TontonTremblay/DOPER_BOP

DOPER BOP -- Per-Object 6DoF Pose Estimation on BOP Datasets

Per-object DOPER-t (keypoint) and RTMDet-tiny (detector) models for all 168 objects across 9 BOP datasets. Trained on synthetic PBR data only. Supports two evaluation modes:

1. Keypoints-only (GT bounding box) -- upper bound on pose accuracy
2. Detection + Pose (full pipeline, no GT) -- RTMDet detection, DOPER-t keypoints, PnP solve

Demo: Full Pipeline (Detection + Keypoints)

28 per-object RTMDet detectors + 28 per-object DOPER-t keypoint models running on a single HOPE val image. No GT bounding boxes.

Results Summary

Keypoints-Only (GT BBox)

Dataset	Objects	Split	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
lm	15	test	3000	0.9820	0.9931	0.9733	0.9766	0.6166	0.7718
lmo	8	test	1517	0.9655	0.9798	0.9565	0.9786	0.5042	0.7479
tless	30	test_primesense	6900	0.6816	0.7030	0.6662	0.9140	0.2245	0.4098
tudl	3	test	600	0.9635	0.9789	0.9431	0.9587	0.3465	0.6038
ycbv	21	test	4125	0.8671	0.8884	0.8503	0.9177	0.2528	0.4584
hb	33	val_primesense	23120	0.9538	0.9630	0.9451	0.9772	0.5961	0.7718
itodd	28	val	123	0.6352	0.6582	0.6123	0.6237	0.0925	0.1037
icbin	2	test	2250	0.8641	0.8921	0.8372	0.9192	0.1548	0.3391
hope	28	val	920	0.8759	0.8937	0.8581	0.8878	0.3331	0.4424

Detection + Pose (Full Pipeline)

Per-object RTMDet-tiny detectors trained with ~30-40% negative images (cross-dataset).

Dataset	Objects	Split	N GT	Det Rate	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
lm	15	test	3000	97.4%	0.9843	0.9950	0.9768	0.9783	0.6264	0.7813
lmo	8	test	1517	94.5%	0.9701	0.9840	0.9622	0.9801	0.5234	0.7602
tless	30	test_primesense	6900	74.1%	0.7079	0.7292	0.6940	0.9129	0.2581	0.4351
tudl	3	test	600	93.8%	0.9064	0.9251	0.8831	0.9380	0.3032	0.5281
ycbv	21	test	4125	85.0%	0.7650	0.7882	0.7473	0.8695	0.2001	0.3671
itodd	28	val	123	66.7%	0.6369	0.6554	0.6202	0.5718	0.1609	0.1177
icbin	2	test	2250	9.1%	0.9116	0.9414	0.8826	0.8452	0.1506	0.2443
hope	28	val	920	57.8%	0.9199	0.9362	0.9065	0.8856	0.3605	0.4596

Det Rate = fraction of GT instances matched by a detection (IoU > 0.1). AUC metrics computed only over detected+solved instances.

Metrics

Metric	Description
ADD	Average Distance of model points (non-symmetric)
ADD-S	Average Distance of closest model points (symmetric-aware, a.k.a. ADI)
MSSD	Maximum Symmetry-Aware Surface Distance (BOP)
MSPD	Maximum Symmetry-Aware Projection Distance (BOP)
AUC	Area Under the recall-vs-threshold Curve (40 thresholds in (0, 10x diameter])
ARMSSD	Average Recall at BOP thresholds: {0.05, 0.10, ..., 0.50} x diameter
ARMSPD	Average Recall at BOP thresholds: {5, 10, ..., 50} pixels

Method

1. 3D Keypoints: 17 symmetry-aware keypoints per object from BOP meshes + models_info.json
2. Keypoint Training: DOPER-t (CSPNeXt-tiny, 256x256, 300 epochs) on projected keypoints from PBR renders
3. Detector Training: RTMDet-tiny (CSPNeXt-tiny, COCO pretrained, 20 epochs) per object, with ~30-40% cross-dataset negative images
4. Inference: RTMDet detects object bbox -> DOPER-t predicts 2D keypoints -> PnP+RANSAC solves 6DoF pose

Training Data

• BOP core (lm, lmo, tless, tudl, ycbv, hb, itodd, icbin): BOP BlenderProc PBR (train_pbr)
• HOPE: Custom BlenderProc synthetic (~40K images/object)

