CVPR25_VisionPAD-青年科学家李镇.pdf

1、VisionPAD:A Vision-Centric Pre-training Paradigm for Autonomous DrivingHaiming Zhang1,2,Wending Zhou1,2,Yiyao Zhu 3,Xu Yan 4,Jiantao Gao 4,Dongfeng Bai 4,Yingjie Cai 4,Bingbing Liu 4,Shuguang Cui 2,1,Zhen Li 2,11 The Future Network of Intelligence Institute,The Chinese University of Hong Kong(Shenzh

2、en),2 School of Science and Engineering,The Chinese University of Hong Kong(Shenzhen),3 HKUST4 Huawei Noahs Ark Lab1Background2Vision-centric 3D Perception Tasks:Inputs:Multi-view camera imagesOutputs:3D bounding boxes(3D object detection),3D semantic occupancy,map segmentationAdvantages:cost-effect

3、iveness,general object representation,suitable for unified modelsMulti-view images3D semantic occupancy prediction3D object detectionMap segmenationChallengesChallenges of Scaling up Vision-centric Perception Models:Lack of large-scale 3D annotations;Time-consuming and high-load to train models with

4、 large parameters;3How to solve?Pre-training is an effective method to rescue.Pre-training in ADContrastiveMAERendering-basedInferior performance and not suitable formulti-view AD data.Time-consuming and all heavily rely onexplicit depth supervision from LiDAR.MotivationCould we design an efficient

5、self-supervised pre-training paradigm solely rely on vision inputs?4First introduce a more efficient anchor-based 3DGSrepresentation in vision-centric perception models;Proposeaphotometricconsistencymoduletoimpose geometric information into the volume featurewithout utilizing the LiDAR sensors;The s

6、elf-supervised volume velocity estimationmodule further enhance the motion cues.VisionPADFramework5Any vision-centric perception models build the volumetric features;An efficient anchor-based 3DGS representation built upon the volume feature by shallow MLPs;Photometric consistency loss and self-supe