FlorinShum / TransT

Transformer Tracking (CVPR2021)

TransT - Transformer Tracking [CVPR2021]

Official implementation of the TransT (CVPR2021) , including training code and trained models.

Tracker

TransT

[Paper] [Models]

This work presents a attention-based feature fusion network, which effectively combines the template and search region features using attention. Specifically, the proposed method includes an ego-context augment module based on self-attention and a cross-feature augment module based on cross-attention. We present a Transformer tracking (named TransT) method based on the Siamese-like feature extraction backbone, the designed attention-based fusion mechanism, and the classification and regression head.

TransT is a very simple and efficient tracker, without online update module, using the same model and hyparameter for all test sets.

Results

Model	LaSOT AUC (%)	TrackingNet AUC (%)	GOT-10k AO (%)	OTB100 AUC (%)	NFS AUC (%)	UAV123 AUC (%)	Speed	Params
TransT-N2	64.2	80.9	69.9	68.1	65.7	67.0	70fps	16.7M
TransT-N4	64.9	81.4	72.3	69.4	65.7	69.1	50fps	23.0M

Installation

This document contains detailed instructions for installing the necessary dependencied for TransT. The instructions have been tested on Ubuntu 18.04 system.

Install dependencies

• Create and activate a conda environment

  conda create -n transt python=3.7
  conda activate transt

• Install PyTorch

  conda install -c pytorch pytorch=1.5 torchvision=0.6.1 cudatoolkit=10.2

• Install other packages

  conda install matplotlib pandas tqdm
  pip install opencv-python tb-nightly visdom scikit-image tikzplotlib gdown
  conda install cython scipy
  sudo apt-get install libturbojpeg
  pip install pycocotools jpeg4py
  pip install wget yacs
  pip install shapely==1.6.4.post2

• Setup the environment
  Create the default environment setting files.

  # Change directory to <PATH_of_TransT>
  cd TransT
  
  # Environment settings for pytracking. Saved at pytracking/evaluation/local.py
  python -c "from pytracking.evaluation.environment import create_default_local_file; create_default_local_file()"
  
  # Environment settings for ltr. Saved at ltr/admin/local.py
  python -c "from ltr.admin.environment import create_default_local_file; create_default_local_file()"

You can modify these files to set the paths to datasets, results paths etc.

• Add the project path to environment variables
  Open ~/.bashrc, and add the following line to the end. Note to change <path_of_TransT> to your real path.

  export PYTHONPATH=<path_of_TransT>:$PYTHONPATH
  

• Download the pre-trained networks
  Download the network for TransT and put it in the directory set by "network_path" in "pytracking/evaluation/local.py". By default, it is set to pytracking/networks.

Quick Start

Traning

• Modify local.py to set the paths to datasets, results paths etc.
• Runing the following commands to train the TransT. You can customize some parameters by modifying transt.py

  conda activate transt
  cd TransT/ltr
  python run_training.py transt transt

Evaluation

• We integrated PySOT for evaluation.

  You need to specify the path of the model and dataset in the test.py.

  net_path = '/path_to_model' #Absolute path of the model
  dataset_root= '/path_to_datasets' #Absolute path of the datasets

  Then run the following commands.

  conda activate TransT
  cd TransT
  python -u pysot_toolkit/test.py --dataset <name of dataset> --name 'transt' #test tracker #test tracker
  python pysot_toolkit/eval.py --tracker_path results/ --dataset <name of dataset> --num 1 --tracker_prefix 'transt' #eval tracker

  The testing results will in the current directory(results/dataset/transt/)

• You can also use pytracking to test and evaluate tracker. The results might be slightly different with PySOT due to the slight difference in implementation (pytracking saves results as integers, pysot toolkit saves the results as decimals).

Acknowledgement

This is a modified version of the python framework PyTracking based on Pytorch, also borrowing from PySOT and detr. We would like to thank their authors for providing great frameworks and toolkits.

Contact

• Xin Chen (email:[e-mail])

  Feel free to contact me if you have additional questions.