Symmetry Handling

Type	Strategy
None	Farthest-point sampling (FPS) on mesh surface
Discrete	Keypoints in fundamental domain, replicated under symmetry transforms
Continuous	Axial keypoints + equidistant ring perpendicular to symmetry axis

Per-Object Results (Keypoints-Only, GT BBox)

LM (15 objects, test)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	102.1	200	0.9990	1.0000	0.9986	1.0000	0.8260	0.9965
2	247.5	200	0.9926	0.9995	0.9779	0.9622	0.6405	0.6985
3	167.4	200	0.9106	0.9558	0.9161	0.9319	0.2665	0.4410
4	172.5	200	0.9936	0.9990	0.9922	0.9980	0.6900	0.9295
5	201.4	200	0.9955	0.9986	0.9915	0.9914	0.8490	0.9080
6	154.5	200	1.0000	1.0000	1.0000	1.0000	0.9285	0.9950
7	124.3	200	0.9335	0.9620	0.9018	0.9450	0.1750	0.4435
8	261.5	200	0.9977	0.9995	0.9936	0.9939	0.8770	0.9205
9	109.0	200	0.9749	0.9949	0.9527	0.9672	0.3800	0.6435
10	164.6	200	0.9911	0.9975	0.9875	0.9878	0.6850	0.8915
11	175.9	200	0.9851	0.9965	0.9730	0.9940	0.4825	0.8785
12	145.5	200	0.9734	0.9950	0.9521	0.9456	0.3175	0.5175
13	278.1	200	0.9982	0.9994	0.9970	0.9968	0.8965	0.9090
14	282.6	200	0.9878	0.9991	0.9695	0.9370	0.4830	0.4875
15	212.4	200	0.9972	1.0000	0.9957	0.9989	0.7515	0.9165
mean			0.9820	0.9931	0.9733	0.9766	0.6166	0.7718

LMO (8 objects, test)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	102.1	187	0.9828	0.9880	0.9790	0.9951	0.6567	0.8995
5	201.4	199	0.9935	0.9987	0.9861	0.9864	0.7111	0.8201
6	154.5	196	0.9858	0.9909	0.9811	0.9971	0.6143	0.9276
8	261.5	200	0.9981	0.9996	0.9949	0.9954	0.8350	0.8840
9	109.0	188	0.9608	0.9826	0.9368	0.9654	0.2883	0.5968
10	164.6	191	0.8719	0.9132	0.8753	0.9677	0.2435	0.6429
11	175.9	154	0.9625	0.9735	0.9524	0.9828	0.4422	0.7721
12	145.5	200	0.9689	0.9920	0.9468	0.9389	0.2430	0.4405
mean			0.9655	0.9798	0.9565	0.9786	0.5042	0.7479

T-LESS (30 objects, test_primesense)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	63.5	900	0.7947	0.8240	0.7741	0.9646	0.1030	0.6047
2	66.2	500	0.1058	0.1158	0.0997	0.9357	0.0000	0.3038
3	65.3	400	0.8078	0.8419	0.7850	0.9506	0.0793	0.4778
4	80.7	650	0.8712	0.8968	0.8537	0.9638	0.2065	0.6006
5	108.7	200	0.9694	0.9824	0.9635	0.9778	0.5290	0.7625
6	108.3	100	0.9235	0.9435	0.9133	0.9665	0.2870	0.6640
7	178.6	250	0.8267	0.8514	0.8046	0.8572	0.3076	0.3760
8	217.2	150	0.9177	0.9372	0.9043	0.8987	0.3347	0.4493
9	144.5	250	0.8684	0.9066	0.8286	0.7812	0.0184	0.0160
10	90.2	150	0.9417	0.9542	0.9355	0.9800	0.5987	0.8020
11	76.6	200	0.9243	0.9453	0.9088	0.9610	0.3530	0.6250
12	86.0	150	0.9362	0.9558	0.9187	0.9440	0.3640	0.5140
13	58.1	150	0.8567	0.8872	0.8393	0.9662	0.1493	0.6033
14	71.9	150	0.2485	0.2687	0.2363	0.9418	0.0000	0.3580
15	68.6	150	0.8617	0.8970	0.8390	0.9553	0.0887	0.5220
16	69.2	200	0.0229	0.0271	0.0217	0.9280	0.0000	0.2115
17	112.8	150	0.1552	0.1765	0.1473	0.8477	0.0013	0.0133
18	111.0	150	0.9515	0.9695	0.9283	0.9027	0.3733	0.4573
19	89.1	200	0.0000	0.0000	0.0000	0.8629	0.0000	0.0005
20	98.9	250	0.4074	0.4462	0.3760	0.8706	0.0000	0.0324
21	92.3	200	0.8999	0.9182	0.8811	0.9386	0.4730	0.6105
22	92.3	200	0.9441	0.9588	0.9289	0.9469	0.5480	0.6815
23	142.6	250	0.9696	0.9815	0.9635	0.9632	0.5932	0.7012
24	84.7	200	0.9069	0.9315	0.8935	0.9665	0.3010	0.6300
25	108.8	100	0.9378	0.9470	0.9335	0.9770	0.7100	0.8230
26	108.8	100	0.9030	0.9297	0.8672	0.8755	0.0720	0.1180
27	152.5	100	0.2220	0.2417	0.2095	0.7245	0.0000	0.0030
28	124.8	200	0.2886	0.3205	0.2627	0.8064	0.0000	0.0015
29	134.2	100	0.8878	0.9270	0.8812	0.8858	0.2430	0.3030
30	88.8	150	0.0967	0.1075	0.0882	0.8798	0.0000	0.0293
mean			0.6816	0.7030	0.6662	0.9140	0.2245	0.4098

