></table>",
|
| 735 |
+
"bbox": [
|
| 736 |
+
509,
|
| 737 |
+
362,
|
| 738 |
+
820,
|
| 739 |
+
486
|
| 740 |
+
],
|
| 741 |
+
"page_idx": 6
|
| 742 |
+
},
|
| 743 |
+
{
|
| 744 |
+
"type": "text",
|
| 745 |
+
"text": "",
|
| 746 |
+
"bbox": [
|
| 747 |
+
173,
|
| 748 |
+
531,
|
| 749 |
+
825,
|
| 750 |
+
616
|
| 751 |
+
],
|
| 752 |
+
"page_idx": 6
|
| 753 |
+
},
|
| 754 |
+
{
|
| 755 |
+
"type": "text",
|
| 756 |
+
"text": "Mask R-CNN with R50-C4 on COCO. To demonstrate the extensibility and flexibility of SoCo, we also evaluate transfer performance using Mask R-CNN with R50-C4 backbone on the COCO benchmark. We report the results of SoCo under 100 epochs and 400 epochs. Table 2 compares the proposed method to previous state-of-the-art methods. Without bells and whistles, SoCo obtains state-of-the-art performance, achieving $4 0 . 9 ~ \\mathrm { A P ^ { b b } } / 3 5 . 3 ~ \\mathrm { A P ^ { m k } }$ and $4 2 . 0 ~ \\mathrm { A P ^ { b b } } / 3 6 . 3 ~ \\mathrm { A P ^ { m k } }$ on the COCO $1 \\times$ and $2 \\times$ schedules, respectively. ",
|
| 757 |
+
"bbox": [
|
| 758 |
+
174,
|
| 759 |
+
621,
|
| 760 |
+
825,
|
| 761 |
+
705
|
| 762 |
+
],
|
| 763 |
+
"page_idx": 6
|
| 764 |
+
},
|
| 765 |
+
{
|
| 766 |
+
"type": "text",
|
| 767 |
+
"text": "Faster R-CNN with R50-C4 on Pascal VOC. We also evaluate the transfer ability of SoCo on the Pascal VOC benchmark. Following [5], we adopt Faster R-CNN with R50-C4 backbone for transfer learning. Table 3 shows the comparison. SoCo obtains an improvement of $+ 6 . 2 \\ \\mathrm { A P } ^ { \\mathrm { b b } }$ against the supervised pretraining baseline, achieving $5 9 . 7 \\mathrm { A P } ^ { \\mathrm { b b } }$ for VOC object detection. ",
|
| 768 |
+
"bbox": [
|
| 769 |
+
174,
|
| 770 |
+
710,
|
| 771 |
+
825,
|
| 772 |
+
767
|
| 773 |
+
],
|
| 774 |
+
"page_idx": 6
|
| 775 |
+
},
|
| 776 |
+
{
|
| 777 |
+
"type": "text",
|
| 778 |
+
"text": "4.4 Ablation Study ",
|
| 779 |
+
"text_level": 1,
|
| 780 |
+
"bbox": [
|
| 781 |
+
174,
|
| 782 |
+
786,
|
| 783 |
+
316,
|
| 784 |
+
801
|
| 785 |
+
],
|
| 786 |
+
"page_idx": 6
|
| 787 |
+
},
|
| 788 |
+
{
|
| 789 |
+
"type": "text",
|
| 790 |
+
"text": "To further understand the advantages of SoCo, we conduct a series of ablation studies that examine the effectiveness of object-level contrastive learning, the effects of alignment between pretraining and detection, and different hyper-parameters. For all ablation studies, we use Mask R-CNN with R50-FPN backbone for transfer learning and adopt a 100-epoch SoCo pretraining schedule. Transfer performance is evaluated under the COCO $1 \\times$ schedule. For the ablation of each hyper-parameter or component, we fix all other hyper-parameters to the following default settings: view $V _ { 3 }$ with $1 1 2 \\times 1 1 2$ resolution, batch size of 2048, momentum coefficient $\\tau = 0 . 9 9$ , proposal number $K = 4$ box jitter and scale-aware assignment are used, FPN and R-CNN head are pretrained and transferred, and selective search is used as the proposal generator. ",
|
| 791 |
+
"bbox": [
|
| 792 |
+
174,
|
| 793 |
+
814,
|
| 794 |
+
825,
|
| 795 |
+
911
|
| 796 |
+
],
|
| 797 |
+
"page_idx": 6
|
| 798 |
+
},
|
| 799 |
+
{
|
| 800 |
+
"type": "table",
|
| 801 |
+
"img_path": "images/c6b2678b7f1e5b8a986f9378ce101eaaeb776ccb7b5dacfb234a6d4f09d456ac.jpg",
|
| 802 |
+
"table_caption": [
|
| 803 |
+
"Table 4: Ablation study on the effectiveness of aligning pretraining to object detection. "
|
| 804 |
+
],
|
| 805 |
+
"table_footnote": [],
|
| 806 |
+
"table_body": "<table><tr><td>Whole Image</td><td>Selective Search</td><td>FPN</td><td>Head</td><td>Scale-aware Assignment</td><td>Box Jitter</td><td>Multi View</td><td>Apbb</td><td>Apmk</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>38.1</td><td>34.4</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>40.6 (+2.5)</td><td>36.8 (+2.4)</td></tr><tr><td>√</td><td></td><td></td><td></td><td></td><td></td><td></td><td>40.2 (+2.1)</td><td>36.2 (+1.8)</td></tr><tr><td>√</td><td></td><td></td><td>广</td><td></td><td></td><td></td><td>41.2 (+3.1)</td><td>37.0 (+2.6)</td></tr><tr><td>√</td><td></td><td></td><td></td><td>1</td><td></td><td></td><td>41.6 (+3.5)</td><td>37.3 (+2.9)</td></tr><tr><td>√</td><td>>>>>>></td><td>>>>>></td><td>1</td><td></td><td></td><td>4</td><td>41.7 (+3.6)</td><td>37.5 (+3.1)</td></tr><tr><td>√</td><td></td><td></td><td>厂</td><td>厂</td><td></td><td></td><td>42.3 (+4.2)</td><td>37.6 (+3.2)</td></tr></table>",
|
| 807 |
+
"bbox": [
|
| 808 |
+
179,
|
| 809 |
+
112,
|
| 810 |
+
820,
|
| 811 |
+
251
|
| 812 |
+
],
|
| 813 |
+
"page_idx": 7
|
| 814 |
+
},
|
| 815 |
+
{
|
| 816 |
+
"type": "table",
|
| 817 |
+
"img_path": "images/4a25bd173b3e70c3c2a8472bec56b3bca0b385bcea78cebfd0d5b15962bfbee6.jpg",
|
| 818 |
+
"table_caption": [
|
| 819 |
+
"Table 5: Ablation studies on hyper-parameters for the proposed SoCo method. y on image size of view $V _ { 3 }$ . (c) Study on proposal generation and proposal number $K$ "
|
| 820 |
+
],
|
| 821 |
+
"table_footnote": [],
|
| 822 |
+
"table_body": "<table><tr><td>Image Size</td><td>Apbb</td><td>Apmk</td></tr><tr><td>96</td><td>42.1</td><td>37.7</td></tr><tr><td>112</td><td>42.3</td><td>37.6</td></tr><tr><td>128</td><td>42.1</td><td>37.7</td></tr><tr><td>160</td><td>42.0</td><td>37.6</td></tr><tr><td>192</td><td>42.2</td><td>37.8</td></tr></table>",
|
| 823 |
+
"bbox": [
|
| 824 |
+
192,
|
| 825 |
+
314,
|
| 826 |
+
398,
|
| 827 |
+
415
|
| 828 |
+
],
|
| 829 |
+
"page_idx": 7
|
| 830 |
+
},
|
| 831 |
+
{
|
| 832 |
+
"type": "table",
|
| 833 |
+
"img_path": "images/4bcc7356b2fcb1403d6e3393ed451e3fa1f3922a59938ea9fc5d4be5c12a6379.jpg",
|
| 834 |
+
"table_caption": [
|
| 835 |
+
"(b) Study on batch size. "
|
| 836 |
+
],
|
| 837 |
+
"table_footnote": [],
|
| 838 |
+
"table_body": "<table><tr><td>Batch Size</td><td>APbb</td><td>Apmk</td></tr><tr><td>512</td><td>41.7</td><td>37.6</td></tr><tr><td>1024</td><td>41.9</td><td>37.6</td></tr><tr><td>2048</td><td>42.3</td><td>37.6</td></tr><tr><td>4096</td><td>41.4</td><td>37.3</td></tr></table>",
|
| 839 |
+
"bbox": [
|
| 840 |
+
194,
|
| 841 |
+
453,
|
| 842 |
+
397,
|
| 843 |
+
537
|
| 844 |
+
],
|
| 845 |
+
"page_idx": 7
|
| 846 |
+
},
|
| 847 |
+
{
|
| 848 |
+
"type": "table",
|
| 849 |
+
"img_path": "images/095b4c978a75b84432bce320205feffff37f3d2db47eb797cf9e505432324cf9.jpg",
|
| 850 |
+
"table_caption": [
|
| 851 |
+
"(d) Study on momentum coefficient $\\tau$ "
|
| 852 |
+
],
|
| 853 |
+
"table_footnote": [],
|
| 854 |
+
"table_body": "<table><tr><td>Selective Search</td><td>Random</td><td>K</td><td>APbb</td><td>Apmk</td></tr><tr><td><</td><td></td><td>1</td><td>41.6</td><td>37.3</td></tr><tr><td></td><td></td><td>486</td><td>42.3 41.6</td><td>37.6</td></tr><tr><td>√</td><td></td><td></td><td></td><td>37.4</td></tr><tr><td>√</td><td></td><td></td><td>41.2</td><td>37.0</td></tr><tr><td></td><td>>>></td><td>148</td><td>41.4</td><td>36.9</td></tr><tr><td></td><td></td><td></td><td>NaN NaN</td><td>NaN NaN</td></tr></table>",
|
| 855 |
+
"bbox": [
|
| 856 |
+
449,
|
| 857 |
+
314,
|
| 858 |
+
799,
|
| 859 |
+
441
|
| 860 |
+
],
|
| 861 |
+
"page_idx": 7
|
| 862 |
+
},
|
| 863 |
+
{
|
| 864 |
+
"type": "table",
|
| 865 |
+
"img_path": "images/a49ae7d1a316c635aaea01b3ce6c8c54a5f6c0fa7e2cc476e7fa9b43a9cba701.jpg",
|
| 866 |
+
"table_caption": [],
|
| 867 |
+
"table_footnote": [],
|
| 868 |
+
"table_body": "<table><tr><td>T</td><td>APbb</td><td>Apmk</td></tr><tr><td>0.98</td><td>35.0</td><td>31.7</td></tr><tr><td>0.99</td><td>42.3</td><td>37.6</td></tr><tr><td>0.993</td><td>41.8</td><td>37.6</td></tr></table>",
|
| 869 |
+
"bbox": [
|
| 870 |
+
539,
|
| 871 |
+
467,
|
| 872 |
+
709,
|
| 873 |
+
540
|
| 874 |
+
],
|
| 875 |
+
"page_idx": 7
|
| 876 |
+
},
|
| 877 |
+
{
|
| 878 |
+
"type": "text",
|
| 879 |
+
"text": "",
|
| 880 |
+
"bbox": [
|
| 881 |
+
176,
|
| 882 |
+
573,
|
| 883 |
+
823,
|
| 884 |
+
602
|
| 885 |
+
],
|
| 886 |
+
"page_idx": 7
|
| 887 |
+
},
|
| 888 |
+
{
|
| 889 |
+
"type": "text",
|
| 890 |
+
"text": "Effectiveness of Aligning Pretraining to Object Detection. We ablate each component of SoCo step by step to demonstrate the the effectiveness of aligning pretraining to object detection. Table 4 reports the studies. The baseline treats the whole image as an instance, without considering any detection properties or architecture alignment. It obtains $3 8 . 1 \\mathrm { \\ A P ^ { b b } }$ and $3 4 . 4 \\mathrm { A P ^ { m k } }$ . Selective search introduces object-level representations. By leveraging generated object proposals and conducting contrastive learning at the object level, our method obtains an improvement of $+ 2 . 5 \\mathrm { \\ A P ^ { b b } }$ and $+ 2 . 4$ $\\mathbf { A P } ^ { \\mathrm { m k } }$ . Next, we further minimize the architectural discrepancy between pretraining and object detection pipeline by introducing FPN and an R-CNN head into the pretraining architecture. We observe that only introducing FPN into pretraining slightly hurt the performance. In contrast, including both FPN and the R-CNN head improves the transfer performance to $4 1 . 2 ~ \\mathrm { A P ^ { b b } } ~ / ~ 3 7 . 0$ $\\mathbf { A P } ^ { \\mathrm { m k } }$ , which verifies the effectiveness of architectural alignment between self-supervised pretraining and downstream tasks. On top of architectural alignment, we further leverage scale-aware assignment which encourages scale-invariant object-level visual representations, and an improvement of $+ 3 . 5$ $\\mathsf { A P } ^ { \\mathrm { b b } }$ / $+ 2 . 9 \\mathrm { \\ A P ^ { \\mathrm { \\bar { m k } } } }$ against the baseline is found. The box jitter strategy slightly elevates performance. Finally, multiple views $( V _ { 3 } )$ further directs SoCo towards learning scale-invariant and translationinvariant representations, and our method achieves $4 2 . 3 \\mathrm { \\ A P ^ { b b } }$ and $\\mathsf { \\bar { 3 } 7 . 6 \\mathbf { A P } ^ { m k } }$ . ",
