Table 6. Runtimes for registering a mesh target shape. (Experiment 2).

Exp.	Alg.	Average Runtimes (sec.) by Test Case
		1	2	3	4	5	6	7	8	9
2A	ICP	0.086	0.094	0.109	0.107	0.119	0.115	0.087	0.097	0.095
	IMLP-CP	0.194	0.204	0.228	0.252	0.287	0.263	0.138	0.157	0.118
	IMLP-MD	0.270	0.299	0.323	0.389	0.437	0.401	0.226	0.249	0.219
	IMLP	0.352	0.377	0.407	0.499	0.543	0.569	0.258	0.239	0.202
2B	ICP	0.142	0.147	0.165	0.145	0.16	0.168	0.135	0.142	0.166
	IMLP-CP	0.285	0.286	0.322	0.346	0.351	0.343	0.221	0.222	0.196
	IMLP-MD	0.339	0.372	0.381	0.462	0.496	0.467	0.298	0.311	0.275
	IMLP	0.444	0.454	0.495	0.588	0.638	0.686	0.425	0.480	0.379

Average runtimes of successful registrations from Experiment 2 are reported, where 300 randomized trials were conducted for each test case. Each test case represents a different generative noise model (Table 4) applied to points of the source shape. Results are reported for initial shape misalignments of [15, 30] mm / degrees (Experiment 2A) and [30, 60] mm / degrees (Experiment 2B). The algorithms compared include single-threaded implementations of ICP [1], IMLP, and the two IMLP variants IMLP-CP and IMLP-MD.