ISP-reID

Codes for 'Identity-Guided Human Semantic Parsing for Person Re-Identification', accepted by ECCV2020 spotlight. We propose the Identity-Guided Human Semantic Parsing approach (ISP) to locate both the human body parts and personal belongings at pixel-level for person re-ID only with person identity labels. ISP-reID is based on the open project, reid strong baseline.

Installation

1. 'cd' to folder where you want to download this repo

2. Run 'git clone https://github.com/CASIA-IVA-Lab/ISP-reID.git'

3. Install dependencies:

   • pytorch>=1.1.0
   • torchvision
   • ignite
   • yacs

4. Install faiss from faiss

Re-ID Dataset Preparation

Create a directory to store reid datasets under this repo or outside this repo. Remember to set your path to the root of the dataset in config/defaults.py for all training and testing or set in every single config file in configs/ or set in every single command.

Take 'DukeMTMC-reID' for exmaple:

Extract dataset to 'DukeMTMC-reID'. The data structure would like:

data
    DukeMTMC-reID
            bounding_box_test/
            bounding_box_train/
            bounding_box_train_parsing_pgt/
            ......

'bounding_box_train_parsing_pgt' is the predicted ground-truth generated by SCHP, which is only used for evaluating the performance of the parsing results of ISP. We have generated the predicted ground-truth for person re-ID datasets on PGT, code: m6n5.

Pre-trained Model Preparation

Download the pre-trained HRNet32 on ImageNet from Model, code: r1o2.

Train

For your convenience, we provide the bash scripts with our recommended settings. Please 'cd' to the root path of this repo and run:

bash ./ISP-*.sh

Reference

We hope that this technique will benefit more computer vision related applications and inspire more works. If you find this technique and repository useful, please cite the paper. Thanks!

@article{zhu2020identity,
  title={Identity-Guided Human Semantic Parsing for Person Re-Identification},
  author={Zhu, Kuan and Guo, Haiyun and Liu, Zhiwei and Tang, Ming and Wang, Jinqiao},
  journal={ECCV},
  year={2020}
}