|
| 891 |
+
"bbox": [
|
| 892 |
+
174,
|
| 893 |
+
608,
|
| 894 |
+
825,
|
| 895 |
+
828
|
| 896 |
+
],
|
| 897 |
+
"page_idx": 7
|
| 898 |
+
},
|
| 899 |
+
{
|
| 900 |
+
"type": "text",
|
| 901 |
+
"text": "Ablation Study on Hyper-Parameters. Table 5 examines sensitivity to the hyper-parameters of SoCo. ",
|
| 902 |
+
"bbox": [
|
| 903 |
+
173,
|
| 904 |
+
835,
|
| 905 |
+
821,
|
| 906 |
+
863
|
| 907 |
+
],
|
| 908 |
+
"page_idx": 7
|
| 909 |
+
},
|
| 910 |
+
{
|
| 911 |
+
"type": "text",
|
| 912 |
+
"text": "Table 5(a) ablates the resolution of view $V _ { 3 }$ . SoCo is found to be insensitive to the image size of $V _ { 3 }$ We use $1 1 2 \\times 1 1 2$ as the default resolution due to its slightly better transfer performance and low training computation cost. ",
|
| 913 |
+
"bbox": [
|
| 914 |
+
174,
|
| 915 |
+
869,
|
| 916 |
+
825,
|
| 917 |
+
911
|
| 918 |
+
],
|
| 919 |
+
"page_idx": 7
|
| 920 |
+
},
|
| 921 |
+
{
|
| 922 |
+
"type": "table",
|
| 923 |
+
"img_path": "images/68a0660a8f4ccf5101315b3ecceddb50a80803f5160b8a92ee87fbbeee30e9c9.jpg",
|
| 924 |
+
"table_caption": [
|
| 925 |
+
"Table 6: Transfer Learning on LVIS dataset using Mask R-CNN with R50-FPN. "
|
| 926 |
+
],
|
| 927 |
+
"table_footnote": [],
|
| 928 |
+
"table_body": "<table><tr><td rowspan=\"2\">Method</td><td rowspan=\"2\">Epoch</td><td colspan=\"5\">1× Schedule</td><td rowspan=\"2\"></td><td colspan=\"6\">2× Schedule</td></tr><tr><td>Apbb</td><td>AP</td><td>AP</td><td>APmk</td><td>AP</td><td>AP7</td><td>Apbb AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>AP</td></tr><tr><td>Supervised</td><td>90</td><td>20.4</td><td>32.9</td><td>21.7</td><td>19.4</td><td>30.6</td><td>20.5</td><td>23.4</td><td>36.9</td><td>24.9</td><td>22.3</td><td>34.7</td><td>23.5</td></tr><tr><td>SoCo*</td><td>400</td><td>26.3</td><td>41.2</td><td>27.8</td><td>25.0</td><td>38.5</td><td>26.8</td><td>28.3</td><td>43.5</td><td>30.7</td><td>26.9</td><td>41.1</td><td>28.7</td></tr></table>",
|
| 929 |
+
"bbox": [
|
| 930 |
+
173,
|
| 931 |
+
111,
|
| 932 |
+
823,
|
| 933 |
+
175
|
| 934 |
+
],
|
| 935 |
+
"page_idx": 8
|
| 936 |
+
},
|
| 937 |
+
{
|
| 938 |
+
"type": "table",
|
| 939 |
+
"img_path": "images/749422280104aeca60c7972e4a5777c932a1f19d1240faf832321a134f8a0dd1.jpg",
|
| 940 |
+
"table_caption": [
|
| 941 |
+
"Table 7: Transfer learning on RetinaNet and FCOS. "
|
| 942 |
+
],
|
| 943 |
+
"table_footnote": [],
|
| 944 |
+
"table_body": "<table><tr><td>Method</td><td>Epoch</td><td>Apbb</td><td>AP8</td><td>AP</td></tr><tr><td>Supervised</td><td>90</td><td>36.3</td><td>55.3</td><td>38.6</td></tr><tr><td>SoCo* 一</td><td>400</td><td>38.3</td><td>57.2</td><td>41.2</td></tr></table>",
|
| 945 |
+
"bbox": [
|
| 946 |
+
173,
|
| 947 |
+
222,
|
| 948 |
+
485,
|
| 949 |
+
287
|
| 950 |
+
],
|
| 951 |
+
"page_idx": 8
|
| 952 |
+
},
|
| 953 |
+
{
|
| 954 |
+
"type": "table",
|
| 955 |
+
"img_path": "images/56b5dfd353782043229bd9a09a3f5c91c66a035e0b47ee3c1065a650186189e4.jpg",
|
| 956 |
+
"table_caption": [
|
| 957 |
+
"(b) Transfer to FCOS. "
|
| 958 |
+
],
|
| 959 |
+
"table_footnote": [],
|
| 960 |
+
"table_body": "<table><tr><td>Method</td><td>Epoch</td><td>Apbb</td><td>AP8</td><td>AP9</td></tr><tr><td>Supervised</td><td>90</td><td>36.6</td><td>56.0</td><td>38.8</td></tr><tr><td>SoCo*</td><td>400</td><td>37.4</td><td>56.3</td><td>39.9</td></tr></table>",
|
| 961 |
+
"bbox": [
|
| 962 |
+
508,
|
| 963 |
+
222,
|
| 964 |
+
820,
|
| 965 |
+
286
|
| 966 |
+
],
|
| 967 |
+
"page_idx": 8
|
| 968 |
+
},
|
| 969 |
+
{
|
| 970 |
+
"type": "text",
|
| 971 |
+
"text": "Table 5(b) ablates the training batch size of SoCo. It can be seen that larger or smaller batch sizes hurt the performance. We use a batch size of 2048 by default. ",
|
| 972 |
+
"bbox": [
|
| 973 |
+
174,
|
| 974 |
+
314,
|
| 975 |
+
821,
|
| 976 |
+
342
|
| 977 |
+
],
|
| 978 |
+
"page_idx": 8
|
| 979 |
+
},
|
| 980 |
+
{
|
| 981 |
+
"type": "text",
|
| 982 |
+
"text": "Table 5(c) ablates the effectiveness of object proposal generation strategies. To demonstrate the superiority of meaningful proposals generated by selective search, we randomly generate $K$ bounding boxes in each of the training images to replace the selective search proposals (namely Random in Table 5(c)). SoCo can still yield satisfactory results using a single random bounding box to learn object-level representations, which is not surprising, since the RoIAlign operation performed on the randomly generated bounding boxes also introduces object-level representation learning. This fact indirectly demonstrates that downstream object detectors can benefit from self-supervised pretraining that involves object-level representations. However, the result is still slightly worse than the case where only one object proposal generated by selective search is used in the pretraining. For the cases where we randomly generate 4 and 8 bounding boxes, the noisy and meaningless proposals cause the pretraining to diverge. From the table, we also find that applying more object proposals $K = 8$ and $K = 1 6$ ) generated by selective search can be harmful. One potential reason is that most of the images in the ImageNet dataset only contain a few objects, so a larger $K$ may create duplicated and redundant proposals. ",
|
| 983 |
+
"bbox": [
|
| 984 |
+
174,
|
| 985 |
+
348,
|
| 986 |
+
825,
|
| 987 |
+
542
|
| 988 |
+
],
|
| 989 |
+
"page_idx": 8
|
| 990 |
+
},
|
| 991 |
+
{
|
| 992 |
+
"type": "text",
|
| 993 |
+
"text": "Table 5(d) ablates the momentum coefficient $\\tau$ of the exponential moving average, and $\\tau = 0 . 9 9$ yields the best performance. ",
|
| 994 |
+
"bbox": [
|
| 995 |
+
176,
|
| 996 |
+
549,
|
| 997 |
+
823,
|
| 998 |
+
577
|
| 999 |
+
],
|
| 1000 |
+
"page_idx": 8
|
| 1001 |
+
},
|
| 1002 |
+
{
|
| 1003 |
+
"type": "text",
|
| 1004 |
+
"text": "4.5 More Experiments ",
|
| 1005 |
+
"text_level": 1,
|
| 1006 |
+
"bbox": [
|
| 1007 |
+
174,
|
| 1008 |
+
594,
|
| 1009 |
+
343,
|
| 1010 |
+
609
|
| 1011 |
+
],
|
| 1012 |
+
"page_idx": 8
|
| 1013 |
+
},
|
| 1014 |
+
{
|
| 1015 |
+
"type": "text",
|
| 1016 |
+
"text": "Transfer Learning on LVIS Dataset. In addition to COCO and PASCAL VOC object detection, we also consider the challenging LVIS v1 dataset [46] to demonstrate the effectiveness and generality of our approach. The LVIS v1 detection benchmark contains 1203 categories in a long-tailed distribution with few training samples. It is considered to be more challenging than the COCO benchmark. We use Mask R-CNN with R50-FPN backbone and follow the standard LVIS $1 \\times$ and $2 \\times$ training schedules. We do not use any training or post-processing tricks. Table 6 shows the results. We can see that our method achieves $+ 5 . 9 \\mathrm { \\ A \\bar { P } ^ { b b } / + \\bar { S } . 6 \\mathrm { \\ A P ^ { m k } } }$ and $+ 4 . 9 \\mathrm { \\ A P ^ { b b } / + 4 . 6 \\ A P ^ { m k } }$ improvements under the LVIS $1 \\times$ and $2 \\times$ training schedules, respectively. ",
|
| 1017 |
+
"bbox": [
|
| 1018 |
+
173,
|
| 1019 |
+
621,
|
| 1020 |
+
825,
|
| 1021 |
+
733
|
| 1022 |
+
],
|
| 1023 |
+
"page_idx": 8
|
| 1024 |
+
},
|
| 1025 |
+
{
|
| 1026 |
+
"type": "text",
|
| 1027 |
+
"text": "Transfer Learning on Different Object Detectors. In addition to two-stage detectors, we further conduct experiments on RetinaNet [47] and FCOS [48], which are representative works for singlestage detectors and anchor-free detectors. We transfer the weights of backbone and FPN layers learned by $\\mathrm { S o C o ^ { * } }$ to RetinaNet and FCOS architectures and follow the standard COCO $1 \\times$ schedule. We use MMDetection [49] as code base and default optimization configs are adopted. Table 7 shows the comparison. ",
|
| 1028 |
+
"bbox": [
|
| 1029 |
+
174,
|
| 1030 |
+
738,
|
| 1031 |
+
825,
|
| 1032 |
+
821
|
| 1033 |
+
],
|
| 1034 |
+
"page_idx": 8
|
| 1035 |
+
},
|
| 1036 |
+
{
|
| 1037 |
+
"type": "text",
|
| 1038 |
+
"text": "Evaluation on Mini COCO. Transfer to the full COCO dataset may be of limited significance due to the extensive supervision available from its large-scale annotated training data. To further demonstrate the generalization ability of SoCo, we conduct an experiment on a mini version of the COCO dataset, named Mini COCO. Concretely, we randomly select $5 \\%$ or $10 \\%$ of the training data from COCO train2017 to form Mini COCO benchmarks. Mask-RCNN with R50-FPN backbone is adopted to evaluate the transfer performance on the COCO $1 \\times$ schedule. COCO val2017 is still used as the evaluation set. The other settings remain unchanged. Table 8 summarizes the results. Compared with the supervised pretraining baseline which achieves $1 9 . 4 \\ : \\mathrm { A P ^ { b b } } / \\ : 1 8 . 3 \\ : \\mathrm { A P ^ { m k } }$ and $2 4 . 7 \\mathrm { A P } ^ { \\mathrm { b b } }$ / 22.9 $\\mathbf { A P } ^ { \\mathrm { m k } }$ on the $5 \\%$ and $10 \\%$ Mini COCO benchmarks, SoCo obtains large improvements of $+ 6 . 6$ $\\mathsf { A P } ^ { \\mathrm { b b } }$ $^ { \\mathrm { b } } / { + 4 . 6 } \\mathrm { A P ^ { \\mathrm { m k } } }$ and $+ 5 . 7$ $\\mathsf { A P } ^ { \\mathrm { b b } }$ / $+ 3 . 9 \\mathrm { A P } ^ { \\mathrm { m k } }$ , respectively. $\\mathrm { S o C o ^ { * } }$ further boosts performance with improvements of $+ 7 . 4 \\mathrm { \\ A P ^ { b b } }$ $' + 5 . 5 \\ : \\mathrm { A P } ^ { \\mathrm { m k } }$ and $+ 6 . 4 \\mathrm { \\ A P ^ { b b } }$ $\\mathrm { ^ { \\prime } { + } 4 . 7 \\ A P ^ { m k } }$ over the supervised pretraining baseline. ",
