hhl's Introduction

Generalizing A Person Retrieval Model Hetero- and Homogeneously

================================================================

Code for Generalizing A Person Retrieval Model Hetero- and Homogeneously (ECCV 2018). [paper]

Preparation

Requirements: Python=3.6 and Pytorch=0.4.0

Install Pytorch
Download dataset
- reid_dataset [GoogleDriver]
- The reid_dataset including Market-1501 (with CamStyle), DukeMTMC-reID (with CamStyle), and CUHK03
- Unzip reid_dataset under 'HHL/data/'

CamStyle Generation

You can train CamStyle model and generate CamStyle imgaes with stargan4reid

Training and test domain adaptation model for person re-ID

Baseline

# For Duke to Market-1501
python baseline.py -s duke -t market --logs-dir logs/duke2market-baseline
# For Market-1501 to Duke
python baseline.py -s market -t duke --logs-dir logs/market2duke-baseline

# For Duke to Market-1501
python HHL.py -s duke -t market --logs-dir logs/duke2market-HHL
# For Market-1501 to Duke
python HHL.py -s market -t duke --logs-dir logs/market2duke-HHL

Results

	Duke to Market		Market to Duke
Methods	Rank-1	mAP	Rank-1	mAP
Baseline	44.6	20.6	32.9	16.9
HHL	62.2	31.4	46.9	27.2

References

[1] Our code is conducted based on open-reid
[2] StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation , CVPR 2018
[3] Camera Style Adaptation for Person Re-identification. CVPR 2018.

Citation

If you find this code useful in your research, please consider citing:

@inproceedings{zhong2018generalizing,
title={Generalizing A Person Retrieval Model Hetero- and Homogeneously},
author={Zhong, Zhun and Zheng, Liang and Li, Shaozi and Yang, Yi},
booktitle ={ECCV},
year={2018}
}