Awesome

Unsupervised Multi-Source Domain Adaptation for Person Re-Identification

Unofficial implementation of CVPR 2021 Oral paper Unsupervised Multi-Source Domain Adaptation for Person Re-Identification

This repo provides the code to implement 3-source (duke+cuhk03+msmt -> market) domain adaptive person re-ID task. It runs on Python 3.6 with Pytorch 1.3.1. For other dependencies, see setup.py.

Installation

# clone this repo
cd MSUDA_REID
python setup.py install

Prepare Datasets

cd examples && mkdir data

Prepare DukeMTMC-reID, Market-1501, CUHK03 and MSMT17 datasets as in MMT.

Train

Two 32GB V100 GPUs are used to train the 3-source adaptation. You can also reduce the batch size to fit your GPU memory.

Stage I: Pre-training on the source domains

bash scripts/multi_src_pretrain.sh 1
bash scripts/multi_src_pretrain.sh 2

Stage II: End-to-end training with MSUDA (RDSBN-MDIF)

bash scripts/multi_src_train_mmt_msuda.sh

Test

Test the trained model with best performance by

bash scripts/test_msuda.sh

Result

Pre-trained models

The best performance model can be downloaded from Baidu Drive password: n0cm.

Place the downloaded model in logs/dukemtmc_cuhk03_msmt17TOmarket1501/resnet50_rcdsbn_mdif-MMT-DBSCAN/

Ack

This repo borrows a lot of code from MMT, thanks!