notesum.ai

Published at October 30

S3PT: Scene Semantics and Structure Guided Clustering to Boost Self-Supervised Pre-Training for Autonomous Driving

cs.CV

cs.AI

cs.RO

Released Date: October 30, 2024

Authors: Maciej K. Wozniak¹, Hariprasath Govindarajan², Marvin Klingner¹, Camille Maurice¹, Ravi Kiran³, Senthil Yogamani⁴

Aff.: ¹Qualcomm Technologies International GmbH; ²Arriver Software AB and Linköping University, Sweden; ³Qualcomm SARL, France; ⁴Qualcomm Technologies, Inc

Arxiv: http://arxiv.org/abs/2410.23085v1

Note	Num. clusters	vMF	SK iters.	Depth weight	Linear (mIoU)	Mask Trans. (mIoU)
CrIBo baseline	32	No	5	–	20.50	41.67
+ semantic distribution consistent clustering	32	Yes	5	–	26.07 ^(+5.57)	45.04 ^(+3.37)
+ Object diversity consistent spatial clustering	8	Yes	5	–	26.13	43.40
	16	Yes	5	–	26.18	43.87
	32	Yes	5	–	26.07	45.04
	64	Yes	5	–	24.86	45.85
	128	Yes	5	–	24.15	46.01
	32	Yes	1	–	25.24	45.47
	128	Yes	1	–	28.10 ^(+2.03)	47.46 ^(+2.42)
+ Depth-guided spatial clustering	128	Yes	1	0.5	36.80	51.10
	128	Yes	1	1.0	37.65	51.51 ^(+4.05)
	128	Yes	1	4.0	38.06 ^(+9.96)	51.04