|
| 1039 |
+
"bbox": [
|
| 1040 |
+
174,
|
| 1041 |
+
827,
|
| 1042 |
+
825,
|
| 1043 |
+
911
|
| 1044 |
+
],
|
| 1045 |
+
"page_idx": 8
|
| 1046 |
+
},
|
| 1047 |
+
{
|
| 1048 |
+
"type": "table",
|
| 1049 |
+
"img_path": "images/bc352656a333ac7b0ce9fa8bb9193bb26f2c1f19d852e1e55184d578615ab345.jpg",
|
| 1050 |
+
"table_caption": [
|
| 1051 |
+
"Table 8: Results on Mini COCO $\\mathbf { 1 } \\times$ schedule. Mask R-CNN with R50-FPN backbone is adopted. "
|
| 1052 |
+
],
|
| 1053 |
+
"table_footnote": [],
|
| 1054 |
+
"table_body": "<table><tr><td rowspan=\"2\">Methods</td><td rowspan=\"2\">Epoch</td><td colspan=\"5\">Mini COCO (5%)</td><td rowspan=\"2\"></td><td colspan=\"5\">Mini C0CO (10%)</td></tr><tr><td>Apbb AP8</td><td>AP9</td><td>APmk</td><td>APk</td><td>APm</td><td>Apbb</td><td>AP8</td><td>AP</td><td>Apmk</td><td>AP APP</td></tr><tr><td>Supervised</td><td>90</td><td>19.4</td><td>36.6</td><td>18.6</td><td>18.3</td><td>33.5</td><td>17.9</td><td>24.7</td><td>43.1</td><td>25.3</td><td>22.9</td><td>40.0</td><td>23.4</td></tr><tr><td>SoCo</td><td>100</td><td>24.6</td><td>42.2</td><td>25.6</td><td>22.1</td><td>38.8</td><td>22.4</td><td>29.2</td><td>47.7</td><td>31.1</td><td>26.1</td><td>44.4</td><td>27.0</td></tr><tr><td>SoCo</td><td>400</td><td>26.0</td><td>43.2</td><td>27.5</td><td>22.9</td><td>40.0</td><td>23.4</td><td>30.4</td><td>48.6</td><td>32.5</td><td>26.8</td><td>45.2</td><td>27.9</td></tr><tr><td>SoCo*</td><td>400</td><td>26.8</td><td>45.0</td><td>28.3</td><td>23.8</td><td>41.4</td><td>24.2</td><td>31.1</td><td>49.9</td><td>33.4</td><td>27.6</td><td>46.4</td><td>28.9</td></tr></table>",
|
| 1055 |
+
"bbox": [
|
| 1056 |
+
173,
|
| 1057 |
+
111,
|
| 1058 |
+
825,
|
| 1059 |
+
196
|
| 1060 |
+
],
|
| 1061 |
+
"page_idx": 9
|
| 1062 |
+
},
|
| 1063 |
+
{
|
| 1064 |
+
"type": "table",
|
| 1065 |
+
"img_path": "images/0a19044b5ecb88e6338dc7fe05507570ee87ea3fd08b922631b26bf2a7d7a7e4.jpg",
|
| 1066 |
+
"table_caption": [
|
| 1067 |
+
"Table 9: Pretraining on non-object-centric dataset. "
|
| 1068 |
+
],
|
| 1069 |
+
"table_footnote": [],
|
| 1070 |
+
"table_body": "<table><tr><td>Pretraining dataset</td><td>Epoch</td><td>Apbb</td><td>AP</td><td>AP</td><td>Apmk</td><td>AP</td><td>APP</td></tr><tr><td>ImageNet</td><td>100</td><td>42.3</td><td>62.5</td><td>46.5</td><td>37.6</td><td>59.1</td><td>40.5</td></tr><tr><td>ImageNet-subset</td><td>100</td><td>36.9</td><td>56.2</td><td>40.1</td><td>33.2</td><td>53.0</td><td>35.6</td></tr><tr><td>COCO train set + unlabeled set</td><td>100</td><td>37.3</td><td>56.5</td><td>40.6</td><td>33.5</td><td>53.4</td><td>36.0</td></tr><tr><td>COCO train set + unlabeled set</td><td>530</td><td>40.6</td><td>61.1</td><td>44.4</td><td>36.4</td><td>58.1</td><td>38.7</td></tr></table>",
|
| 1071 |
+
"bbox": [
|
| 1072 |
+
191,
|
| 1073 |
+
234,
|
| 1074 |
+
808,
|
| 1075 |
+
320
|
| 1076 |
+
],
|
| 1077 |
+
"page_idx": 9
|
| 1078 |
+
},
|
| 1079 |
+
{
|
| 1080 |
+
"type": "text",
|
| 1081 |
+
"text": "",
|
| 1082 |
+
"bbox": [
|
| 1083 |
+
173,
|
| 1084 |
+
347,
|
| 1085 |
+
825,
|
| 1086 |
+
431
|
| 1087 |
+
],
|
| 1088 |
+
"page_idx": 9
|
| 1089 |
+
},
|
| 1090 |
+
{
|
| 1091 |
+
"type": "text",
|
| 1092 |
+
"text": "Pretraining on Non-object-centric Dataset. We use COCO training set and unlabeled set as pretraining data. Since ImageNet is ${ \\sim } 5 . 3$ times larger than COCO, we pretrain our method on COCO for 100 epochs and prolonged 530 epochs respectively. Furthermore, we also generate a dataset named ImageNet-Subset which contains the same number of images of COCO to verify the impact of different pretraining datasets. Table 9 shows the results. Compared with the ImageNet pretrained model, the COCO pretrained model under 530 epochs is lower by 1.7 AP, possibly because the scale of COCO is smaller than ImageNet. Compared with the ImageNet-Subset pretrained model, the COCO pretrained model under 100 epochs is $+ 0 . 4$ AP higher, which demonstrates the importance of domain alignment between the pretraining dataset and detection dataset. ",
|
| 1093 |
+
"bbox": [
|
| 1094 |
+
173,
|
| 1095 |
+
436,
|
| 1096 |
+
825,
|
| 1097 |
+
563
|
| 1098 |
+
],
|
| 1099 |
+
"page_idx": 9
|
| 1100 |
+
},
|
| 1101 |
+
{
|
| 1102 |
+
"type": "text",
|
| 1103 |
+
"text": "Longer Finetuning. We transfer $\\mathrm { { \\bf S o C o ^ { * } } }$ to Mask R-CNN with R50-FPN under the COCO $4 \\times$ schedule. Table 10 shows the results. Our model continues to improve the performance with the $4 \\times$ finetuning schedule, surpassing the supervised baseline by $+ 2 . 6 \\ : \\mathrm { \\bar { A P } ^ { b b } } / + 1 . \\bar { 9 } \\ : \\mathrm { A P ^ { m k } }$ . ",
|
| 1104 |
+
"bbox": [
|
| 1105 |
+
174,
|
| 1106 |
+
568,
|
| 1107 |
+
825,
|
| 1108 |
+
611
|
| 1109 |
+
],
|
| 1110 |
+
"page_idx": 9
|
| 1111 |
+
},
|
| 1112 |
+
{
|
| 1113 |
+
"type": "table",
|
| 1114 |
+
"img_path": "images/73108f255e627592c1831d2302e1dbd0c8ddb960fd322417d4b0784af7fbd197.jpg",
|
| 1115 |
+
"table_caption": [
|
| 1116 |
+
"Table 10: Finetuning on COCO $4 \\times$ schedule using Mask R-CNN with R50-FPN. "
|
| 1117 |
+
],
|
| 1118 |
+
"table_footnote": [],
|
| 1119 |
+
"table_body": "<table><tr><td>Method</td><td>Epoch</td><td>APbb</td><td>APb 50</td><td>AP</td><td>Apmk</td><td>AP 50</td><td>APP</td></tr><tr><td>Supervised</td><td>90</td><td>41.9</td><td>61.5</td><td>45.4</td><td>37.7</td><td>58.8</td><td>40.5</td></tr><tr><td>SoCo*</td><td>400</td><td>44.5</td><td>64.2</td><td>48.8</td><td>39.6</td><td>61.5</td><td>42.4</td></tr></table>",
|
| 1120 |
+
"bbox": [
|
| 1121 |
+
254,
|
| 1122 |
+
646,
|
| 1123 |
+
741,
|
| 1124 |
+
713
|
| 1125 |
+
],
|
| 1126 |
+
"page_idx": 9
|
| 1127 |
+
},
|
| 1128 |
+
{
|
| 1129 |
+
"type": "text",
|
| 1130 |
+
"text": "5 Conclusion ",
|
| 1131 |
+
"text_level": 1,
|
| 1132 |
+
"bbox": [
|
| 1133 |
+
173,
|
| 1134 |
+
741,
|
| 1135 |
+
299,
|
| 1136 |
+
758
|
| 1137 |
+
],
|
| 1138 |
+
"page_idx": 9
|
| 1139 |
+
},
|
| 1140 |
+
{
|
| 1141 |
+
"type": "text",
|
| 1142 |
+
"text": "In this paper, we propose a novel object-level self-supervised pretraining method named Selective Object COntrastive learning (SoCo), which aims to align pretraining to object detection. Different from prior image-level contrastive learning methods which treat the whole image as an instance, SoCo treats each object proposal generated by the selective search algorithm as an independent instance, enabling SoCo to learn object-level visual representations. Further alignment is obtained in two ways. One is through network alignment between pretraining and downstream object detection, such that all layers of the detectors can be well-initialized. The other is by accounting for important properties of object detection such as object-level translation invariance and scale invariance. SoCo achieves state-of-the-art transfer performance on COCO detection using a Mask R-CNN detector. Experiments on two-stage and single-stage detectors demonstrate the generality and extensibility of SoCo. ",
|
| 1143 |
+
"bbox": [
|
| 1144 |
+
173,
|
| 1145 |
+
772,
|
| 1146 |
+
826,
|
| 1147 |
+
911
|
| 1148 |
+
],
|
| 1149 |
+
"page_idx": 9
|
| 1150 |
+
},
|
| 1151 |
+
{
|
| 1152 |
+
"type": "text",
|
| 1153 |
+
"text": "References ",
|
| 1154 |
+
"text_level": 1,
|
| 1155 |
+
"bbox": [
|
| 1156 |
+
174,
|
| 1157 |
+
90,
|
| 1158 |
+
266,
|
| 1159 |
+
106
|
| 1160 |
+
],
|
| 1161 |
+
"page_idx": 10
|
| 1162 |
+
},
|
| 1163 |
+
{
|
| 1164 |
+
"type": "text",
|
| 1165 |
+
"text": "[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A largescale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009. \n[2] Zhirong Wu, Yuanjun Xiong, X Yu Stella, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. \n[3] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. [4] 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[5] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. \n[6] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning. arXiv preprint arXiv:2005.10243, 2020. [7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. arXiv preprint arXiv:2006.09882, 2020. \n[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. arXiv preprint arXiv:2011.10566, 2020. \n[9] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020. \n[10] Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. arXiv preprint arXiv:2011.10043, 2020. \n[11] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. arXiv preprint arXiv:2011.09157, 2020. \n[12] Ceyuan Yang, Zhirong Wu, Bolei Zhou, and Stephen Lin. Instance localization for selfsupervised detection pretraining. arXiv preprint arXiv:2102.08318, 2021. \n[13] Olivier J Hénaff, Skanda Koppula, Jean-Baptiste Alayrac, Aaron van den Oord, Oriol Vinyals, and João Carreira. Efficient visual pretraining with contrastive detection. arXiv preprint arXiv:2103.10957, 2021. \n[14] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104(2):154–171, 2013. \n[15] Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117–2125, 2017. \n[16] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497, 2015. \n[17] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. \n[18] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, Pierre-Antoine Manzagol, and Léon Bottou. