Wasserstein distance-based auto-encoder tracking

This repository is based on the manuscript "Wasserstein distance-based auto-encoder tracking", under reviewing in the journal NEPL.

Environment

• Anaconda
• Pytorch
• visdom
• Agumentor

conda install pytorch torchvision cudatoolkit=10.2 -c pytorch

Run

You can train the Auto Encoder by yourself using: https://github.com/wahahamyt/train_WAE.git.

This code also integrated in this repository.

Pretrained Auto-Encoder Weights

We also provided the pretrained auto encoder weights:

• Baidu Disk:
  • 2048 z Dimension (2.3GB)：https://pan.baidu.com/s/1yEFpTF9oOHFZtXAedxr0Ow Code: 91pn
  • 256 z Dimension (493MB): https://pan.baidu.com/s/1zMhTGUIYcqroXDKUFPvBqg Code: 7t41
• Google Drive:
  • 2048 z Dimension (2.3GB)：https://drive.google.com/file/d/1g1GMVQeEKSiBboLDm26qArRP5_g8O-dO/view?usp=sharing
  • 256 z Dimension (493MB): https://drive.google.com/file/d/1DeGPatiyGh52TWl4O7cOrdMZCFvDEvdT/view?usp=sharing

After downloaded the weights, it should be renamed as last, then moved to the folder net/weights/.

Start tracking:

Pycharm is recommended for avoiding some path issues (At least 6GB of GPU memory is required [under 2048 z dimension]).

tracking/run_tracker.py

Evaluation

OTB protocal: http://cvlab.hanyang.ac.kr/tracker_benchmark/datasets.html

LaSOT protocal: https://github.com/HengLan/LaSOT_Evaluation_Toolkit

TC128 protocal: http://www.dabi.temple.edu/~hbling/data/TColor-128/TColor-128.html

Citation

If you're using this code in a publication, please cite our paper.

Xu, L., Wei, Y., Dong, C. et al. Wasserstein Distance-Based Auto-Encoder Tracking. Neural Process Lett (2021). https://doi.org/10.1007/s11063-021-10507-9