[WACV'24] ODM3D: Alleviating Foreground Sparsity for Semi-Supervised Monocular 3D Object Detection
Introduction
This repository contains a PyTorch implementation of ODM3D based on the OpenPCDet codebase. ODM3D is a LiDAR-to-RGB cross-modal knowledge distillation framework. It exploits unlabelled data and LiDAR point clouds to boost a monocular 3D object detector, leading to new state-of-the-art monocular detection performance on the KITTI dataset.
KITTI Validation Results
Methods | 3D Car Easy@R40 | 3D Car Mod@R40 | 3D Car Hard@R40 | BEV Car Easy@R40 | BEV Car Mod@R40 | BEV Car Hard@R40 |
---|---|---|---|---|---|---|
CMKD (ECCV'22) | 30.20 | 21.50 | 19.40 | - | - | - |
Mix-Teaching (CSVT'23) | 29.74 | 22.27 | 19.04 | 37.45 | 28.99 | 25.31 |
LPCG (ECCV'22) | 31.15 | 23.42 | 20.60 | - | - | - |
ODM3D (WACV'24) | 35.09 | 23.84 | 20.57 | 41.24 | 30.53 | 25.70 |
Checkpoints
We provide the pre-trained models for reproducing our experimental results. Currently, the provided model has better performance on the Car category and higher training efficiency than the one reported in our paper.
Method | 3D Car Easy@R40 | 3D Car Mod@R40 | 3D Car Hard@R40 | BEV Car Easy@R40 | BEV Car Mod@R40 | BEV Car Hard@R40 | Student | Teacher |
---|---|---|---|---|---|---|---|---|
ODM3D-R50 | 35.32 | 23.96 | 21.16 | 43.37 | 30.82 | 27.07 | model | model |
Installation
You may follow the guide provided in CMKD.
Getting Started
Organisation
You may follow the guide provided in CMKD. Make sure you use this file as the training data instead of the original one, since the former contains pseudo-labels for unlabelled scenes. Image backbone weights (e.g., deeplabv3_resnet50_coco-cd0a2569.pth) shall be automatically downloaded into folder "checkpoints", if not already there, as the training begins.
Training
- First conduct stage-1 training with:
python train_odm3d.py --cfg_file cfgs/kitti_models/ODM3D/odm3d_s1.yaml --pretrained_lidar_model ../checkpoints/scd-teacher-kitti.pth
- Then conduct stage-2 training with:
python train_odm3d.py --cfg_file cfgs/kitti_models/ODM3D/odm3d_s2.yaml --pretrained_lidar_model ../checkpoints/scd-teacher-kitti.pth --pretrained_img_model ${PATH_TO_STAGE1_MODEL}
- To resume training from a previously saved checkpoint, do:
python train_odm3d.py --cfg_file cfgs/kitti_models/ODM3D/odm3d_s1.yaml --pretrained_lidar_model ../checkpoints/scd-teacher-kitti.pth -ckpt ${PATH_TO_CKPT}
python train_odm3d.py --cfg_file cfgs/kitti_models/ODM3D/odm3d_s2.yaml --pretrained_lidar_model ../checkpoints/scd-teacher-kitti.pth --ckpt ${PATH_TO_CKPT}
Evaluation
- Evaluate a pretrained model with:
python test_odm3d.py --cfg_file cfgs/kitti_models/ODM3D/odm3d_s2.yaml --ckpt ${PATH_TO_MODEL}
Citation
If you find our paper helpful to your work, you may cite it as:
@inproceedings{odm3d,
author = {Weijia Zhang and Dongnan Liu and Chao Ma and Weidong Cai},
title = {Alleviating Foreground Sparsity for Semi-Supervised Monocular 3D Object Detection},
booktitle = {WACV},
year = {2024}
}