notesum.ai

Published at November 20

BelHouse3D: A Benchmark Dataset for Assessing Occlusion Robustness in 3D Point Cloud Semantic Segmentation

cs.CV

cs.AI

cs.RO

Released Date: November 20, 2024

Authors: Umamaheswaran Raman Kumar¹, Abdur Razzaq Fayjie, Jurgen Hannaert², Patrick Vandewalle¹

Aff.: ¹Dept. of Electrical Engineering (ESAT), KU Leuven, Belgium; ²23Frog, Belgium

Arxiv: http://arxiv.org/abs/2411.13251v1

Refer to caption

		IID			OOD (occlusion)
		mIoU	OA		mIoU	OA
Pointnet [58]		38.38	66.16		26.97 ( $-30\%$ )	68.89 ( $+4\%$ )
Pointnet++ [59]		71.97	82.33		36.56 ( $-49\%$ )	76.06 ( $-8\%$ )
DGCNN [73]		72.57	84.85		40.34 ( $-44\%$ )	76.92 ( $-11\%$ )
Stratified Transformer [40]		79.12	89.12		44.54 ( $-44\%$ )	78.24 ( $-12\%$ )
Point TransformerV2 [74]		82.32	88.44		55.95 ( $-32\%$ )	80.83 ( $-9\%$ )