ABMT

This is the official PyTorch implementation of the WACV 2021 paper:Enhancing Diversity in Teacher-Student Networks via Asymmetric branches for Unsupervised Person Re-identification (WACV 2021).

[Video]

Installation

git clone https://github.com/chenhao2345/ABMT
cd ABMT
python setup.py install

Prepare Datasets

cd examples && mkdir data

Download the raw datasets DukeMTMC-reID, Market-1501, MSMT17, and then unzip them under the directory like

ABMT/examples/data
├── dukemtmc-reid
│   └── DukeMTMC-reID
├── market1501
└── msmt17
    └── MSMT17_V1(or MSMT17_V2)

Prepare Pre-trained Models

When training with the backbone of IBN-ResNet-50, you need to download the ImageNet pre-trained model from this link and save it under the path of logs/pretrained/.

mkdir logs && cd logs
mkdir pretrained

The file tree should be

MMT/logs
└── pretrained
    └── resnet50_ibn_a.pth.tar

Example #1:

Transferring from DukeMTMC-reID to Market-1501 on the backbone of ResNet-50, i.e. Duke-to-Market (ResNet-50).

Stage I: Pre-training on the source domain

sh ABMT_source_pretrain.sh dukemtmc-reid market1501 resnet50_AB 

Stage II: End-to-end training with ABMT

sh ABMT_target_adaptation.sh dukemtmc-reid market1501 resnet50_AB

Example #2:

Transferring from DukeMTMC-reID to Market-1501 on the backbone of IBN-ResNet-50, i.e. Duke-to-Market (IBN-ResNet-50).

Stage I: Pre-training on the source domain

sh ABMT_source_pretrain.sh dukemtmc-reid market1501 resnet_ibn50a_AB

Stage II: End-to-end training with ABMT

sh ABMT_target_adaptation.sh dukemtmc-reid market1501 resnet_ibn50a_AB

Example #3:

Fully unsupervised learning on Market-1501 on the backbone of ResNet-50.

End-to-end training with ABMT

sh ABMT_fully_unsupervised.sh market1501 resnet50_AB

Results

Unsupervised Domain Adaptation Results

Fully Unsupervised Results

Acknowledgment

This implementation is based on MMT (ICLR 2020).

Citation

@InProceedings{Chen_2021_WACV,
author = {Chen, Hao and Lagadec, Benoit and Bremond, Francois},
title = {Enhancing Diversity in Teacher-Student Networks via Asymmetric branches for Unsupervised Person Re-identification},
booktitle = {The IEEE Winter Conference on Applications of Computer Vision (WACV)},
year = {2021}
}