Real-time 'Actor-Critic' tracking

Code for Real-time 'Actor-Critic' tracking accepted by ECCV 2018

Introduction

We propose a novel tracking algorithm with real-time performance based on the ‘Actor-Critic’ framework.
Results on OTB100

Requirements

Tensorflow 1.4.0 (Train) and Pytorch 0.3.0 (Test)
CUDA 8.0 and cuDNN 6.0
Python 2.7

Usage

Train

Please download the ILSVRC VID dataset, and put the VID folder into $(ACT_root)/train/
(We adopt the same videos as meta_trackers. You can find more details in ilsvrc_train.json.)
Run the $(ACT_root)/train/DDPG_train.py to train the 'Actor and Critic' network.

Test

Please run $(ACT_root)/tracking/run_tracker.py for demo.

Citation

If you find ACT useful in your research, please kindly cite our paper:

@InProceedings{Chen_2018_ECCV,
author = {Chen, Boyu and Wang, Dong and Li, Peixia and Wang, Shuang and Lu, Huchuan},
title = {Real-time 'Actor-Critic' Tracking},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}

Contact

If you have any questions, please feel free to contact bychen@mail.dlut.edu.cn

Acknowledgments

Many parts of this code are adopted from other related works (py-MDNet and meta_trackers)

ZhangNYG/ACT