Generalizing Person Re-identification

Implementation of ECCV2020 paper Generalizing Person Re-Identification by Camera-Aware Instance Learning and Cross-Domain Mixup.

Dependencies

• python 3.6
• pytorch 1.3
• apex
• ignite 0.2.0

Preparation

Download and extract Market-1501, DukeMTMC-reID, CUHK03 and MSMT17. Replace the root paths of corresponding datasets in the config file configs/default/dataset.py.

Train

bash train.sh GPU_ID_0,GPU_ID_1 PATH_TO_YOUR_YAML_FILE

Our code is validated under 2-GPUs setting. GPU_ID_0 and GPU_ID_1 are the indices of the selected GPUs. PATH_TO_YOUR_YAML_FILE is the path to your config yaml file. We also offer the template of config file configs/duke2market.yml, configs/market2duke.yml, configs/single_domain.yml. You can optionally adjust the hyper-parameters in the config yaml file. All of our experiments are conducted under the mix-precision training to reduce the burden of GPU memory, i.e, we set the flag fp16=true.

During the training, the checkpoint files and logs will be saved in ./checkpoints and ./logs directories, respectively.

Test

In our code, the model is evaluated on the target domain at intervals automatically. You can also evaluate the trained model manually by running:

python3 eval.py GPU_ID PATH_TO_CHECKPOINT_FILE [--dataset DATASET]

PATH_TO_CHECKPOINT_FILE is the path to the checkpoint file of the trained model. DATASET is the name of the target dataset. Its value can be {market,duke,cuhk,msmt}. As an intermediate product, the feature matrix of target images is stored in the ./features directory.

Citation

@inproceedings{luo2020generalizing,
  title={Generalizing Person Re-Identification by Camera-Aware Invariance Learning and Cross-Domain Mixup},
  author={Luo, Chuanchen and Song, Chunfeng and Zhang, Zhaoxiang},
  booktitle={European Conference on Computer Vision},
  year={2020}
}