09-04-2024: C++ implementation of SMAT is available here

07-09-2023: The paper is available on arXiv now

28-08-2023: The pretrained tracker model is released

17-08-2023: The SMAT tracker training and inference code is released

14-08-2023: The paper is accepted at WACV2024

Install the dependency packages using the environment file smat_pyenv.yml.

Generate the relevant files:

python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir ./output

After running this command, modify the datasets paths by editing these files

lib/train/admin/local.py  # paths about training
lib/test/evaluation/local.py  # paths about testing

• Set the path of training datasets in lib/train/admin/local.py
• Place the pretrained backbone model under the pretrained_models/ folder
• For data preparation, please refer to this
• Uncomment lines 63, 67, and 71 in the base_backbone.py file. Long story short: The code is opitmized for high inference speed, hence some intermediate feature-maps are pre-computed during testing. However, these pre-computations are not feasible during training.
• Run

python tracking/train.py --script mobilevitv2_track --config mobilevitv2_256_128x1_ep300 --save_dir ./output --mode single

• The training logs will be saved under output/logs/ folder

The pretrained tracker model can be found here

• Update the test dataset paths in lib/test/evaluation/local.py
• Place the pretrained tracker model under output/checkpoints/ folder
• Run

python tracking/test.py --tracker_name mobilevitv2_track --tracker_param mobilevitv2_256_128x1_ep300 --dataset got10k_test or trackingnet or lasot

• Change the DEVICE variable between cuda and cpu in the --tracker_param file for GPU and CPU-based inference, respectively
• The raw results will be stored under output/test/ folder

To evaluate the tracker on a sample video, run

python tracking/video_demo.py --tracker_name mobilevitv2_track --tracker_param mobilevitv2_256_128x1_ep300 --videofile *path-to-video-file* --optional_box *bounding-box-annotation*

• We use the Separable Self-Attention Transformer implementation and the pretrained MobileViTv2 backbone from ml-cvnets. Thank you!
• Our training code is built upon OSTrack and PyTracking
• To generate the evaluation metrics for different datasets (except, server-based GOT-10k and TrackingNet), we use the pysot-toolkit

If our work is useful for your research, please consider citing:

@inproceedings{gopal2024separable,
  title={Separable self and mixed attention transformers for efficient object tracking},
  author={Gopal, Goutam Yelluru and Amer, Maria A},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={6708--6717},
  year={2024}
}