BEVFormer: a Cutting-edge Baseline for Camera-based Detection

BEVFormer.mp4

BEVFormer: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers, ECCV 2022

Paper in arXiv | Paper in Chinese | OpenPerceptionX

Blog in Chinese | Video Talk and Slides (in Chinese)

BEV Perception Survey | Github repo

News

[2022/6/16]: We added two BEVformer configurations, which require less GPU memory than the base version. Please pull this repo to obtain the latest codes.
[2022/6/13]: We release an initial version of BEVFormer. It achieves a baseline result of 51.7% NDS on nuScenes.
[2022/5/23]: 🚀🚀Built on top of BEVFormer, BEVFormer++, gathering up all best practices in recent SOTAs and our unique modification, ranks 1st on Waymo Open Datast 3D Camera-Only Detection Challenge. We will present BEVFormer++ on CVPR 2022 Autonomous Driving Workshop.
[2022/3/10]: 🚀BEVFormer achieve the SOTA on nuScenes Detection Task with 56.9% NDS (camera-only)!

Abstract

In this work, the authors present a new framework termed BEVFormer, which learns unified BEV representations with spatiotemporal transformers to support multiple autonomous driving perception tasks. In a nutshell, BEVFormer exploits both spatial and temporal information by interacting with spatial and temporal space through predefined grid-shaped BEV queries. To aggregate spatial information, the authors design a spatial cross-attention that each BEV query extracts the spatial features from the regions of interest across camera views. For temporal information, the authors propose a temporal self-attention to recurrently fuse the history BEV information. The proposed approach achieves the new state-of-the-art 56.9% in terms of NDS metric on the nuScenes test set, which is 9.0 points higher than previous best arts and on par with the performance of LiDAR-based baselines.

Methods

Getting Started

Model Zoo

Backbone	Method	Lr Schd	NDS	mAP	memroy	Config	Download
R50	BEVFormer-tiny_fp16	24ep	35.9	25.7	-	config	model/log
R50	BEVFormer-tiny	24ep	35.4	25.2	6500M	config	model/log
R101-DCN	BEVFormer-small	24ep	47.9	37.0	10500M	config	model/log
R101-DCN	BEVFormer-base	24ep	51.7	41.6	28500M	config	model/log

Catalog

Bibtex

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{li2022bevformer,
  title={BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers},
  author={Li, Zhiqi and Wang, Wenhai and Li, Hongyang and Xie, Enze and Sima, Chonghao and Lu, Tong and Qiao, Yu and Dai, Jifeng}
  journal={arXiv preprint arXiv:2203.17270},
  year={2022}
}

Acknowledgement

Many thanks to these excellent open source projects:

timothylimyl/BEVFormer