Visual Tracking by TridentAlign and Context Embedding (TACT)

Test code for "Visual Tracking by TridentAlign and Context Embedding"

Janghoon Choi, Junseok Kwon, and Kyoung Mu Lee

arXiv paper

Overall Framework

<img src="./_figs/overview.png">

Results on LaSOT test set

<img src="./_figs/plots.png">

• Link to LaSOT dataset
• Raw results available on Google drive

Dependencies

• Ubuntu 18.04
• Python==2.7.17
• numpy==1.16.5
• pytorch==1.3.0
• matplotlib==2.2.4
• opencv==4.1.0.25
• moviepy==1.0.0
• tqdm==4.32.1

Usage

Prerequisites

• Download network weights from Google drive
• Copy network weight files ckpt_res18.tar and ckpt_res50.tar to ckpt/ folder
• Choose between TACT-18 and TACT-50 by modifying the cfgs/cfg_test.py file (default: TACT-50)

To test tracker on LaSOT test set

• Download LaSOT dataset from link
• Modify cfgs/cfg_test.py file to local LaSOTBenchmark folder path
• Run python test_tracker.py

To test tracker on an arbitrary sequence

• Using run_track_seq() function in tracker_batch.py, tracker can run on an arbitrary sequence
• Provide the function with following variables
  • seq_name : name of the given sequence
  • seq_path : path to the given sequence
  • seq_imlist : list of image file names of the given sequence
  • seq_gt : ground truth box annotations of the given sequence (may only contain annotation for initial frame, [x_min,y_min,width,height] format)

Raw results on other datasets

• Link to raw results on Google drive
  • Results for test sets of LaSOT, OxUvA, GOT-10k, TrackingNet

Citation

If you find our work useful for your research, please consider citing the following paper:

@article{choi2020tact,
  title={Visual tracking by tridentalign and context embedding},
  author={Choi, Janghoon and Kwon, Junseok and Lee, Kyoung Mu},
  journal={arXiv preprint arXiv:2007.06887},
  year={2020}
}