Awesome

SIFTrack

Selective Information Flow for Transformer Tracking

[Models][Raw Results]

Environment

Our experiments are conducted with Ubuntu 18.04.6 and CUDA 11.4 (python=3.8).

Preparation

• Download the pre-trained MAE ViT-Base weights and place the file into the pretrained_models directory under SIFTrack project path.

• Download the training datasets ( GOT-10k, TrackingNet, LaSOT, COCO2017) and testing dataset (UAV123) to your disk.

  Organized directory should look like:
  
    ```
    --GOT10k/
    	|--test
    	|--train
    	|--val
    --TrackingNet/
    	|--TRAIN_0
    	|...
    	|--TEST
    --LaSOT/
    	|--airplane-1
    	|...
    	|--zebra-20
    --COCO/
    	|--annotations
    	|--images
    --UAV123/
    	|--anno
    	|--data_seq
    ```
  

• Run the following command to set paths for this project

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

  After running this command, you can also modify paths by editing these two files

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

Training

• Multiple GPU training

  python tracking/train.py --script siftrack --config vitb_256_mae_ce_all_dataset --save_dir ./output --mode multiple --nproc_per_node 4 --use_wandb 1
  

  Replace --config with the desired model config under experiments/siftrack. We use wandb to record detailed training logs.

• Single GPU training

  python tracking/train.py --script siftrack --config vitb_256_mae_ce_all_dataset --save_dir ./output --mode single --use_wandb 1
  

• For GOT-10k Training, set the --config to vitb_256_mae_ce_got10k.

Evaluation

• Make sure you have prepared the trained model or download the model weights from Google Drive (Put the downloaded weights on $PROJECT_ROOT$/output/checkpoints/train/siftrack)

  • GOT10K-test

    python tracking/test.py siftrack vitb_256_mae_ce_got10k --dataset got10k_test --threads 0 --num_gpus 1
    python lib/test/utils/transform_got10k.py --tracker_name siftrack --cfg_name vitb_256_mae_ce_got10k
    

    Then upload $PROJECT_ROOT$/output/test/tracking_results/siftrack/vitb_256_mae_ce_got10k/got10k_submit.zip to the online evaluation server.

  • TrackingNet

    python tracking/test.py siftrack vitb_256_mae_ce_all_dataset --dataset trackingnet --threads 0 --num_gpus 1
    python lib/test/utils/transform_trackingnet.py --tracker_name siftrack vitb_256_mae_ce_all_dataset
    

    Then upload $PROJECT_ROOT$/output/test/tracking_results/siftrack/vitb_256_mae_ce_all_dataset/trackingnet_submit.zip to the online evaluation server.

• LaSOT

  python tracking/test.py siftrack vitb_256_mae_ce_all_dataset --dataset lasot --threads 0 --num_gpus 1
  

  Then evaluate the raw results using the official MATLAB toolkit.

  • UAV123

    python tracking/test.py siftrack vitb_256_mae_ce_all_dataset --dataset uav --threads 0 --num_gpus 1
    python tracking/analysis_results.py
    

Acknowledgments

• Our work is implemented using OSTrack. Thanks for their excellent open-source works

Citation

If our paper or code contributes to your research, please consider citing our work:

@article{KUGARAJEEVAN2025125381,
title = {Selective Information Flow for Transformer Tracking},
author = {Janani Kugarajeevan and Kokul Thanikasalam and Amirthalingam Ramanan and Subha Fernando},
journal = {Expert Systems with Applications},
volume = {259},
pages = {125381},
year = {2025},
doi = {https://doi.org/10.1016/j.eswa.2024.125381},
url = {https://www.sciencedirect.com/science/article/pii/S0957417424022486}
}