TUD-L (3 objects, test)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	430.3	200	0.9711	0.9835	0.9545	0.9439	0.3925	0.5605
2	175.7	200	0.9324	0.9547	0.9096	0.9689	0.2365	0.6570
3	352.4	200	0.9870	0.9985	0.9651	0.9633	0.4105	0.5940
mean			0.9635	0.9789	0.9431	0.9587	0.3465	0.6038

YCB-V (21 objects, test)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	172.1	300	0.9468	0.9800	0.9470	0.9587	0.1273	0.5660
2	269.6	225	0.9967	0.9996	0.9943	0.9876	0.6609	0.7849
3	198.4	375	0.9925	0.9975	0.9906	0.9956	0.6568	0.8539
4	120.5	450	0.9518	0.9670	0.9436	0.9778	0.4098	0.7502
5	196.5	150	0.9512	0.9622	0.9373	0.9408	0.3940	0.5693
6	89.8	300	0.7816	0.8251	0.7290	0.8697	0.0230	0.1147
7	142.5	75	0.9780	0.9923	0.9743	0.9943	0.3600	0.8467
8	114.1	75	0.9910	1.0000	0.9863	1.0000	0.5920	0.9480
9	129.5	225	0.8569	0.8846	0.8296	0.9054	0.2316	0.4089
10	197.8	150	0.7247	0.7437	0.6943	0.9063	0.0400	0.2027
11	259.5	225	0.7109	0.7407	0.6770	0.6787	0.0262	0.0236
12	259.6	300	0.7168	0.7312	0.6992	0.8213	0.2130	0.2900
13	161.9	150	0.8455	0.8883	0.8307	0.8267	0.0587	0.0527
14	125.0	150	0.9562	0.9832	0.9252	0.9317	0.0500	0.3567
15	226.2	300	0.9940	0.9972	0.9878	0.9849	0.5987	0.7610
16	237.3	75	0.6860	0.7153	0.6660	0.8560	0.0000	0.0320
17	204.0	75	0.9407	0.9567	0.9337	0.9587	0.1373	0.5240
18	121.4	150	0.9830	0.9893	0.9788	0.9983	0.5080	0.9033
19	174.7	150	0.9033	0.9292	0.8903	0.9232	0.0753	0.4100
20	217.1	150	0.9152	0.9480	0.8945	0.9010	0.1467	0.2180
21	102.9	75	0.3873	0.4257	0.3460	0.8553	0.0000	0.0093
mean			0.8671	0.8884	0.8503	0.9177	0.2528	0.4584

