Awesome
<div align="center"> <h1>RayDN</h1> <h1>Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection</h1> </div>
<a href="https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/6549_ECCV_2024_paper.php"><img src="https://img.shields.io/badge/ECCV2024-Paper-<color>"></a>
<video src="figs/RayDN.mp4"></video>
Introduction
This repository is an official implementation of our ECCV 2024 paper Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection. This repository contains Pytorch training code, evaluation code and pre-trained models.
Framework
<div align="center"> <img src="figs/framework.png" width="800"/> </div><br/>Getting Started
Our code is built based on StreamPETR. Please follow StreamPETR to setup enviroment and prepare data step by step.
Training and Inference
You can train the model following:
tools/dist_train.sh projects/configs/RayDN/raydn_eva02_800_bs2_seq_24e.py 8
You can evaluate the detection model following:
tools/dist_test.sh projects/configs/RayDN/raydn_eva02_800_bs2_seq_24e.py work_dirs/raydn_eva02_800_bs2_seq_24e/latest.pth 8 --eval bbox
Results on NuScenes Val Set.
| Model | Setting | Pretrain | Lr Schd | NDS | mAP | Config | Download |
|---|---|---|---|---|---|---|---|
| RayDN | R50 - 428q | NuImg | 60ep | 56.1 | 47.1 | config | ckpt |
| RayDN | EVA02-L - 900q | EVA02 | 24ep | 62.4 | 54.1 | config | ckpt |
Acknowledgements
We thank these great works and open-source codebases: MMDetection3d, StreamPETR, DETR3D, PETR.
Citation
If you find RayDN is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.
@article{liu2024ray,
title={Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D Object Detection},
author={Liu, Feng and Huang, Tengteng and Zhang, Qianjing and Yao, Haotian and Zhang, Chi and Wan, Fang and Ye, Qixiang and Zhou, Yanzhao},
journal={arXiv preprint arXiv:2402.03634},
year={2024}
}