Table 2.

Inference times of the different keypoint extraction methods and segmentation networks on GPU

KPs & Model	KP extr. [s]	Inference [s]	Mesh rec. [s]	Total [s]	Seg. Pts
PointNet (0.48 M par., 1.00 B MACs)
Förstner	0.10 ± 0.03	0.06 ± 0.02	$\underset{\_}{0.95\pm 0.05}$	1.10 ± 0.06	442
CNN	$\underset{\_}{0.20\pm 0.12}$	0.06 ± 0.02	1.62 ± 0.11	1.89 ± 0.17	3777
DGCNN (0.65 M par., 4.36 B MACs)
Förstner	0.10 ± 0.03	$\underset{\_}{0.11\pm 0.02}$	0.92 ± 0.04	$\underset{\_}{1.13\pm 0.05}$	331
CNN	$\underset{\_}{0.20\pm 0.12}$	0.12 ± 0.02	1.53 ± 0.09	1.86 ± 0.16	3283
PointTransformer (7.77 M par., 0.41 B MACs)
Förstner	0.10 ± 0.03	1.18 ± 0.13	0.92 ± 0.03	2.20 ± 0.14	331
CNN	$\underset{\_}{0.20\pm 0.12}$	1.32 ± 0.02	1.53 ± 0.09	3.05 ± 0.15	3272
nnU-Net (31.20 M par., 534.21 B MACs)
	–	2.77 ± 1.121	37.00 ± 15.24	39.77 ± 15.28	8845

¹nnU-Net time measured without test-time augmentation or ensembling

PointNet [4], DGCNN [5], PointTransformer [6], and nnU-Net [2] all segment a different amount of points, influencing the duration of PSR mesh reconstruction. Bold denotes the lowest, underlining the second lowest times