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research, 11(12), 2010. \n[19] Ruslan Salakhutdinov and Geoffrey Hinton. Deep boltzmann machines. In Artificial intelligence and statistics, pages 448–455. PMLR, 2009. \n[20] Gustav Larsson, Michael Maire, and Gregory Shakhnarovich. Colorization as a proxy task for visual understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6874–6883, 2017. \n[21] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In Proceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015. \n[22] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2536–2544, 2016. \n[23] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. \n[24] Olivier Henaff. Data-efficient image recognition with contrastive predictive coding. In International Conference on Machine Learning, pages 4182–4192. PMLR, 2020. \n[25] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Selfsupervised learning via redundancy reduction. arXiv preprint arXiv:2103.03230, 2021. \n[26] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. arXiv preprint arXiv:2104.14294, 2021. \n[27] Mathilde Caron, Piotr Bojanowski, Julien Mairal, and Armand Joulin. Unsupervised pretraining of image features on non-curated data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2959–2968, 2019. \n[28] Priya Goyal, Mathilde Caron, Benjamin Lefaudeux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021. \n[29] Pedro O Pinheiro, Amjad Almahairi, Ryan Y Benmaleck, Florian Golemo, and Aaron Courville. Unsupervised learning of dense visual representations. arXiv preprint arXiv:2011.05499, 2020. \n[30] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 328–335, 2014. \n[31] Zhigang Dai, Bolun Cai, Yugeng Lin, and Junying Chen. Up-detr: Unsupervised pre-training for object detection with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1601–1610, 2021. \n[32] Songtao Liu, Zeming Li, and Jian Sun. Self-emd: Self-supervised object detection without imagenet. arXiv preprint arXiv:2011.13677, 2020. \n[33] Bharat Singh and Larry S. Davis. An analysis of scale invariance in object detection - snip, 2018. \n[34] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014. \n[35] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440–1448, 2015. \n[36] 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, pages 770–778, 2016. \n[37] 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[38] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Icml, 2010. \n[39] Yang You, Igor Gitman, and Boris Ginsburg. Scaling sgd batch size to 32k for imagenet training. arXiv preprint arXiv:1708.03888, 6:12, 2017. \n[40] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. \n[41] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, Lukasz Wesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaiming He. Accurate, large minibatch sgd: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. \n[42] 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. In European conference on computer vision, pages 740–755. Springer, 2014. \n[43] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338, 2010. \n[44] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019. \n[45] Tete Xiao, Colorado J Reed, Xiaolong Wang, Kurt Keutzer, and Trevor Darrell. Region similarity representation learning. arXiv preprint arXiv:2103.12902, 2021. \n[46] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5356–5364, 2019. \n[47] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. \n[48] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627–9636, 2019. \n[49] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. ",
|
| 1166 |
+
"bbox": [
|
| 1167 |
+
173,
|
| 1168 |
+
95,
|
| 1169 |
+
828,
|
| 1170 |
+
919
|
| 1171 |
+
],
|
| 1172 |
+
"page_idx": 10
|
| 1173 |
+
},
|
| 1174 |
+
{
|
| 1175 |
+
"type": "text",
|
| 1176 |
+
"text": "",
|
| 1177 |
+
"bbox": [
|
| 1178 |
+
168,
|
| 1179 |
+
41,
|
| 1180 |
+
828,
|
| 1181 |
+
917
|
| 1182 |
+
],
|
| 1183 |
+
"page_idx": 11
|
| 1184 |
+
},
|
| 1185 |
+
{
|
| 1186 |
+
"type": "text",
|
| 1187 |
+
"text": "",
|
| 1188 |
+
"bbox": [
|
| 1189 |
+
169,
|
| 1190 |
+
88,
|
| 1191 |
+
828,
|
| 1192 |
+
739
|
| 1193 |
+
],
|
| 1194 |
+
"page_idx": 12
|
| 1195 |
+
}
|
| 1196 |
+
]
|
parse/train/8PA2nX9v_r2/8PA2nX9v_r2_middle.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/8PA2nX9v_r2/8PA2nX9v_r2_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/9p2ekP904Rs/9p2ekP904Rs_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:009dfe8454d3edb43e857cbe4c4420329de42653d7e242cf1e9b670dc5ecfa9c
|
| 3 |
+
size 3632966
|
parse/train/9p2ekP904Rs/9p2ekP904Rs_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2ff6fc7fbf7f749df0957a639fbef318ae1a262b0363b1cd0a72347e92c38202
|
| 3 |
+
size 3493219
|
parse/train/9p2ekP904Rs/9p2ekP904Rs_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d43c196ba358ea551cc1bd6bf151721676c0154b17dabe705f6fefe67a697e8d
|
| 3 |
+
size 3639638
|
parse/train/B18WgG-CZ/B18WgG-CZ_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:37c14cfd2e5cba6b5e3cd44ae566fa01fc118087d5688b60ad4164bac1d07888
|
| 3 |
+
size 730993
|
parse/train/B18WgG-CZ/B18WgG-CZ_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:986baddf78b081284c0474ee3bac9c9d8ef5d6776e8b684f82e603d45871395e
|
| 3 |
+
size 571835
|
parse/train/B18WgG-CZ/B18WgG-CZ_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3258f04fb0a1598d4236ca4d3df11c3e555cdbc5f570a7604f653dfa2a6cc2f3
|
| 3 |
+
size 729114
|
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c0efe724dbd76626551c01619b9b4a46b67428d14c132cd332cca04afe61a926
|
| 3 |
+
size 1242178
|
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:5a9daaa6faea990b323867a9098b27f75c88de1d531bd1ad5953ce4962809baa
|
| 3 |
+
size 1102720
|
parse/train/B1Gi6LeRZ/B1Gi6LeRZ_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a958fd299c40f773facd9b91d891d1244adef391f8e018eeadd6684412f66c5e
|
| 3 |
+
size 1246312
|
parse/train/B1eY_pVYvB/B1eY_pVYvB_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3388075e890bae286193a27f75d607ba5a042f70b9ee618157bcfdc876a67211
|
| 3 |
+
size 5983196
|
parse/train/B1eY_pVYvB/B1eY_pVYvB_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2ba3ff2a44ff66b3d00d9f97f19adfc3f2045c7f6dc02af066547c8f615b856b
|
| 3 |
+
size 5836948
|
parse/train/B1eY_pVYvB/B1eY_pVYvB_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:092b96cb5f723680fe873ac97a61289900d4e30ef164d2b633ee6d644d98eac5
|
| 3 |
+
size 5984199
|
parse/train/B1gjs6EtDr/B1gjs6EtDr_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d00a3c34f9830cea7841e531e4ae19f9c09fca3d27caff3023fe9f5e8afde34c
|
| 3 |
+
size 525818
|
parse/train/B1gjs6EtDr/B1gjs6EtDr_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2cb5b315951c97d8881cb63e568124d7cf896d797a79c2b811603a4e52370ad9
|
| 3 |
+
size 316632
|
parse/train/B1gjs6EtDr/B1gjs6EtDr_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c0ea5304b65e246986cd1aa40175dc90880db15726e42a49bb084e8d23a5f026
|
| 3 |
+
size 530805
|
parse/train/B1l3M64KwB/B1l3M64KwB_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d555fc17d8ef40e6efc2f326b2b46e8c5c111fca4cc03712e71fe6949aa546ad
|
| 3 |
+
size 1190468
|
parse/train/B1l3M64KwB/B1l3M64KwB_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e9a8e54009d0ab518460cb7d9953ec23129afab0df7aa99432c33ec563a277bc
|
| 3 |
+
size 1074148
|
parse/train/B1l3M64KwB/B1l3M64KwB_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b31f043c429e783f692916d22d99914fbebf7ae19786f0a72716de9c5fe41530
|
| 3 |
+
size 1191808
|
parse/train/B1lgUkBFwr/B1lgUkBFwr_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ae0979d75994c6bd5f2e87c5c77515b912d7b193a4bd546dba4a36ce0ca10cd9
|
| 3 |
+
size 3392928
|
parse/train/B1lgUkBFwr/B1lgUkBFwr_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:632b588916d846a2e728d0693ffe7e6dc94e1ab13f7e584cbe30887b51f3ef4d
|
| 3 |
+
size 3132888
|
parse/train/B1xtFpVtvB/B1xtFpVtvB.md
ADDED
|
@@ -0,0 +1,256 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# IMPROVING THE GENERALIZATION OF VISUAL NAVIGATION POLICIES USING INVARIANCE REGULARIZATION
|
| 2 |
+
|
| 3 |
+
Anonymous authors Paper under double-blind review
|
| 4 |
+
|
| 5 |
+
# ABSTRACT
|
| 6 |
+
|
| 7 |
+
Training agents to operate in one environment often yields overfitted models that are unable to generalize to the changes in that environment. However, due to the numerous variations that can occur in the real-world, the agent is often required to be robust in order to be useful. This has not been the case for agents trained with reinforcement learning $( R L )$ algorithms. In this paper, we investigate the overfitting of RL agents to the training environments in visual navigation tasks. Our experiments show that deep RL agents can overfit even when trained on multiple environments simultaneously. We propose a regularization method which combines RL with supervised learning methods by adding a term to the RL objective that would encourage the invariance of a policy to variations in the observations that ought not to affect the action taken. The results of this method, called invariance regularization, show an improvement in the generalization of policies to environments not seen during training.
