/UDT_pytorch

Primary LanguagePythonMIT LicenseMIT

UDT_pytorch

This repository contains a Python reimplementation of the Unsupervised Deep Tracking

Ning Wang, Yibing Song, Chao Ma, Wengang Zhou, Wei Liu, and Houqiang Li

to appear in CVPR 2019

Acknowledges

The results of this implementation may be slightly different from the original UDT. The results in our paper are obtained using the MatconvNet implementation.

Our baseline method is DCFNet and many parts of the code are from DCFNet_pytorch. For more details, you can refer to DCFNet_pytorch. The main differences are (1) unsupervised data preprocessing (please check the train/dataset/ folder); (2) our training code utilizes forward tracking and backward verification to train the model (please check the train/train_UDT.py file). (3) This implementation is a simplified version, and I will update it later.

Requirements

Requirements for PyTorch 0.4.0 and opencv-python

conda install pytorch torchvision -c pytorch
conda install -c menpo opencv

Training data (VID) and Test dataset (OTB).

Test

cd UDT_pytorch/track 
ln -s /path/to/your/OTB2015 ./dataset/OTB2015
ln -s ./dataset/OTB2015 ./dataset/OTB2013
cd dataset & python gen_otb2013.py
python UDT.py --model ../train/work/checkpoint.pth.tar

Train

  1. Download training data. (ILSVRC2015 VID)

  2. Prepare training data for dataloader.

    cd UDT_pytorch/train/dataset
    python parse_vid.py <VID_path>  # save all vid info in a single json
    python crop_image.py  # crop and generate a json for dataloader
  3. Training. (on multiple GPUs ⚡ ⚡ ⚡ ⚡)

    cd UDT_pytorch/train/
    CUDA_VISIBLE_DEVICES=0,1,2,3 python train_UDT.py
    

Fine-tune hyper-parameter

  1. After training, you can simple test the model with default parameter.

    cd UDT_pytorch/track/
    python UDT.py --model ../train/work/crop_125_2.0/checkpoint.pth.tar
  2. Search a better hyper-parameter.

    CUDA_VISIBLE_DEVICES=0 python tune_otb.py  # run on parallel to speed up searching
    python eval_otb.py OTB2013 * 0 10000

License

Licensed under an MIT license.

Citation

If you find this work useful for your research, please consider citing our work and DCFNet:

@inproceedings{Wang_2019_Unsupervised,
    title={Unsupervised Deep Tracking},
    author={Wang, Ning and Song, Yibing and Ma, Chao and Zhou, Wengang and Liu, Wei and Li, Houqiang},
    booktitle={The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2019}
}

@article{wang17dcfnet,
    Author = {Qiang Wang, Jin Gao, Junliang Xing, Mengdan Zhang, Weiming Hu},
    Title = {DCFNet: Discriminant Correlation Filters Network for Visual Tracking},
    Journal = {arXiv preprint arXiv:1704.04057},
    Year = {2017}
}

Contact

If you have any questions, please feel free to contact wn6149@mail.ustc.edu.cn