This repository includes Matlab code for the following paper:
Learning a Visual Tracker from a Single Movie without Annotation (AAAI-2019)
Lingxiao Yang, David Zhang, Lei Zhang
This paper proposes a method that off-line pre-learns a DCFNet [1] in an unsuperivsed manner. It formulates the optimization as a two-stage learning pipeline, which iteratively updates the tracker by alternating between target localization and network optimization. Besides, the tracker is learned from a single movie, for which the data could be easily obtained other vthan collecting thousands of video clips or millions of images.
The trained network can be directly used in DCFNet software [2]. You can simply use the given model in 'model/' to reproduce similar results as shown in below.
1. MatConvNet-1.0-beta24 or latest
2. git clone https://github.com/ZjjConan/UL-Tracker-AAAI2019.git
Please follow this website to compile MatConvNet with GPU support.
Options: To extract SSW proposals, we use OpenCV3.0 and mexopencv, please follow the mexopencv's website to compile them with opencv extra contrib module.
We provide the extracted proposals and images for The Shawshank Redemption in the following links. For other movies, interested readers can pre-process these movies by the provided codes in this repository.
Note: This is a little difference to the original paper, we downsample each images twice for fast propossing. We found that the final performance is slightly better than original one (maybe it is because of a good seed for random initialization ^_^). Results are shown below:
| OTB-2013 (AUC) | OTB-2015 (AUC) | VOT-2015 (A) | VOT-2015 (R) | VOT-2015 (EAO) | |
|---|---|---|---|---|---|
| DCFNet [1] | 62.24 | 58.37 | 0.53 | 1.68 | 0.2174 |
| AAAI-2019 Paper | 62.47 (1.37) | 57.45 (0.64) | 0.53 | 1.70 | 0.2234 |
| This Repository | 63.22 | 59.01 | 0.56 | 1.55 | 0.2473 |
These comparisons are based on the DCFNet_1.0. Since DCFNet_2.0 shows significant better performance than its 1.0 version, currently we are willing to achieve more robust tracking by optimizing DCFNet_2.0.
- Get frames and proposals from the above links, and compile all required softwares.
- Use ul_group_frames to get clip index for each frame.
- Use ul_create_imdb to create a imdb.mat for training.
- Use ul_demo_trainnet to train the network.
Moe detail parameter settings can be found in each ul_demo_* files.
Please check ul_demo_preprocess to find the pre-process steps for new data.
Lingxiao Yang
Email: lingxiao.yang717@gmail.com
If you find UL-Tracker useful in your research, please consider cite:
@inproceedings{yang2019learning,
title={Learning a Visual Tracker from a Single Movie without Annotation},
author={Yang, Lingxiao and Zhang, David and Zhang, Lei},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={33},
pages={9095--9102},
year={2019}
}
[1] Wang Q, Gao J, Xing J, et al. Dcfnet: Discriminant correlation filters network for visual tracking [J]. arXiv preprint arXiv:1704.04057, 2017.