SCSAtt Tracker

Efficient visual tracking with stacked channel-spatial attention learning (SCSAtt), IEEE Access 2020. Paper link

Summary

Siamese stacked channel-spatial attention learning for visual tracking where channel attention emphasizes 'what' informative part of the target image has to focus and spatial attention responsible for 'where' the informative part is located. Therefore, combining these two attention modules learn 'what' and 'where' to focus or suppress the target information to locate it efficiently. This process help to improve the overall baseline tracking performance.

This work implemented using python with the PyTorch deep learning framework and performed all experiments on a desktop with Intel(R) Core(TM) i7-8700 CPU @ 3.20 GHz and Nvidia GeForce RTX 2080 Super GPU.

To run this code, please consider the following instructions:

N.B. To re-implement this work, it is recommended to install python 3.6 with the PyTorch >= 1.0.

Training

First, we need to configure the training dataset path from config.py file to start training the network. After completing the training dataset configuratiion, run the train.py file.

Testing

To evaluate the model performance, configure testing benchmark path from config.py file and keep the pre-trained model in the "model" folder. We will use the pre-trained model during testing. Finally, run the test.py file to get the success and precision plots.

N.B. To evaluate the other popoular benchmarks, configure the "experiments" variable from test.py file accordingly: Example: benchmark_path is data/OTB, data/UAV123, data/TC128

    ExperimentOTB('benchmark_path', version=2015), #OR version='tb100'
    ExperimentOTB('benchmark_path', version='tb50'),
    ExperimentVOT('benchmark_path', version=2016),
    ExperimentVOT('benchmark_path', version=2017),
    ExperimentVOT('benchmark_path', version=2018),
    ExperimentTColor128('benchmark_path'),
    ExperimentUAV123('benchmark_path', version='UAV123'),
    ExperimentUAV123('benchmark_path', version='UAV20L'),
    ExperimentGOT10k('benchmark_path', subset='test'),
    ExperimentDTB70('benchmark_path'),
    ExperimentNfS('benchmark_path', fps=30),
    ExperimentNfS('benchmark_path', fps=240)

Performance evaluation results

All of the results are computed and compared based on the official OTB and VOT toolkit (OTB (50/100), VOT (2016~2018)

OTB2015 / OTB50/ TColor128 / UAV123

Dataset	Success Score	Precision Score
OTB100	0.641	0.855
OTB50	0.602	0.822
TC128	0.549	0.744
UAV123	0.547	0.776

OTB50

OTB100

Dependencies/ Prerequisite

Install GOT-10k toolkit using the instructions from original site(recommended)before running this code: GOT-10K Toolkit installation guide

OR follow the simple steps to install:

Install the toolkit using pip (recommended):

pip install --upgrade got10k

Stay up-to-date run the following command:

pip install --upgrade git+https://github.com/got-10k/toolkit.git@master

Acknowledgments

Thanks for their amazing SiamFC work

Citation

If you find useful please cite as,

@ARTICLE{9102303,  author={M. M. {Rahman} and M. {Fiaz} and S. K. {Jung}},  
journal={IEEE Access},   
title={Efficient Visual Tracking With Stacked Channel-Spatial Attention Learning},   
year={2020},  
volume={8}, 
pages={100857-100869}
}