A Robust Convolutional Neural Network for 6D Object Pose Estimation from RGB Image with Distance Regularization Voting Loss

Table 2

The performance on the occlusion LINEMOD dataset for pose estimation based on ADD (-S) scores.

Methods	Ape	Can	Cat	Driller	Duck	Egg box	Glue	Hole puncher	Mean
YOLO6D [3]	2.48	17.48	0.67	7.66	1.14	—	10.08	5.45	6.42
Pix2Pose [33]	22.00	44.70	22.70	44.70	15.00	25.20	32.40	49.50	32.00
PoseCNN [5]	9.60	45.20	0.93	41.40	19.60	22.00	38.50	22.10	24.90
PVNet [8]	15.81	63.30	16.68	65.65	25.24	50.17	49.62	39.67	40.77
DPVL [11]	19.23	69.76	21.06	71.58	34.27	47.32	39.65	45.27	43.52
PDAL + AFAM [12]	25.47	68.20	22.26	68.33	32.61	45.28	49.28	47.51	44.87
[7]	29.58	69.10	21.74	70.10	32.08	47.58	55.24	52.22	47.23
Ours	22.44	73.31	24.23	75.07	38.60	51.43	44.08	50.11	47.40

The bold values given in Table 2 indicate the high performance on the occlusion among the compared methods for pose estimation. It is also required to bold the values of our proposed method from column number 2 to 6 (Can to Egg box).