DVPE

[IJCAI2024] DVPE: Divided View Position Embedding for Multi-View 3D Object Detection

Introduction

This repository is an official implementation of DVPE.

Getting Started

We provide the documentation to get started:

• Getting Started

Results on NuScenes Val Set

Model	Setting	Pretrain	Epoch	NDS	mAP	Config
DVPE	R50-704x256-900q	R50-nuImage	24	0.5542	0.4547	config
DVPE	R50-704x256-900q	R50-ImageNet	90	0.555	0.4496	config
DVPE	R50-704x256-428q	R50-nuImage	60	0.559	0.4658	config
DVPE	R101-1408x512-900q	R101-nuImage	60	0.6032	0.5206	config

Results on NuScenes Test Set

Model	Setting	Pretrain	Epoch	NDS	mAP
DVPE	V99-1600x640-900q	V99-DD3D	60	0.6445	0.5723

Acknowledgement

Many thanks to these excellent open-source projects:

• StreamPETR
• PETR
• mmdet3d

Citation

If you find this project useful in your research, please cite our paper:

@article{wang2024dvpe,
  title={DVPE: Divided View Position Embedding for Multi-View 3D Object Detection},
  author={Wang, Jiasen and Li, Zhenglin and Sun, Ke and Liu, Xianyuan and Zhou, Yang},
  journal={arXiv preprint arXiv:2407.16955},
  year={2024}
}