lianqing11 / mvc_monodet

MVC-MonoDet

This repo contains the implementations of MVC-MonoDet. Our implementations are built on top of MMDetection3D.

Prerequisite

Please install/build the following package (follow the getting_started):

• install mmcv==1.4.8
• build mmseg in software/
• build mmdet in software/
• build mmdet3d in software/
• build mvc_monodet with python setup.py develop

We provide a simple script to conduct the above envirnment setup. For the details, please check the README.md of each package in software/.

# 1. install MMVC with version of 1.4.8
# you may need to change the cuda version with replace cu111 and torch version with replace torch1.9.0
pip install mmcv-full==1.4.8 -f https://download.openmmlab.com/mmcv/dist/cu111/torch1.9.0/index.html

# 2. build mmdet
cd software/mmdet/
python setup.py develop

# 3. build mmseg
cd ../mmseg/
python setup.py develop

# 4. build mmdet3d
cd ../mmdet3d/
python setup.py develop

# 5. build the DCNv2
cd ../../det3d/models/backbones/DCNv2_t18/
bash install.sh
cd ../../../../

Data

Please follow data_prepartion to prepare the training data.

Training ane evaluation.

1. Training:

We provide the config of the standard supervised and semi-supervised training in configs/centernet/semi

For the semi-supervised training setting, we provide the pre-trained script that first trained using the supervised training config in google drive.

2. Evaluation:

For the evaluation, please follow mmdet3d to evaluate the trained model.

For the model in MVC-MonoDet:

Backbone	mAP (Easy)
Baseline	21.99
MVC-MonoDet (w/ semi-supervised training)	26.85
If you find this repo useful for your research, please consider citing the papers

@inproceedings{
   mvc-monodet,
   title={Semi-Supervised Monocular 3D Object Detection by Multi-View Consistency},
   author={Lian, Qing and Xu, Yanbo and Yao, Weilong and Chen, Yingcong and Zhang, Tong},
   booktitle={ECCV},
   year={2022}
}

Acknowledgement

The codebase is heavily leanred from sevreal open source repo: MMDetection3D, Detr3D, MonoFlex.