|
| 8 |
+
|
| 9 |
+
# 1 INTRODUCTION
|
| 10 |
+
|
| 11 |
+
Learning control policies from high-dimensional sensory input has been gaining more traction lately due to the popularity of deep reinforcement learning (DRL) Mnih et al. (2015); Levine et al. (2015); Zhang et al. (2018b); Rakelly et al. (2019), which enables learning the perception and control modules simultaneously. However, most of the work done in RL chooses to evaluate the learned policies in the same environment in which training occurred Cobbe et al. (2018).
|
| 12 |
+
|
| 13 |
+
Using the same environments to train and test agents does not give any insight into the generalization abilities of the learned policy. There could be a number of changes in the environment at test time that would degrade the agent’s performance. Variations could appear in the visual aspects that determine the agent’s observation, the physical structure that determines the agent’s state and even some aspects that are related to the agent’s goal (Figure 1). For example, different observations of the same room are encountered at different times of the day (different lighting conditions). New obstacles could be present. Levels of a game could be different, yet playing a few levels should often be enough to figure out how to play the rest. Such variations might result in a new environment where the control model that defined the training environment has changed. A robust policy should generalize from its experience and perform the same skills in the presence of these variations.
|
| 14 |
+
|
| 15 |
+
DRL agents have been notorious for overfitting to their training environments Cobbe et al. (2018). An agent could have drastically different performance on testing environments even if it manages to maximize the reward during training Zhang et al. (2018a). Supervised learning algorithms have been shown to have some generalization guarantees when adding proper regularization Mohri et al. (2018). However, these guarantees are weakened in reinforcement learning algorithms where the source of the data is not i.i.d.. In order to make use of the progress of DRL algorithms in practice we need policies that are robust to possible changes in the sensory inputs, surrounding structure and even some aspects of the task.
|
| 16 |
+
|
| 17 |
+
In this paper we study the notion of generalization that is appropriate for visual navigation control policies that are learned with DRL. We present: (1) a study of the generalization of visual control policies to certain changes in the underlying dynamical system; (2) an alternative training method that combines DRL with supervised learning, thus using DRL to learn a controller while leveraging the generalization properties of supervised learning. In our experiments we use the VizDoom platform
|
| 18 |
+
|
| 19 |
+
Kempka et al. (2016) which is easily customizable and enables the generation of numerous variants of a given environment.
|
| 20 |
+
|
| 21 |
+
# 2 PRELIMINARIES
|
| 22 |
+
|
| 23 |
+
Visual navigation for mobile robots combines the domains of vision and control. Navigation can be described as finding a suitable and safe path between a starting state and a goal state Bonin-Font et al. (2008). Classical approaches split the problem into a sequence of sub-tasks, such as map construction, localization, planning and path following Bonin-Font et al. (2008). However, each sub-task requires some handengineering that is specific to the environment and task which makes it hard to adapt it to different scenarios without performing some tuning. Deep learning approaches enable the use of highly non-linear classifiers that can adapt their inner representations to learn to robustly solve complicated tasks Goodfellow et al. (2016).
|
| 24 |
+
|
| 25 |
+
In this work, we use reinforcement learning algorithms coupled with deep learning approaches to solve the task of navigating an agent towards
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
Figure 1: The figure shows how environments may differ in their visual aspects, like textures of the surfaces. The textures provide a differentiator for each environment, where without them the environments would have shared the same state space.
|
| 29 |
+
|
| 30 |
+
a goal object using only its visual observations as input. The field of view of the agent is limited, i.e., it does not observe the full environment, and we do not provide an explicit map of the environment to that agent.
|
| 31 |
+
|
| 32 |
+
# 2.1 PROBLEM STATEMENT
|
| 33 |
+
|
| 34 |
+
We model the problem of visual navigation as a partially observed Markov decision process (POMDP) (Spaan, 2012). A POMDP is given by a tuple
|
| 35 |
+
|
| 36 |
+
$$
|
| 37 |
+
\mathcal { P } : = \langle \mathcal { S } , \mathcal { A } , \Omega , R , T , O , P _ { 0 } \rangle ,
|
| 38 |
+
$$
|
| 39 |
+
|
| 40 |
+
where $s$ is the set of states, $\mathcal { A }$ is the set of actions and $\Omega$ is the set of observations, all which are assumed to be finite sets. The reward function is $R : S \times A \to \mathbb { R }$ . The conditional transition probability mass function is $T : \mathcal { S } \times \mathcal { A } \times \mathcal { S } [ 0 , 1 ]$ , with the interpretation that $T ( s , a , s ^ { \prime } ) =$ $\bar { p } ( s _ { t + 1 } = \cdot | s _ { t } = s , a _ { t } = a )$ is the probability that the next state is $s ^ { \prime }$ given that the current state is $s$ and that action $a$ is taken. The conditional observation probability mass function is $O : S \times A \times \Omega [ 0 , 1 ]$ , with the interpretation that $O ( s , a , o ) = p \bar { ( } o _ { t } = o | s _ { t } = s , a _ { t - 1 } = a )$ is the probability of observing $o$ in state $s$ when the last action taken was $a$ , and we allow for a special observation probability $\bar { O ( s , o ) } = p ( o _ { 0 } = o | s _ { 0 } = s )$ when in the initial state $s$ and no action has yet been taken. Finally, $P _ { 0 }$ is the initial state probability mass function, so that $P _ { 0 } ( s ) = p ( s _ { 0 } = s )$ is the probability that the initial state is $s$ .
|
| 41 |
+
|
| 42 |
+
In DRL, we work with a parameterized policy $\pi _ { \boldsymbol { \theta } } ( h , a ) \ : = \ : p _ { \boldsymbol { \theta } } ( a _ { t } \ : = \ : a | h _ { t } \ : = \ : h )$ with parameters $\theta ~ \in ~ \Theta$ , giving the probability of taking action $a$ given observation-action history $h _ { t } \ : =$ $\left( o _ { 0 } , a _ { 0 } , o _ { 1 } , a _ { 1 } , \ldots , a _ { t - 1 } , o _ { t } \right)$ . The objective is to adjust the parameters $\theta$ to attain a high value for the discounted reward
|
| 43 |
+
|
| 44 |
+
$$
|
| 45 |
+
J _ { \mathcal { P } } ( \theta ) : = \mathbb { E } _ { \mathcal { P } } ^ { \pi _ { \theta } } \left[ \sum _ { t = 0 } ^ { \infty } \gamma ^ { t } R ( s _ { t } , a _ { t } ) \right]
|
| 46 |
+
$$
|
| 47 |
+
|
| 48 |
+
with discount factor $\gamma ~ \in ~ [ 0 , 1 )$ . The expectation is over state-observation-action sequences $( s _ { t } , o _ { t } , a _ { t } ) _ { t = 0 } ^ { \infty }$ where the initial state $s _ { 0 }$ is drawn from $P _ { 0 }$ and other elements of such a sequence are drawn from $T , O$ and $\pi _ { \theta }$ respectively (Sutton & Barto, 1998).
|
| 49 |
+
|
| 50 |
+
Many methods for attempting to approximate optimal policies have been proposed. For instance, policy gradient methods perform gradient ascent on estimates of the expected discounted reward. In
|
| 51 |
+
|
| 52 |
+
this work we use the proximal policy optimization (PPO) algorithm, which arguably shows relatively robust performance on a wide range of different tasks (Schulman et al., 2017).
|
| 53 |
+
|
| 54 |
+
# 2.2 FORMALIZING GENERALIZATION
|
| 55 |
+
|
| 56 |
+
As in classification, we wish to learn from a finite training set but still perform well on previouslyunseen examples from a test set. To formalize this, we have a distribution $\mathcal { D }$ over POMDPs, representing multiple environments or tasks, and we sample $n ^ { \mathrm { t r a i n } }$ POMDPs from this distribution $\mathcal { P } _ { 1 } , \mathcal { P } _ { 2 } , \ldots , \mathcal { P } _ { n ^ { \mathrm { t a i n } } }$ . In the context of navigation, these POMDPs might differ in terms of their observation distributions, perhaps representing views of the same environment at different times of day or year, in terms of their transition distributions, perhaps representing maps with different geometries, or in terms of their reward distributions, perhaps corresponding to the specification of different goal states. Given this sample, we then learn a policy $\pi _ { \theta }$ from a finite collection of state-observation-action sequences from these POMDPs. In order to have a meaningful common policy across these POMDPs, we require that they have common state, action and observation spaces $s , A$ and $\Omega$ . By analogy with the notion of generalization risk in classification (Mohri et al., 2018), we say that policy $\pi _ { \theta }$ generalizes well if it attains a high value for the expectation of the discounted reward over the full distribution of POMDPs, which we call the discounted generalization reward, so that $\mathbb { E } _ { \mathcal { P } \sim \mathcal { D } } J _ { \mathcal { P } } ( \boldsymbol { \theta } )$ is high in some sense.This is our own terminology as we did not find a semantically-equivalent term in the literature.
|
| 57 |
+
|
| 58 |
+
It is not hard to see that the discounted generalization reward is actually the discounted reward $J _ { \mathcal { P } ^ { D } } ( \theta )$ of a single larger POMDP $\mathcal { P } ^ { \mathcal { D } }$ , whose state space may however no longer be finite. To see this, let us associate a unique identifier $i ( \mathcal { P } )$ with any POMDP sampled from $\mathcal { D }$ and let ${ \boldsymbol { \mathcal { T } } } ^ { \mathcal { D } }$ be the set of all such unique identifiers. In the large POMDP $\mathcal { P } ^ { \mathcal { D } }$ , the state space is the Cartesian product $\boldsymbol { S } \times \boldsymbol { \mathcal { T } } ^ { \mathcal { D } }$ of the original states and these unique identifiers, but the action and observation spaces are just $\mathcal { A }$ and $\Omega$ . The initial state distribution is obtained by first sampling a POMDP $\mathcal { P } \sim \mathcal { D }$ and then sampling $s _ { \mathrm { e n v } } \sim P _ { 0 } ^ { \mathcal { P } }$ from that POMDP’s initial state distribution. The initial state in the large POMDP is then the concatenation $( s _ { \mathrm { e n v } } , i ( \mathcal { P } ) )$ . Thus one might succinctly state the problem of generalization in POMDPs as follows: given a distribution $\mathcal { D }$ over POMDPs with common state, action and observation spaces and access to a sample of state-observation-action sequences from a sample of POMDPs drawn from $\mathcal { D }$ , choose a policy $\pi _ { \theta }$ that obtains a high value for the discounted reward $J _ { \mathcal { P } ^ { D } } ( \theta )$ .
|
| 59 |
+
|
| 60 |
+
# 3 RELATED WORK
|
| 61 |
+
|
| 62 |
+
Training in synthetic environments enables the simulation of huge amounts of experience in a span of a few hours. Simulations are convenient to use when training reinforcement learning agents that are often highly sample inefficient Sutton & Barto (1998). There is, frequently, a gap between the synthetic world and the real-world, mainly due to the manner in which the simulators depict the real-world dynamics and visual appearances. Often, these simulated worlds capture the richness and noise of the real-world with low-fidelity Tobin et al. (2017). Many have tried to propose transfer learning techniques to bridge the reality gap in order to still make use of fast simulators for training Taylor & Stone (2009).
|
| 63 |
+
|
| 64 |
+
One popular method to bridge the reality gap is by randomizing some aspects of the training environment Sadeghi & Levine (2016). This domain randomization technique has been shown to be successful for the transfer of grasping policies from simulated training environments to the real-world Tobin et al. (2017). However, the learned models resulting from that work are not control policies, but perception modules. Previous work has showed some success in transferring the perception module learned in simulation to the real world, but not the controller.
