Hao Shi · Chengshan Pang · Jiaming Zhang · Kailun Yang · Yuhao Wu · Huajian Ni · Yining Lin · Rainer Stiefelhagen · Kaiwei Wang
CoBEV (Complementary-BEV) is a new end-to-end monocular 3D object detector that integrates geometry-centric depth and semantic-centric height cues to construct robust BEV representations for roadside perception. CoBEV outperforms BEVHeight with a margin of 3.80% / 7.92% /6.09% across Vehicle, Pedestrian and Cyclist on DAIR-V2X-I. In scenarios where camera parameters are directly perturbed by noise affecting focal length, roll, and pitch, CoBEV exhibits an enhanced detection capability, surpassing BEVHeight by an average of 3.93%. Our goal is to enhance the perception range of intelligent vehicles and improve the responsiveness of the overall transportation system by constructing robust features within a unified BEV space.
- Release the pretrained models
Table of Contents
Train CoBEV with 8 GPUs
python [EXP_PATH] --amp_backend native -b 8 --gpus 8
Eval CoBEV with 8 GPUs
python [EXP_PATH] --ckpt_path [CKPT_PATH] -e -b 8 --gpus 8
The pretrained models can be found via the following links:
This project is developed based on the code of the following projects.
If our work is helpful to you, please consider citing us by using the following BibTeX entry:
@article{shi2024cobev,
title={Cobev: Elevating roadside 3d object detection with depth and height complementarity},
author={Shi, Hao and Pang, Chengshan and Zhang, Jiaming and Yang, Kailun and Wu, Yuhao and Ni, Huajian and Lin, Yining and Stiefelhagen, Rainer and Wang, Kaiwei},
journal={IEEE Transactions on Image Processing},
year={2024},
publisher={IEEE}
}