README.md

A Simple Codebase for Image-based Person Re-identification

Requirements: Python 3.6, Pytorch 1.6.0, yacs

Supported losses

Classification Losses

• CrossEntropy Loss
• CrossEntropy Loss with Label Smooth
• CosFace Loss
• ArcFace Loss
• Circle Loss

Pairwise Losses

• Triplet Loss
• Contrastive Loss
• Pairwise CosFace Loss
• Pairwise Circle Loss

Supported models

• ResNet-50
• ResNet-50-IBN
• IANet

Get Started

• Replace _C.DATA.ROOT and _C.OUTPUT in configs/default.py with your own data path and output path, respectively.
• Run train.sh

Some Results

Market-1501

classification loss	pairwise loss	backbone	top-1	mAP
CrossEntropy	Triplet	ResNet-50	94.5	86.6
CrossEntropy	Contrastive	ResNet-50	94.3	86.4
CrossEntropy	Cosface	ResNet-50	94.3	86.2
CELabelSmooth	Triplet	ResNet-50	95.0	87.4
CELabelSmooth	Contrastive	ResNet-50	94.5	87.1
CELabelSmooth	Cosface	ResNet-50	94.1	86.4
Cosface	Triplet	ResNet-50	95.1	86.7
Cosface	Cosface	ResNet-50	94.5	87.1
Arcface	Triplet	ResNet-50	94.2	86.3
Circle	Circle	ResNet-50	94.7	87.3

MSMT

classification loss	pairwise loss	backbone	top-1	mAP
CrossEntropy	Triplet	ResNet-50	78.9	57.0
CrossEntropy	Contrastive	ResNet-50	79.3	56.7
CrossEntropy	Cosface	ResNet-50	78.2	55.2
CELabelSmooth	Triplet	ResNet-50	79.9	58.0
CELabelSmooth	Contrastive	ResNet-50	80.3	58.7
CELabelSmooth	Cosface	ResNet-50	79.2	56.6
Cosface	Triplet	ResNet-50	78.1	54.1
Cosface	Cosface	ResNet-50	78.8	55.9
Arcface	Triplet	ResNet-50	78.2	54.2
Circle	Circle	ResNet-50	79.7	57.0

Citation

If you use our code in your research or wish to refer to the baseline results, please use the following BibTeX entry.

@InProceedings{CVPR2019IANet
author = {Hou, Ruibing and Ma, Bingpeng and Chang, Hong and Gu, Xinqian and Shan, Shiguang and Chen, Xilin},
title = {Interaction-And-Aggregation Network for Person Re-Identification},
booktitle = {CVPR},
year = {2019}
}