notesum.ai

Published at December 6

Towards Flexible 3D Perception: Object-Centric Occupancy Completion Augments 3D Object Detection

cs.CV

cs.AI

Released Date: December 6, 2024

Authors: Chaoda Zheng¹, Feng Wang², Naiyan Wang³, Shuguang Cui⁴, Zhen Li⁴

Aff.: ¹FNii-Shenzhen; ²TuSimple; ³Xiaomi EV; ⁴SSE, CUHK-Shenzhen

Arxiv: http://arxiv.org/pdf/2412.05154v1

Refer to caption

Tracklet Inputs	Method	IoU %	mIoU (track) %	mIoU (box) %
GT track	Baseline	61.35	62.19	63.46
GT track	Ours	69.15	64.05	67.91
GT track + noise	Baseline	50.39	45.21	48.59
	Ours	64.92	60.70	63.78
	Ours-E	69.30	64.11	68.04
FSD track	Baseline	44.28	34.77	42.61
	Ours	62.84	54.12	61.58
	Ours-E	68.38	60.96	67.22
CP track	Baseline	40.45	26.69	37.29
	Ours	57.99	44.94	55.10
	Ours-E	65.80	56.81	64.29
FSDv2 track	Ours (no train)	61.41	47.69	60.76