|
| 65 |
+
|
| 66 |
+
Cobbe et al. (2018) conduct a large scale study on generalization using a new environment, that resembles an arcade game, which they call CoinRun. They experiment by training on different background images and different level structures. They test with different regularization strategies and network architectures finding that the RL agent has a surprising tendency to overfit even to large training sets. Zhang et al. (2018a) reach a similar conclusion, when learning in grid-world environments, and state that the agents have a tendency to memorize levels of the training set. Unlike Cobbe et al. (2018), however, they argue that the methods that inject stochasticity into the dynamics of the system to prevent memorization, such as sticky actions Machado et al. (2017) and random initializations Hausknecht & Stone (2015), often do not help. In our work we are interested in generalization when navigating under partial observability unlike the fully observable CoinRun or grid-world environments.
|
| 67 |
+
|
| 68 |
+
Domain adaptation methods have also been used for simulated to real transfer. They allow models trained on a source domain to generalize to a target domain. Bousmalis et al. (2017) train a generative model to adapt the synthetic images of the simulator to appear like the real environment. It was shown to successfully transfer a grasping policy trained in simulation to the real world. However, they do not discuss whether the policy generalizes when variations happen in the target domain.
|
| 69 |
+
|
| 70 |
+
Another aspect of generalization is the transfer of learned skills to solve different tasks. In other words, generalization to the goal of the trained agent $g$ . Achieving different tasks would require the agent to have the ability to maximize different reward functions. Schaul et al. (2015) consider working with value functions that contain the goal $g$ as part of the agent’s state. They call them universal value functions. The reward will then become a function of a state-action-goal tuple $( s , a , g )$ instead of a classical state-action pair. In the paper, the authors present universal value function approximators (UVFA). A method that attempts to learn a universal value function estimate $V _ { \theta } ( s , g )$ They show that UVFA’s can generalize for unseen state-goal pairs in grid-world setup.
|
| 71 |
+
|
| 72 |
+
Deep reinforcement learning has been used to train control policies. These DRL based methods generally propose to learn motor control commands from raw camera images, thus mapping pixels to commands that control the robot’s motors Levine et al. (2015). DRL algorithms have been used for various navigation tasks such as goal conditioned navigation Mirowski et al. (2016); Zhu et al. (2016) and mapless navigation Mirowski et al. (2018).
|
| 73 |
+
|
| 74 |
+
# 4 GENERALIZATION IN VISUAL CONTROL
|
| 75 |
+
|
| 76 |
+
Control policies learned from high-dimensional visual input are often brittle and lack the robustness to operate in novel situations. Our main contribution is to propose a regularization term that can be added to the RL objective to improve the robustness of the learned policy to variations in the observations, presented in Section 4.2. However, to motivate the necessity of our proposed method we study domain randomization in Section 4.1; one of the current main practices that aims at learning a policy that generalizes well.
|
| 77 |
+
|
| 78 |
+
# 4.1 DOMAIN RANDOMIZATION
|
| 79 |
+
|
| 80 |
+
Domain randomization is typically used to train policies that can generalize to variations and noise in the observations. It is done by training on several POMDP’s that share the same $s , A , \Omega$ spaces, however they could in their observation distribution. The motivation behind domain randomization is that it is assumed to be an effective technique to provide a policy that is invariant to the changes that would appear in the observations. We explore the problem of navigating the agent towards a goal object with random noise added to the agent’s observations. If the agent is able to perform the task in an environment defined by a POMDP $\mathcal { P } _ { 1 }$ then it should still be able to perform the task in another POMDP $\mathcal { P } _ { 2 }$ , if certain features $f$ of the environment that are specific to successfully achieving the task exist and are invariant to these variations, i.e., $f ( \mathcal { P } _ { 1 } ) = f ( \mathcal { P } _ { 2 } )$ .
|
| 81 |
+
|
| 82 |
+
In Section 5.1, we study domain randomization when added to RL training and the ability of resulting policies to generalize in unseen POMDPs. We want to investigate if the policy does in fact overfit to the training POMDPs and whether we mitigate that overfitting by training the policies on multiple POMDPs.
|
| 83 |
+
|
| 84 |
+
# 4.2 INVARIANCE REGULARIZATION (PROPOSED METHOD)
|
| 85 |
+
|
| 86 |
+
In the previous sections, we discussed how overfitting to the training environment can be a big problem in RL. Furthermore, we should be careful not to jump to the conclusion that training on different environments will ensure policies that generalize well to new environments. It is merely an assumption that has been shown to empirically hold up when used in a supervised learning context.
|
| 87 |
+
|
| 88 |
+
However, we show in this work that this assumption might not hold for reinforcement learning techniques. This is compatible with the findings in Cobbe et al. (2018) and Zhang et al. (2018a).
|
| 89 |
+
|
| 90 |
+
We reason that in order to generalize well, the training objective should include a term that encourages policy generalization. Therefore, putting the weight of the problem of generalizing explicitly in the objective function itself. Formally, a function $h$ of variable $x$ is invariant to a transformation $\phi$ of $x$ if $h ( x ) = h ( \phi ( x ) )$ . We can deduce the same definition for the invariance of a policy $\pi$ to changes in the observation given by some transformation $\tau$ , $\pi ( o ) = \pi ( { \mathcal T } ( o ) )$ . We add this regularization penalty term to the RL objective as shown in Equation (1):
|
| 91 |
+
|
| 92 |
+
$$
|
| 93 |
+
\operatorname* { m a x } _ { \theta } \quad \quad L _ { p p o } ( { \cal O } ; \pi _ { \theta } ) - \frac { \lambda } { N M } \sum _ { i = 1 } ^ { N } \sum _ { j = 1 } ^ { M } d ( \pi _ { \theta } ( o _ { i } ) , \pi _ { \theta } ( \mathcal { T } _ { j } ( o _ { i } ) ) ) ,
|
| 94 |
+
$$
|
| 95 |
+
|
| 96 |
+
where $L _ { p p o }$ is the PPO objective Schulman et al. (2017), $\theta$ is the set of parameters that define the policy $\pi _ { \theta }$ , $d$ is a distance function between the two conditional distributions, and $\lambda$ is a weighting coefficient of the penalty. $O = \{ o _ { 1 } , o _ { 2 } , . . . o _ { N } \}$ is a sequence of $N$ observations.
|
| 97 |
+
|
| 98 |
+
$\tau$ is a transformation of the observations. Given an observation $o$ and a transformation on that observation $\tau$ where the transformation still holds the semantic context of the underlying state, but with added visual variations. We can think of the difference between observing a room with observation $o$ and observing the same room with observation $\mathcal { T } ( o )$ as the color of the wall for example. Therefore, let us say that we observe $o$ in POMDP $\mathcal { P }$ and observe $\mathcal { T } ( o )$ in POMDP $\mathcal { P } ^ { \mathcal { T } }$ then ${ \dot { \boldsymbol { f } } } ( { \mathcal { P } } ) = f ( { \mathcal { P } } ^ { \tau } )$ , where $f ( \mathcal P )$ is the set of invariant features of the environment defined by the POMDP $\mathcal { P }$ . We further discuss the nature of $\tau$ in the experiments section. $M$ is the number of transformations of each observations.
|
| 99 |
+
|
| 100 |
+
The penalty $d$ in Equation 1 resembles adding a constraint on the PPO objective, where the new objective dictates that the policy should simultaneously obtain a high reward while behaving similarly for the observations $o$ and $\mathcal { T } ( o )$ . The idea is similar, in spirit, to trust region policy optimization Schulman et al. (2015) where a penalty term, resembling that which would result from imposing a trustregion constraint, is added to ensure monotonic improvement of the average return with each policy update. We call this method in Equation 1 invariance regularization (IR) since the regularization term indicates the invariance of the learned policy to a transformation of given observations.
|
| 101 |
+
|
| 102 |
+
We propose two ways to solve the RL problem in Equation 1. The first is to directly optimize the full objective by adding the penalty to the original PPO loss. The second method splits the training process to two stages of training RL first and then performing a supervised learning step to minimize $d ( \pi ( o ) , \pi ( T ( o ) ) )$ which presents an elegant form that combines reinforcement learning with supervised learning, more details of the second method is available in Appendix A.
|
| 103 |
+
|
| 104 |
+
In the next section, we will discuss experiments using both methods. Before that, we will describe a study on the effectiveness of domain randomization as a mean to reducing overfitting in DRL agents.
|
| 105 |
+
|
| 106 |
+
# 5 EXPERIMENTS
|
| 107 |
+
|
| 108 |
+
In this section we present the results of two experiments. The first is about training RL with domain randomization. We discuss the ability of the learned policies to generalize to unseen environments when trained on variations of the training environment. The next part presents the results obtained when using the invariance regularization (IR) method, proposed in Section 4.2, with domain randomization and shows that it improves the success rate considerably.
|
| 109 |
+
|
| 110 |
+
We performed these experiments because we are interested in the following questions: (1) Does training on environments with random variations (as domain randomization suggests) learn a representation of the invariant $f$ with which the policy can generalize to other environments that share the same invariant features? (2) Can we find a training algorithm that would empirically guarantee finding these invariant features $f$ ?
|
| 111 |
+
|
| 112 |
+
# 5.1 DOMAIN RANDOMIZATION
|
| 113 |
+
|
| 114 |
+
We leverage the customizability of VizDoom maps Kempka et al. (2016) with hundreds of unique textures to generate train/test scenarios. The agent is required to reach an object in order to get a
|
| 115 |
+
|
| 116 |
+
<table><tr><td rowspan=1 colspan=1>Num training envs:</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>10</td><td rowspan=1 colspan=1>50</td><td rowspan=1 colspan=1>100</td><td rowspan=1 colspan=1>500</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=5>PPO</td></tr><tr><td rowspan=1 colspan=1>RGB</td><td rowspan=1 colspan=1>0.21 ± 0.04</td><td rowspan=1 colspan=1>0.17± 0.04</td><td rowspan=1 colspan=1>0.35 ± 0.13</td><td rowspan=1 colspan=1>0.35 ± 0.16</td><td rowspan=1 colspan=1>0.34 ± 0.14</td></tr><tr><td rowspan=1 colspan=1>RGB-D</td><td rowspan=1 colspan=1>0.05±0.04</td><td rowspan=1 colspan=1>0.89 ±0.05</td><td rowspan=1 colspan=1>0.90±0.05</td><td rowspan=1 colspan=1>0.61 ±0.37</td><td rowspan=1 colspan=1>0.77 ± 0.33</td></tr><tr><td rowspan=1 colspan=1>Grayscale</td><td rowspan=1 colspan=1>0.36 ± 0.04</td><td rowspan=1 colspan=1>0.33 ± 0.13</td><td rowspan=1 colspan=1>0.37± 0.04</td><td rowspan=1 colspan=1>0.47±0.14</td><td rowspan=1 colspan=1>0.41 ± 0.22</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=5>PPO-IR (split)</td></tr><tr><td rowspan=1 colspan=1>RGB</td><td rowspan=1 colspan=1></td><td rowspan=1 colspan=1>0.64 ± 0.05</td><td rowspan=1 colspan=1>0.69 ±0.03</td><td rowspan=1 colspan=1>0.72 ± 0.016</td><td rowspan=1 colspan=1>0.75 ± 0.02</td></tr><tr><td rowspan=1 colspan=1>RGB-D</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>0.85 ± 0.02</td><td rowspan=1 colspan=1>0.90±0.05</td><td rowspan=1 colspan=1>0.94±0.01</td><td rowspan=1 colspan=1>0.95± 0.02</td></tr><tr><td rowspan=1 colspan=1>Grayscale</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>0.69±0.01</td><td rowspan=1 colspan=1>0.76±0.02</td><td rowspan=1 colspan=1>0.75 ±0.02</td><td rowspan=1 colspan=1>0.76±0.02</td></tr><tr><td rowspan=1 colspan=1></td><td rowspan=1 colspan=5>PPO-IR (full objective)</td></tr><tr><td rowspan=1 colspan=1>RGB</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>0.79± 0.05</td><td rowspan=1 colspan=1>0.79 ±0.03</td><td rowspan=1 colspan=1>0.81 ±0.03</td><td rowspan=1 colspan=1>0.81±0.02</td></tr><tr><td rowspan=1 colspan=1>RGB-D</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>0.98±0.01</td><td rowspan=1 colspan=1>0.97 ± 0.01</td><td rowspan=1 colspan=1>0.99 ± 0.01</td><td rowspan=1 colspan=1>0.99± 0.01</td></tr><tr><td rowspan=1 colspan=1>Grayscale</td><td rowspan=1 colspan=1>1</td><td rowspan=1 colspan=1>0.79± 0.03</td><td rowspan=1 colspan=1>0.79±0.01</td><td rowspan=1 colspan=1>0.79±0.02</td><td rowspan=1 colspan=1>0.80±0.02</td></tr></table>
|
| 117 |
+
|
| 118 |
+
Table 1: Average success rate and standard deviation of agents, that are trained on a different number of randomly environments, when tested on 50 test environments whose textures are not seen during training. The bold values represent the algorithm that resulted in the best average success rate according to an amount of training environments and an input type. We see that our method brings stability to the average results and improves generalization even when no depth is added.
