This repository is an official implementation of our 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.
Our code is built based on StreamPETR. Please follow StreamPETR to setup enviroment and prepare data step by step.
You can train the model following:
tools/dist_train.sh projects/configs/SelfDN/selfdn_r50_704_bs2_seq_428q_nui_60e.py 8
You can evaluate the detection model following:
tools/dist_test.sh projects/configs/SelfDN/selfdn_r50_704_bs2_seq_428q_nui_60e.py work_dirs/selfdn_r50_704_bs2_seq_428q_nui_60e/latest.pth 8 --eval bbox
| Model | Setting | Pretrain | Lr Schd | NDS | mAP | Config | Download |
|---|---|---|---|---|---|---|---|
| SelfDN | R50 - 428q | NuImg | 24ep | 53.8 | 44.3 | config | ckpt |
| SelfDN | R50 - 428q | NuImg | 60ep | 56.8 | 47.9 | config | ckpt |
| SelfDN | EVA02-L - 900q | EVA02 | 24ep | 62.9 | 55.1 | config | ckpt |
We thank these great works and open-source codebases: MMDetection3d, StreamPETR, DETR3D, PETR, RayDN.