HB (33 objects, val_primesense)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC
1	232.6	680	0.9906	0.9962	0.9803
2	257.4	340	1.0000	1.0000	1.0000
3	166.5	1020	0.9737	0.9834	0.9640
4	179.0	680	0.9849	0.9908	0.9793
5	205.4	680	0.9864	0.9899	0.9825
6	121.4	340	0.9741	0.9968	0.9471
7	263.7	340	0.9984	1.0000	0.9940
8	186.8	680	0.9931	0.9968	0.9903
9	166.6	680	0.9139	0.9348	0.8892
10	180.8	680	0.9338	0.9546	0.9307
11	238.5	340	0.9153	0.9433	0.9260
12	156.9	1360	0.9688	0.9823	0.9537
13	145.3	1020	0.9582	0.9733	0.9455
14	243.7	680	0.9922	0.9965	0.9857
15	113.0	1700	0.9483	0.9611	0.9374
16	101.6	1020	0.9519	0.9679	0.9296
17	132.8	1360	0.9448	0.9571	0.9293
18	211.1	680	0.1994	0.2258	0.1693
19	185.6	680	0.9987	0.9993	0.9982
20	244.8	340	1.0000	1.0000	1.0000
21	212.6	340	0.9979	0.9993	0.9963
22	190.2	1360	0.9874	0.9933	0.9823
23	233.9	1020	0.9987	0.9997	0.9971
24	252.3	340	0.9940	0.9999	0.9889
25	202.9	680	0.9753	0.9853	0.9612
26	183.8	680	0.9785	0.9893	0.9613
27	264.4	340	0.9989	0.9998	0.9983
28	477.5	340	1.0000	1.0000	1.0000
29	198.0	680	0.9720	0.9970	0.9489
30	416.2	340	1.0000	1.0000	1.0000
31	158.0	340	0.9912	0.9971	0.9817
32	201.8	680	0.9610	0.9701	0.9487
33	187.2	680	0.9945	0.9974	0.9916
mean			0.9538	0.9630	0.9451

ITODD (28 objects, val)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	64.1	4	0.9125	0.9437	0.8562	0.5500	0.0000	0.0000
2	51.5	3	0.4833	0.5000	0.4500	0.7167	0.0000	0.0000
3	142.2	4	0.0000	0.0000	0.0000	0.2750	0.0000	0.0000
4	139.4	3	0.2583	0.3000	0.2250	0.3083	0.0000	0.0000
5	158.6	6	0.7167	0.7583	0.6833	0.2500	0.0000	0.0000
6	85.3	5	0.9650	0.9750	0.9650	0.9750	0.1800	0.6000
7	38.5	5	0.0000	0.0200	0.0000	0.8250	0.0000	0.0000
8	68.9	3	0.8417	0.8917	0.7917	0.5083	0.0000	0.0000
9	94.8	3	0.0000	0.0000	0.0000	0.4000	0.0000	0.0000
10	55.7	4	0.9563	0.9688	0.9313	0.8938	0.0000	0.1500
11	140.1	5	0.5950	0.6300	0.5850	0.4550	0.0600	0.0000
12	107.7	4	0.0000	0.0000	0.0000	0.1437	0.0000	0.0000
13	128.1	4	0.9750	0.9875	0.9563	0.7438	0.1750	0.1000
14	102.9	3	0.0000	0.0000	0.0000	0.3917	0.0000	0.0000
15	114.2	3	0.9583	0.9833	0.9167	0.4750	0.0333	0.0000
16	193.1	3	0.9833	1.0000	0.9667	0.5917	0.2333	0.0000
17	77.8	3	0.7333	0.7833	0.6833	0.6167	0.0000	0.0000
18	108.5	3	0.2083	0.2417	0.1917	0.6583	0.0000	0.0000
19	121.4	3	0.7917	0.8417	0.7500	0.5083	0.0000	0.0000
20	122.0	4	0.9812	0.9875	0.9750	0.9062	0.6250	0.4250
21	171.2	3	0.9917	1.0000	0.9833	0.7833	0.5333	0.1333
22	267.5	3	0.9333	0.9667	0.8833	0.3333	0.0000	0.0000
23	56.9	1	1.0000	1.0000	0.9750	0.9750	0.4000	0.7000
24	65.0	6	0.3333	0.3583	0.3292	0.8708	0.0000	0.0000
25	48.5	6	0.9375	0.9625	0.9083	0.7917	0.0000	0.3000
26	66.8	6	0.9792	0.9875	0.9667	0.9458	0.3500	0.4333
27	55.7	5	0.7200	0.7750	0.6700	0.6800	0.0000	0.0000
28	24.1	18	0.5306	0.5681	0.5014	0.8917	0.0000	0.0611
mean			0.6352	0.6582	0.6123	0.6237	0.0925	0.1037

