Awesome

SFTransT: Learning Spatial-Frequency Transformer for Visual Object Tracking

The official implement of SFTransT. Arxiv, IEEE T-CSVT

TL;DR

SFTransT follows the Siamese matching framework which takes the template and search frame as input. The Swin-Tiny network is adopted as the backbone, and the cross-scale features are fused as embedded features. Then, a Multi-Head Cross-Attention (MHCA) module is used to boost the interactions between the dual features. The output will be fed into our core component Spatial-Frequency Transformer, which models the Gaussian spatial prior and low-/high-frequency feature information simultaneously. More in detail, the GGN is adopted to predict the Gaussian spatial attention which will be added to the self-attention matrix. Then, the GPHA is designed to decompose them into low- and high-pass branches to achieve all-pass information propagation. Finally, the enhanced features will be fed into the classification and regression head for target object tracking.

Installation

1. Create and activate a conda environment

conda create -n sftranst python=3.7
conda activate sftranst

2. Install the necessary packages. Please install them line by line to ensure the success.

conda install -c pytorch pytorch=1.5 torchvision=0.6.1 cudatoolkit=10.2
conda install matplotlib pandas tqdm
pip install opencv-python tb-nightly visdom scikit-image tikzplotlib gdown
conda install cython scipy
sudo apt-get install libturbojpeg
pip install pycocotools jpeg4py
pip install wget yacs
pip install shapely==1.6.4.post2 
pip install timm
pip install einops

3. Add the softlink of datasets into the path './dataset/'

     |--dataset
        |--got10k
        |--lasot
        |--trackingnet
        |--.......

4. Setup Environment.

# Environment settings for ltr. Saved at ltr/admin/local.py
cd SFTransT
python -c "from ltr.admin.environment import create_default_local_file; create_default_local_file()"

Training

1. download pretrained model of Swin-Tiny, and put into the ltr/models/backbone/

2. run commmend

cd SFTransT/ltr
conda activate sftranst
python run_training.py --train_module sftranst  --train_name sftranst_cfa_gpha_mlp  

Test and Eval

1. For UAV, OTB， GOT10k

cd SFTransT/pysot_toolkit
conda activate sftranst
python eval_global.py --cuda 0  --begin 99 --end 100 --interval 1 --folds sftranst_cfa_gpha_mlp --subset test

2. For other datasets, like LaSOT:

python test_global.py --dataset LaSOT --cuda 5 --epoch 300  --win 0.50

Acknowledgement

This is a combination version of the python tracking framework PyTracking and PySOT-Toolkit.
Thanks for the TransT which firstly introduce the Transformer into visual tracking.

Citation

@ARTICLE{tang2022learning,
  author={Tang, Chuanming and Wang, Xiao and Bai, Yuanchao and Wu, Zhe and Zhang, Jianlin and Huang, Yongmei},
  journal={IEEE Transactions on Circuits and Systems for Video Technology}, 
  title={Learning Spatial-Frequency Transformer for Visual Object Tracking}, 
  year={2023},
  doi={10.1109/TCSVT.2023.3249468}}

or

@article{tang2022learning,
  title={Learning Spatial-Frequency Transformer for Visual Object Tracking},
  author={Tang, Chuanming and Wang, Xiao and Bai, Yuanchao and Wu, Zhe and Zhang, Jianlin and Huang, Yongmei},
  journal={arXiv preprint arXiv:2208.08829},
  year={2022}
}