|
| 119 |
+
|
| 120 |
+
reward. We train an actor-critic style agent Konda & Tsitsiklis (1999) to solve the task. The network consists of three convolutional layers and 2 fully connected layers, followed by the policy and value function estimator layers. The policy output is a four-dimensional fully-connected layer, where the four dimensions corresponds to four actions; move forward, turn right, turn left and do nothing. The ouput of the policy layer is a log-probability of each action. The value layer is a single unit that predicts the value function. This network architecture was proposed by Mnih et al. (2015). ReLUs are used as the non-linear operations in all layers Nair & Hinton (2010). As mentioned, we optimize the PPO objective Schulman et al. (2017) with a binary reward function $( + 1$ if goal is reached, 0 otherwise) and a discount factor $\gamma = 0 . 9 9$ .
|
| 121 |
+
|
| 122 |
+
We generate the variations of the training environment by changing the textures on the surfaces using the numerous textures provided by VizDoom Kempka et al. (2016). We train agents on a subset of 1, 10, 50, 100 and 500 rooms from the generated environments and test on 50 rooms with textures from a hold-out set which are not seen in training. We detail this experimental setup in Appendix B. We experiment with different types of visual input; RGB, RGB-D and Grayscale. The number of training iterations is fixed at $5 \times 1 0 ^ { 6 }$ to ensure repeatability of our experiment. The results are therefore potentially pessimistic, and in future work we would like to choose the number of iterations for each network independently so as to maximize generalization performance. The agent and the goal object are initialized at random positions in the environment at the start of each episode.
|
| 123 |
+
|
| 124 |
+
The role of depth. Adding a depth channel to the observation plays a significant role in generalization. Depth is invariant to many changes in the visible spectrum of the observations. This might lead the training agent to partly find an invariance in observations in its implicit perception model, which in this case can be as simple as focusing on the depth channel only. Therefore, it was not surprising to see, in Table 1, that the depth agents (RGB-D) generalize better than the agents without any depth information.
|
| 125 |
+
|
| 126 |
+
Table 1 shows the success rate of the PPO models with respect to the number of training environments used and the input type (RGB, RGB-D). The results are averaged over 5 seeds, a standard practice in the RL literature today. We notice the superior performance of the agent with depth than the agent without depth. The fact that the RGB agent is not able to generalize well even when exposed to numerous environments tells us that it might not be learning the invariance relating the environments. On the other hand, the RGB-D agents perform well on the testing environments even when the agents are only exposed to 10 random training environments.
|
| 127 |
+
|
| 128 |
+
Looking at the RGB and RGB-D experiments, the agents trained on 100 and 500 environments generalize worse on average than the ones trained on 10 and 50, which indicates that some agents might be overfitting. This is inspite of the fact that these agents are able to maximize the reward in the training set regardless of the set size. Looking at the max statistic of these results (not shown in this paper) the 100 and 500 experiments outperform the rest. However, the 100 and 500 experiments have a higher variance in the success rates of different seeds than the 10 and 50 experiments. High variance in the test results of the 100/500 RGB-D experiments shows that some seeds are able to achieve a near perfect score on the testing environment and others completely fail, thus there is then no empirical guarantee that RL agents will generalize when exposed to numerous environments.
|
| 129 |
+
|
| 130 |
+
The average success rate for the RGB input without the depth shows that domain randomization alone might not be an effective method to adapt the policy to variations in the observations, at least not in the context of RL. In fact, it shows little progress, e.g., the RGB agent exposed to one environment achieves around a $20 \%$ success on the testing environments and the agents exposed to $5 0 +$ environments achieve less than $40 \%$ success. These results are consistent when running with a grayscale channel (see Table 1).
|
| 131 |
+
|
| 132 |
+
While training by randomizing the environment did show some success in making supervised learning models generalize better, it fails to do so in RL policies. It is clear from these results, that adding random variations and relying solely on the RL objective is not enough to ensure generalization. Much of the success of domain randomization in previous works Tobin et al. (2017) was reported using supervised learning. Also, the generalization abilities of machine learning algorithms have been linked to supervised learning setups. Therefore, it would make sense to adapt supervised learning techniques to regularize the models trained with DRL.
|
| 133 |
+
|
| 134 |
+
# 5.2 INVARIANCE REGULARIZATION EXPERIMENTS
|
| 135 |
+
|
| 136 |
+
In this section we will discuss the results obtained from training the agent using the method proposed in Section 4.2. As mentioned in Section 4.2, we propose two methods of using the proposed IR penalty. The first is to add to the PPO objective as in Equation 1, this method is referred to as (full objective) in the results. The second, which is referred to as (split) in the results, is to split the objective into two parts; RL step and a supervised learning step (more details available in Appendix A). The value of $\lambda$ in Equation 1 used in all IR (full objective) experiments is 1.0.
|
| 137 |
+
|
| 138 |
+
As for the nature of transformation $\tau$ of the observations, we tested with the same randomly textured environments from VizDoom, that were used in the previous section, in order to be able to make fair comparisons with the pure RL and domain randomization agents. Regarding the distance penalty term $d$ in Equation 1, we did preliminary experiments with the KL divergence, $L 1$ , $L 2$ and cross-entropy losses and the KL divergence returned the best results. Table 1 shows the results for combining PPO with the IR penalty using the two proposed implementations.
|
| 139 |
+
|
| 140 |
+
Observing the split method’s results, we see that the proposed training procedure returns stable success rates that are improves as more environments are added. The split version was able to outperform vanilla PPO and substantially improve the generalization especially in the cases of RGB/Grayscale inputs. Training with the full objective, however, returned the best results that outperform vanilla PPO with domain randomization and the split version of the IR algorithm. Similar to the split version, training on the full objective shows stable performance for the different inputs across different seeds.
|
| 141 |
+
|
| 142 |
+
The results, in Table 1, also show that the trained models, on the full objective, are achieving test success rate, with only 10 training environments, that is close if not identical to the agents trained on 50, 100 and 500 environments. These results suggests that training with the full objective version of the IR algorithms does not require a large number of environments to learn the invariant features. Notice the average testing success rate is similar across the different number of training environments since the model learns the invariant features from only 10 environments and adding more environments that share the same invariant features will not make a difference. We can verify that hypothesis when looking at the RGB-D testing results in the full objective part. All agents achieve a near perfect score which we attribute to the availability of an invariant feature map in the input (the depth channel) which only the agents trained with the full objective are able to catch.
|
| 143 |
+
|
| 144 |
+
# 5.2.1 COMPARISONS WITH OTHER REGULARIZATION TECHNIQUES
|
| 145 |
+
|
| 146 |
+
Regularization has been shown to help in the generalization of supervised learning models Srivastava et al. (2014). Using regularization in supervised learning often improves the performance of the trained models on test sets. However, regularization has not been frequently used in DRL setups, possibly due to the previously-common poor practise of testing and training in the same environment so there is no generalization gap Cobbe et al. (2018).
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
Figure 2: Different regularization methods tested on the RGB input. (Left): The average success rate show that some of these methods achieve similar results to ours in some instances. (Right): The lower SPL of the other regularization methods relative to ours indicates some randomness in their learned policies
|
| 150 |
+
|
| 151 |
+
We compare our method with some regularization techniques that are frequently used; dropout, batchnorm and $L 2$ . The first experiment has a dropout layer added after each convolutional layer Srivastava et al. (2014), the second has a batchnorm layer added after every convolutional layer Ioffe & Szegedy (2015) and the last uses $L 2$ regularization. We choose the dropout probability to be 0.1 and the $L 2$ weight to be $1 0 ^ { - 4 }$ , the same values that were proposed by Cobbe et al. (2018). As in the previous setup, we train five models (different seeds) for each technique and evaluate on 50 environments whose textures are sampled from a hold-out set. We report the experiments done with RGB input only as it poses a harder problem and a larger gap than RGB-D.
|
| 152 |
+
|
| 153 |
+
Figure 2 (left) shows the average success rate over 5 seeds for the four methods. We see that our proposed method is the only one that is steadily improving when more environments are added. The batchnorm models performed worst while dropout and $L 2$ achieved similar success rates to the split version of our method given 50 and 500 training environments. However, the entropy of the learned policies is substantially higher when dropout and $L 2$ are added to the model.
|
| 154 |
+
|
| 155 |
+
We hypothesize that the high entropy policies are able to generalize by acting randomly in some instances and this makes them more robust in certain situations. We show the success weighted shortest path length $( S P L )$ in Figure 2 (right). A random behavior that displays robustness (has a high success probability) would return a relatively lower SPL due to the fact that this random behavior will probably not take the shortest possible path to the goal. Details of the formulation of SPL are available in Appendix C. Figure 2 (right) shows that the dropout and L2 agents have a lower SPL than the IR agents indicating that these policies with higher entropy are inefficient.
|
| 156 |
+
|
| 157 |
+
# 6 DISCUSSION AND CONCLUSIONS
|
| 158 |
+
|
| 159 |
+
We present a study of the generalization capabilities of visual navigation agents trained with deep reinforcement learning algorithms. We formalize what it means to generalize in the context of a POMDP. We find that the tendency of RL agent to overfit even when exposed to large training sets is quite visible. We show that using domain randomization with RL, without adding invariant features to the input such as the depth maps, is not enough to generalize. In the second part, we proposed Invariance Regularization (IR), a method that attempts to regularize the RL model with a supervised learning loss. It improves the generalization success and displays stable performance across different seeds.
|
| 160 |
+
|
| 161 |
+
In this work, we focused our experimentation on generalization to changes in the input observation. However, it is also interesting to generalize the learned skills to different architectural designs of the environment, just as one one wishes to generalize to different levels of the game as proposed in the retro competition Nichol et al. (2018). Another avenue of future work is to explore the appropriate transformation function $\tau$ of the observations.One might consider an adaptive form of $\tau$ learned with data augmentation Cubuk et al. (2018) or adversarial examples Goodfellow et al. (2015).
|
| 162 |
+
|
| 163 |
+
# REFERENCES
|
| 164 |
+
|
| 165 |
+
Peter Anderson, Angel X. Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir Roshan Zamir. On evaluation of embodied navigation agents. CoRR, abs/1807.06757, 2018. URL http://arxiv.org/abs/1807.06757.
|
| 166 |
+
|
| 167 |
+
Francisco Bonin-Font, Alberto Ortiz, and Gabriel Oliver. Visual navigation for mobile robots: A survey. Journal of Intelligent and Robotic Systems, 53:263–296, 11 2008. doi: 10.1007/ s10846-008-9235-4.
|
| 168 |
+
|
| 169 |
+
Konstantinos Bousmalis, Alex Irpan, Paul Wohlhart, Yunfei Bai, Matthew Kelcey, Mrinal Kalakrishnan, Laura Downs, Julian Ibarz, Peter Pastor, Kurt Konolige, Sergey Levine, and Vincent Vanhoucke. Using simulation and domain adaptation to improve efficiency of deep robotic grasping. CoRR, abs/1709.07857, 2017. URL http://arxiv.org/abs/1709.07857.