IC-BIN (2 objects, test)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	136.6	1800	0.8560	0.8871	0.8380	0.9422	0.2424	0.4592
2	220.6	450	0.8723	0.8971	0.8364	0.8962	0.0671	0.2191
mean			0.8641	0.8921	0.8372	0.9192	0.1548	0.3391

HOPE (28 objects, val)

ID	Diam (mm)	N GT	ADD-AUC	ADD-S-AUC	MSSD-AUC	MSPD-AUC	ARMSSD	ARMSPD
1	107.9	45	0.8717	0.8867	0.8644	0.9417	0.4178	0.5844
2	153.9	45	0.8994	0.9150	0.8733	0.8478	0.3133	0.4000
3	115.0	45	0.8622	0.8778	0.8533	0.9439	0.4378	0.6200
4	89.6	20	0.9525	0.9713	0.9400	0.9475	0.3050	0.5300
5	98.0	30	0.8350	0.8458	0.8250	0.9417	0.5300	0.6633
6	208.7	25	0.9910	0.9960	0.9880	0.9370	0.5920	0.4880
7	89.6	40	0.8325	0.8663	0.8150	0.9187	0.1250	0.3600
8	115.6	25	0.8990	0.9190	0.8780	0.8470	0.2760	0.3600
9	206.0	25	0.8610	0.8910	0.8210	0.7250	0.2000	0.1880
10	89.7	35	0.6257	0.6493	0.5979	0.8479	0.1571	0.2286
11	153.9	40	0.9238	0.9338	0.9094	0.9256	0.5225	0.6525
12	207.1	50	0.9535	0.9710	0.9310	0.7805	0.3500	0.2800
13	153.4	30	0.9767	0.9917	0.9550	0.8525	0.3767	0.3733
14	204.5	25	0.9660	0.9760	0.9500	0.8920	0.4120	0.3520
15	75.7	35	0.8814	0.9150	0.8521	0.8929	0.0657	0.3943
16	161.5	50	0.9560	0.9660	0.9420	0.8980	0.5120	0.5220
17	205.7	25	0.7150	0.7290	0.6790	0.7060	0.1960	0.1840
18	122.8	30	0.9658	0.9783	0.9458	0.9133	0.4167	0.4467
19	89.2	20	0.9650	0.9800	0.9550	0.9750	0.3600	0.7350
20	89.9	30	0.7575	0.7900	0.7350	0.9050	0.0900	0.3400
21	89.2	55	0.8286	0.8509	0.8141	0.9223	0.2855	0.5655
22	152.4	20	0.7838	0.8137	0.7475	0.8462	0.0650	0.1350
23	151.3	40	0.8206	0.8475	0.7900	0.7162	0.0975	0.1500
24	151.3	10	1.0000	1.0000	0.9975	0.9750	0.8000	0.6500
25	252.8	35	0.9786	0.9836	0.9714	0.9257	0.5286	0.5400
26	107.1	35	0.7529	0.7679	0.7457	0.9364	0.3771	0.5800
27	76.1	35	0.7736	0.7971	0.7636	0.9443	0.1629	0.4886
28	82.9	20	0.8962	0.9150	0.8875	0.9525	0.3550	0.5750
mean			0.8759	0.8937	0.8581	0.8878	0.3331	0.4424

File Structure

{dataset}/obj_{NNNNNN}/
  best_coco_AP_epoch_NNN.pth   # DOPER-t keypoint checkpoint
  keypoints_3d.json             # 17 symmetry-aware 3D keypoints (mm)
  bop_summary.json              # Evaluation metrics
  vis_grid.jpg                  # Qualitative results (GT bbox eval)

Detector checkpoints and BOP submission CSVs are in the bop_submission/ directory.

Usage

from huggingface_hub import hf_hub_download

# Download keypoint model + 3D keypoints for ycbv object 1
ckpt = hf_hub_download("TontonTremblay/DOPER_BOP", "ycbv/obj_000001/best_coco_AP_epoch_200.pth", repo_type="dataset")
kpts = hf_hub_download("TontonTremblay/DOPER_BOP", "ycbv/obj_000001/keypoints_3d.json", repo_type="dataset")

BOP Submission

Pre-computed detection+pose results for all 9 datasets in BOP CSV format:

• bop_submission/doper-t_bop_results_det_v2.zip -- RTMDet + DOPER-t + PnP (no GT bbox)

Citation

If you use these models, please cite the DOPER project and the BOP benchmark.