Non-local Attention Association Scheme for Online Multi-Object Tracking

This is the implementation of the paper Non-local attention association scheme for online multi-object tracking. We integrate non-local attention [1] for multi-object tracking. The code framework for MOT benefits from the MDP [2].

Prerequisites

• Cuda 8.0
• Cudnn 5.1
• Python 2.7
• Keras 2.0.5
• Tensorflow 1.1.0

For example:

conda create -n mot anaconda python=2.7
conda activate mot
conda install -c menpo opencv
pip install tensorflow-gpu==1.1.0
pip install keras==2.0.5

Usage

1. Download the DMAN model and put it into the "model/" folder.
2. Download the MOT16 dataset, unzip it to the "data/" folder.
3. Cd to the "ECO/" folder, run the script install.m to compile libs for the ECO tracker
4. Run the socket server script:

cd nlaa
python calculate_similarity.py

5. Run the socket client script DMAN_demo.m in Matlab.

Citation

If you use this code, please consider citing:

@article{nlaa,
	author={Haidong Wang and Saizhou Wang and Jingyi Lv and Chenming Hu and Zhiyong Li},
	title={Non-local Attention Association Scheme for Online Multi-Object Tracking.},
    journal={Image and Vision Computing},
	volume=102,
	year=2020,
}

References

[1] Zhang, Yulun et al. "Residual Non-local Attention Networks for Image Restoration.",ICLR 2019 (2019)

[2] Xiang, Y., Alahi, A., Savarese, S.: Learning to track: Online multi-object tracking by decision making. In: ICCV (2015)