README.md

Unsupervised Person Re-identification: Clustering and Fine-tuning

Setup

All our code is implemented in Keras, Tensorflow (Python). Installation instructions are as follows:

pip install --user tensorflow-gpu
pip install --user keras
pip install --user sklearn

Baseline (Fine-tuned ResNet-50)

We provide the fine-tuned models as follows:

1. Duke 2. Market 3. CUHK03 4. Duke + Market 5. Duke + CUHK03 6. Market + CUHK03

Progressive Unsupervised Learning (PUL)

To reappear Duke -> Market:

1. Rename the above fine-tuned "Duke" model as "0.ckpt", which is treated as original model for PUL;

2. Create directory "checkpoint" under the folder "PUL", and move the original model "0.ckpt" into the "checkpoint";

3. Modify PUL/unsupervised.py or PUL/semi-supervised.py and PUL/evaluate.py to train and evaluate Duke -> Market.

If you find this code useful, consider citing our work:

@article{fan17unsupervised,
  author    = {Hehe Fan and Liang Zheng and Yi Yang},
  title     = {Unsupervised Person Re-identification: Clustering and Fine-tuning},
  journal   = {arXiv preprint arXiv:1705.10444},
  year      = {2017}
}