PolarBEVDet: Exploring Polar Representation for Multi-View 3D Object Detection in Bird's-Eye-View

News

• [2025/06/23] We add support for the occupancy prediction task
• [2025/06/23] We add support for Waymo dataset
• [2025/01/23] We release the code and pretrained weights for NuScenes dataset
• [2024/12/04] PolarBEVDet is on Arxiv.

Getting Started

• Installation
• Data Preparation
• Train and Eval

Abstract

Multi-view 3D object detection built upon the Lift-Splat-Shoot (LSS) mechanism provides an economical and deployment-friendly solution for autonomous driving. However, all existing LSS-based methods transform multi-view image features into a Cartesian Bird’s-Eye-View (BEV) representation without considering the non-uniform distribution of image infor- mation. Consequently, this leads to information loss in the near and computational redundancy in the far. Furthermore, these methods struggle to exploit view symmetry, increasing the difficulty of representation learning. In this paper, to fundamentally remove these limitations, we propose to replace the Cartesian BEV representation with the polar BEV representation, which naturally adapts to the image information distribution and effortlessly preserves view symmetry by regular convolution. To achieve this, we elaborately tailor three modules: a polar view transformer to generate the polar BEV representation, a polar temporal fusion module for fusing historical polar BEV features and a polar detection head to predict the polar-parameterized representation of the object. In addition, we design a 2D auxiliary detection head and a spatial attention enhancement module to improve the quality of feature extraction in perspective view and BEV, respectively. Finally, we integrate the above improvements into a novel multi-view 3D object detector, PolarBEVDet. Experiments on nuScenes and Waymo show that PolarBEVDet achieves superior performance, and the polar BEV representation can be seamlessly substituted into different LSS-based detectors with consistent performance improvement.

Benchmark Results

results on nuScenes validation set:

Setting	Pretrain	NDS	MAP	Weights
r50_704x256_24e	ImageNet	53.0	43.2	gdrive
r50_704x256_60e	ImageNet	55.3	45.0	gdrive
r50_704x256_nuImg_60e	nuImg	56.7	46.9	gdrive

results on Waymo validation set:

Setting	Pretrain	mAPL	mAPH	mAP
r101_832x1920_24e	nuImg	43.8	54.9	59.6

Regrettably, we are unable to provide the model weights due to Waymo Dataset License Agreement

results on Occ3D-nuScenes validation set:

Setting	mIoU	RayIoU	Weights
r50_704x256_24e	33.7	39.4	gdrive

Visualization

Acknowledgement

Many thanks to these excellent open-source projects:

• 3D Detection:BEVDet, SOLOFusion, StreamPETR
• Codebase: MMDetection3D

Bibtex

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

@article{yu2024polarbevdet,
  title={PolarBEVDet: Exploring Polar Representation for Multi-View 3D Object Detection in Bird's-Eye-View},
  author={Yu, Zichen and Liu, Quanli and Wang, Wei and Zhang, Liyong and Zhao, Xiaoguang},
  journal={arXiv preprint arXiv:2408.16200},
  year={2024}
}