Relative Pose Estimation for RGB-D Human Input Scans via Implicit Function Reconstruction

Table 1

The relative pose estimation evaluation on our dataset. We divide all the data into four categories according to the overlap rate: 0-5%, 10%-30%, 40%-70%, and over 80%, and report the mean error of rotation angle (Rot. ()) and translation (Trans. (m)), respectively.

	Overlap (0-5%)		Overlap (10-30%)		Overlap (40-70%)		Overlap (80%)
	Rot.	Trans.	Rot.	Trans.	Rot.	Trans.	Rot.	Trans.
Super4PCs [7]	126.08	1.715	103.95	1.468	38.99	0.574	4.46	0.084
Greg [10]	124.41	1.679	92.50	1.538	64.47	1.109	18.44	0.325
ICP [9]	165.97	2.252	122.02	1.943	67.86	1.208	27.75	0.540
Greg+ICP	126.72	1.664	103.88	1.461	35.10	0.508	1.27	0.0262
Ours-initial	7.15	0.138	6.54	0.130	5.90	0.106	1.29	0.0259
Ours	6.16	0.127	5.88	0.119	4 .60	0.084	1.15	0.0209