Tracking Vision Transformer With Class and Regression Tokens

Paper

Table of Contents

1. Projects path
2. Training
3. Tracking
4. Checkpoint
5. Raw results
6. Acknowledgments
7. Citation

Projects path

Create the file for tracker configuration

python tracking/create_default_local_file.py --workspace_dir . --data_dir <data_dir> --save_dir .

The above script will create local.py files for tarining and testing in the following paths:

lib/train/admin/local.py  # Training
lib/pytracking/evaluation/local.py  # Testing

Please edit them with appropriate paths

Training

For training:

# Multiple GPU
python tracking/train.py --script vit_crt --config baseline --save_dir . --mode multiple --nproc_per_node 4

# Single GPU
python tracking/train.py --script vit_crt --config baseline --save_dir . --mode single

Tracking

• VOT2022-STB

Trax wrapper is in the following directory:

<your_path>/ViTCRTracking/pytracking/vot/vit_crt_exp_stb.py

• OTB/LaSOT/TrackingNet/GOT10K:

python tracking/test.py vit_crt baseline --dataset <dataset_name> --threads N --num_gpus N

Original PySOT and GOT10K toolkit are provided for comparison

Checkpoint

To run the experiments it is possible to use the following checkpoint:

• Google Drive
• OneDrive

Put it in

<your_path>/ViTCRTracking/checkpoints/train/vit_crt/baseline

Raw results

Raw results for each dataset reported in the article are provided to the following link:

• Google Drive
• OneDrive

Specifically they are:

• GOT10K
• TrackingNet
• LaSOT
• OTB100
• NfS
• TC128
• UAV123

Acknowledgments

• Thanks to PyTracking library that allows easily build a tracker.
• Thanks to STARK as starting point for our code

Citation

@article{ViTCRT,
title = {Tracking Vision Transformer With Class and Regression Tokens},
author = {Emanuel Di Nardo and Angelo Ciaramella},
journal = {Information Sciences},
year = {2022},
doi = {https://doi.org/10.1016/j.ins.2022.11.055}
}