yanbeic / Deep-Association-Learning

Tensorflow Implementation on Paper [BMVC2018]Deep Association Learning for Unsupervised Video Person Re-identification

Deep-Association-Learning

Tensorflow Implementation of the paper Chen et al. Deep Association Learning for Unsupervised Video Person Re-identification. BMVC2018. You may refer to our poster for a quick overview.

Getting Started

Prerequisites:

• Datasets: PRID2011 [3], iLIDS-VIDS [4], MARS [5].
• Python 2.7.
• Tensorflow version >= 1.4.0. (For model training)
• Matlab. (For model evaluation)

Data preparation:

1. Download ImageNet pretrained models: mobilenet_v1 [1], resnet_v1_50 [2].

2. Convert image data to tfrecords. (Need to supply your paths in the following .sh file. Check the TODO comments in the .sh file.)

bash scripts/tf_convert_data.sh

Running Experiments

Training:

Train models and extract features. (Need to supply your paths in the following .sh file. Check the TODO comments in the .sh file.)

Model implementation include the following .py files:

• train_dal.py: build and run the training graph.
• association.py: build the anchor learning graph and compute the association losses.
• network.py: define the network.
• utils.py: data preparation.

For example, to train the DAL model using mobilenet_v1 on MARS, run the the following scripts.

bash scripts/train_MARS.sh

Note that you may modify the type of deep model by changing the flag --model_name (eg. --model_name=resnet_v1_50). You can also modify the number of gpus by changing the flag --num_gpus. (eg. --num_gpus=2).

Testing:

Test model performance in matlab. Evaluation codes are placed under the directory evaluation.

For examples, to test the DAL model performance trained on MARS in matlab, run the following command.

clear; model_name = 'mobilenet_b64_dal'; CMC_mAP_MARS

Citation

Please refer to the following if this repository is useful for your research.

Bibtex:

@inproceedings{chen2018bmvc,
  title={Deep Association Learning for Unsupervised Video Person Re-identification},
  author={Chen, Yanbei and Zhu, Xiatian and Gong, Shaogang},
  booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
  year={2018}
}

License

This project is licensed under the MIT License - see the LICENSE.md file for details.

References

[1] Howard et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv 2017.
[2] He et al. Deep Residual Learning for Image Recognition. CVPR 2016.
[3] Hirzer et al. Person Re-Identification by Descriptive and Discriminative Classification. SCIA 2011.
[4] Wang et al. Person Re-Identification by Video Ranking. ECCV 2014.
[5] Zheng et al. MARS: A Video Benchmark for Large-Scale Person Re-identification. ECCV 2016.

Acknowledgements

This repository is partially built upon the tensorflow/models repository. The evaluation code (cmc & mAP) is partially borrowed from the MARS-evaluation repository.