|
| 170 |
+
|
| 171 |
+
Karl Cobbe, Oleg Klimov, Christopher Hesse, Taehoon Kim, and John Schulman. Quantifying generalization in reinforcement learning. CoRR, abs/1812.02341, 2018. URL http://arxiv. org/abs/1812.02341.
|
| 172 |
+
|
| 173 |
+
Ekin Dogus Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, and Quoc V. Le. Autoaugment: Learning augmentation policies from data. CoRR, abs/1805.09501, 2018. URL http://arxiv. org/abs/1805.09501.
|
| 174 |
+
|
| 175 |
+
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http: //www.deeplearningbook.org.
|
| 176 |
+
|
| 177 |
+
Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. CoRR, abs/1412.6572, 2015.
|
| 178 |
+
|
| 179 |
+
Matthew Hausknecht and Peter Stone. The impact of determinism on learning Atari 2600 games. In AAAI Workshop on Learning for General Competency in Video Games, January 2015.
|
| 180 |
+
|
| 181 |
+
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. CoRR, abs/1502.03167, 2015. URL http://arxiv.org/ abs/1502.03167.
|
| 182 |
+
|
| 183 |
+
Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. ViZ- ´ Doom: A Doom-based AI research platform for visual reinforcement learning. In IEEE Conference on Computational Intelligence and Games, pp. 341–348, Santorini, Greece, Sep 2016. IEEE. URL http://arxiv.org/abs/1605.02097.
|
| 184 |
+
|
| 185 |
+
Vijay R. Konda and John N. Tsitsiklis. Actor-critic algorithms. In Advances in Neural Information Processing Systems 12, pp. 1008–1014, 1999. URL http://papers.nips.cc/paper/ 1786-actor-critic-algorithms.
|
| 186 |
+
|
| 187 |
+
Sergey Levine, Chelsea Finn, Trevor Darrell, and Pieter Abbeel. End-to-end training of deep visuomotor policies. CoRR, abs/1504.00702, 2015. URL http://arxiv.org/abs/1504. 00702.
|
| 188 |
+
|
| 189 |
+
Marlos C. Machado, Marc G. Bellemare, Erik Talvitie, Joel Veness, Matthew J. Hausknecht, and Michael Bowling. Revisiting the arcade learning environment: Evaluation protocols and open problems for general agents. CoRR, abs/1709.06009, 2017. URL http://arxiv.org/abs/ 1709.06009.
|
| 190 |
+
|
| 191 |
+
Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J. Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, Dharshan Kumaran, and Raia Hadsell. Learning to navigate in complex environments. CoRR, abs/1611.03673, 2016. URL http://arxiv.org/abs/1611.03673.
|
| 192 |
+
|
| 193 |
+
Piotr Mirowski, Matthew Koichi Grimes, Mateusz Malinowski, Karl Moritz Hermann, Keith Anderson, Denis Teplyashin, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, and Raia Hadsell. Learning to navigate in cities without a map. CoRR, abs/1804.00168, 2018. URL http://arxiv.org/abs/1804.00168.
|
| 194 |
+
|
| 195 |
+
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin A. Riedmiller, Andreas Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518:529–533, 2015.
|
| 196 |
+
|
| 197 |
+
Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar. Foundations of Machine Learning. MIT Press, second edition, 2018.
|
| 198 |
+
|
| 199 |
+
Vinod Nair and Geoffrey E. Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on International Conference on Machine Learning, ICML’10, pp. 807–814, USA, 2010. Omnipress. ISBN 978-1-60558-907-7. URL http://dl.acm.org/citation.cfm?id $\equiv$ 3104322.3104425.
|
| 200 |
+
|
| 201 |
+
Alex Nichol, Vicki Pfau, Christopher Hesse, Oleg Klimov, and John Schulman. Gotta learn fast: A new benchmark for generalization in RL. CoRR, abs/1804.03720, 2018. URL http://arxiv. org/abs/1804.03720.
|
| 202 |
+
|
| 203 |
+
Kate Rakelly, Aurick Zhou, Deirdre Quillen, Chelsea Finn, and Sergey Levine. Efficient off-policy meta-reinforcement learning via probabilistic context variables. CoRR, abs/1903.08254, 2019. URL http://arxiv.org/abs/1903.08254.
|
| 204 |
+
|
| 205 |
+
Fereshteh Sadeghi and Sergey Levine. Real single-image flight without a single real image. CoRR, abs/1611.04201, 2016. URL http://arxiv.org/abs/1611.04201.
|
| 206 |
+
|
| 207 |
+
Tom Schaul, Dan Horgan, Karol Gregor, and David Silver. Universal value function approximators. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ICML’15, pp. 1312–1320. JMLR.org, 2015. URL http://dl.acm. org/citation.cfm?id=3045118.3045258.
|
| 208 |
+
|
| 209 |
+
John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region policy optimization. CoRR, abs/1502.05477, 2015. URL http://arxiv.org/abs/1502. 05477.
|
| 210 |
+
|
| 211 |
+
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017. URL http://arxiv.org/abs/ 1707.06347.
|
| 212 |
+
|
| 213 |
+
M.T.J. Spaan. Partially observable Markov decision processes. In Reinforcement Learning, pp. 387–414. Springer, 2012.
|
| 214 |
+
|
| 215 |
+
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15:1929–1958, 2014. URL http://jmlr.org/papers/v15/ srivastava14a.html.
|
| 216 |
+
|
| 217 |
+
Richard S. Sutton and Andrew G. Barto. Introduction to Reinforcement Learning. MIT Press, Cambridge, MA, USA, 1st edition, 1998. ISBN 0262193981.
|
| 218 |
+
|
| 219 |
+
Matthew E. Taylor and Peter Stone. Transfer learning for reinforcement learning domains: A survey. J. Mach. Learn. Res., 10:1633–1685, December 2009. ISSN 1532-4435. URL http: //dl.acm.org/citation.cfm?id=1577069.1755839.
|
| 220 |
+
|
| 221 |
+
Joshua Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. CoRR, abs/1703.06907, 2017. URL http://arxiv.org/abs/1703.06907.
|
| 222 |
+
|
| 223 |
+
Chiyuan Zhang, Oriol Vinyals, Rémi Munos, and Samy Bengio. A study on overfitting in deep reinforcement learning. CoRR, abs/1804.06893, 2018a. URL http://arxiv.org/abs/ 1804.06893.
|
| 224 |
+
|
| 225 |
+
Marvin Zhang, Sharad Vikram, Laura Smith, Pieter Abbeel, Matthew J. Johnson, and Sergey Levine. SOLAR: deep structured latent representations for model-based reinforcement learning. CoRR, abs/1808.09105, 2018b. URL http://arxiv.org/abs/1808.09105.
|
| 226 |
+
|
| 227 |
+
Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J. Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. CoRR, abs/1609.05143, 2016. URL http://arxiv.org/abs/1609.05143.
|
| 228 |
+
|
| 229 |
+
# APPENDIX
|
| 230 |
+
|
| 231 |
+
A IR SPLIT VERSION
|
| 232 |
+
|
| 233 |
+
# Algorithm 1 RL with iterative supervision
|
| 234 |
+
|
| 235 |
+
Initialize $k _ { 1 } , k _ { 2 } , \theta _ { 0 } , \mathcal { T } _ { i = \{ 1 . . . , N \} } , e n v$
|
| 236 |
+
while not converged do for $i = 1 \ldots , k _ { 1 }$ do // Train $\pi _ { \theta }$ on env on the RL objective $\theta _ { i } \gets \operatorname* { m a x } _ { \theta } L _ { p p o } ( o ^ { e n v } ; \pi _ { \theta _ { i - 1 } } )$ end for for $j = 1 \ldots , k _ { 2 }$ do // Train $\pi$ on env and $\tau ( e n v )$ Sample $\{ o _ { t } ^ { e n v } , \pi _ { \theta _ { k _ { 1 } } } ( o _ { t } ^ { e n v } ) \}$ Generate $\{ o _ { t } ^ { \mathcal { T } _ { i } ( e n v ) } , \pi _ { \theta _ { k _ { 1 } } } ( o _ { t } ^ { \mathcal { T } _ { i } ( e n v ) } ) \} ^ { i = 1 \dots N }$ $\begin{array} { r } { \theta _ { j } \operatorname* { m i n } _ { \theta } d ( \pi _ { \theta _ { k _ { 1 } } } ( o _ { e n v } ) | | \pi _ { \theta _ { j - 1 } } ( \mathcal T _ { i } ( o _ { e n v } ) ) } \end{array}$ end for
|
| 237 |
+
end while
|
| 238 |
+
return $\pi _ { \theta }$
|
| 239 |
+
|
| 240 |
+
The first part consists of training RL on the observations of the original training environment, while the second part can be seen as a supervised learning objective on the transformed observations, as shown in Algorithm 1.
|
| 241 |
+
|
| 242 |
+
The first step trains RL on one environment and then use the actions that the trained policy would have taken in that environment to tune the model with supervised learning on the textured environments. In the reported experiments using the split version, the model is trained with one iteration of the algorithm. Therefore, the training process has two stages, train RL then train with a supervised learning setup, without iterating between both.
|
| 243 |
+
|
| 244 |
+
# B EXPERIMENTAL SETUP
|
| 245 |
+
|
| 246 |
+
As stated in Section 5.1, we run training on a subset of 1, 10, 50, 100 and 500 rooms where the surfaces in each room are sampled from the variety of textures available in Vizdoom. The resulting policies are tested on 50 rooms with textures from a hold-out set which are not seen in training. During training we run several agents in parallel to quickly collect observation-action-reward data in multiple environments. Another advantage of this parallelization is the ability to run each agent on a variation of the training environment. Due to hardware limitations, we cannot run one agent for each environment, at least not when we have a large number of training environments, i.e., 100 or 500. Therefore, each agent samples one environment from the training set and runs on it for some $n$ episodes before sampling another one $n = 2 5$ episodes).
|
| 247 |
+
|
| 248 |
+
# C SUCCESS WEIGHTED SHORTEST PATH LENGTH (SPL)
|
| 249 |
+
|
| 250 |
+
SPL was proposed by Anderson et al. (2018) as a way of measuring the navigation agents success rates while taking into account the time it takes agents to succeed 1.
|
| 251 |
+
|
| 252 |
+
$$
|
| 253 |
+
S P L = \frac { 1 } { N } \sum _ { i = 1 } ^ { N } S _ { i } \frac { l _ { i } } { p _ { i } } ,
|
| 254 |
+
$$
|
| 255 |
+
|
| 256 |
+
where $N$ is the number of runs, $S _ { i }$ is the binary indicator of the success of episode $i , l _ { i }$ is the length of the shortest possible path and $p _ { i }$ is the length of the path taken by the agent.
|
parse/train/B1xtFpVtvB/B1xtFpVtvB_model.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
parse/train/BJ6anzb0Z/BJ6anzb0Z_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:929de525f551b8b3b330d3f97413fb3bb0289b0ce7942a312bd077a484eb14a5
|
| 3 |
+
size 4051327
|
parse/train/BJ6anzb0Z/BJ6anzb0Z_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7433e895372442775f6d7af9cb368d92dfe3c96a214c3649c055045e43d99a5d
|
| 3 |
+
size 3953680
|
parse/train/BJ6anzb0Z/BJ6anzb0Z_span.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd30b1fa4632e75d821605a39c9e6578aa08aa10c502a6bd42ece3dc78dbb01e
|
| 3 |
+
size 4046224
|
parse/train/BJuWrGW0Z/BJuWrGW0Z_layout.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:725fc2cb57eb5ba24784c102f5a133f1cd0ce8b1d6c47d5d1d3886dddd54b211
|
| 3 |
+
size 692696
|
parse/train/BJuWrGW0Z/BJuWrGW0Z_origin.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d79e8adbf28b375f1ff73bd225f84062cc8ab92bb430f12f8d565dcfa2bb24de
|
| 3 |
+
size 550952
|