hqucv / dmtrack

Distractor-Aware Fast Tracking via Dynamic Convolutions and MOT Philosophy

DMTrack

Official implementation of our CVPR2021 paper: Distractor-Aware Fast Tracking via Dynamic Convolutions and MOT Philosophy.

Paper on arXiv: 2104.12041.

Installation

To reproduce our Python environment, please follow command lines below (we use CUDA toolkit 10.0 CUDNN 7.6):

** notice: Please use the specific versions of mmdetection and mmcv module included in this repo. **

git clone https://github.com/hqucv/dmtrack

cd dmtrack

conda env create -f env.yml

conda activate dmtrack

pip install scikit-build

pip install cmake

cd _submodules/mmcv

python setup.py develop

pip install pillow==6.2.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

cd ../mmdetection

python setup.py develop

# installing dependencies for mmdetection
pip install -r requirements.txt

pip install pycocotools

Run Training

(Assuming all datasets are stored in ~/data)

# training for DMTrack-GlobalSearch (with 2 Tian Xp 12G)
CUDA_VISIBLE_DEVICES=0,1 python tools/train_dmtrack.py --config configs/dmtrackGS_dla34_fpn.py --gpus 2

Run Tracking

(Assuming all datasets are stored in ~/data).

python tools/test_dmtrack.py

Metric

LaSOT Testset

Name	Inf. Time	Success	Precision	Download
DMtrack-GlobalSearch	32 FPS	53.0	54.2	Google Drive: link Baidu Yun: link password: 81fh
DMTrack-ReID	31 FPS	57.4	58.0

TODO

• Releasing config and weight for an improved GlobalTrack
• Clear up code for MOT training
• Releasing model for DMTrack-ReID
• More comparisons

Issues

Please report issues in this repo if you have any problems.

Cite

@inproceedings{dmtrack,
  title={Distractor-Aware Fast Tracking via Dynamic Convolutions and MOT Philosophy},
  author={Zhang, Zikai and Zhong, Bineng and Zhang, Shengping and Tang, Zhenjun and Liu, Xin and Zhang, Zhaoxiang},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1024--1033},
  year={2021}
}

Reference

Thanks for these great works!

• GlobalTrack
• FairMOT
• CondInst