notesum.ai

Published at November 22

VisionPAD: A Vision-Centric Pre-training Paradigm for Autonomous Driving

cs.CV

cs.RO

Released Date: November 22, 2024

Authors: Haiming Zhang¹, Wending Zhou¹, Yiyao Zhu², Xu Yan³, Jiantao Gao³, Dongfeng Bai³, Yingjie Cai³, Bingbing Liu³, Shuguang Cui⁴, Zhen Li⁴

Aff.: ¹FNii, Shenzhen; ²HKUST; ³Huawei Noah's Ark Lab; ⁴SSE, CUHK-Shenzhen

Arxiv: http://arxiv.org/abs/2411.14716v1

Refer to caption

Methods	Venue	Pre-train Modal	CS	CBGS	NDS (%) $\uparrow$	mAP (%) $\uparrow$
BEVFormer-S [20]	ECCV’22	-		✓	44.8	37.5
SpatialDETR [8]	ECCV’22	-			42.5	35.1
PETR [23]	ECCV’22	-		✓	44.2	37.0
Ego3RT [25]	ECCV’22	-			45.0	37.5
3DPPE [33]	ICCV’23	-		✓	45.8	39.1
BEVFormerV2 [44]	CVPR’23	-			46.7	39.6
CMT-C [40]	ICCV’23	-		✓	46.0	40.6
FCOS3D	ICCVW’21	-			38.4	31.1
UVTR [17]	NeurIPS’22	-			45.0	37.2
UVTR+UniPAD	CVPR’24	C			44.8 ^↓0.2	38.5 ^↑1.3
UVTR+VisionPAD (Ours)	-	C			46.7 ^↑1.7	41.0 ^↑3.8
UVTR [17]	NeurIPS’22	-	✓		48.8	39.2
UVTR+UniPAD	CVPR’24	C	✓		48.6 ^↓0.2	40.5 ^↑0.7
UVTR+UniPAD	CVPR’24	C+L	✓		50.2 ^↑1.4	42.8 ^↑3.6
UVTR+VisionPAD (Ours)	-	C	✓		49.7 ^↑0.9	41.2 ^↑2.0
UVTR+VisionPAD (Ours)	-	C+L	✓		50.4 ^↑1.6	43.1 ^↑3.9