This repository contains PyTorch codes for the ACM MM 2022 paper "Revisiting Stochastic Learning for Generalizable Person Re-identification"
Paper Links from CVTEAM webpage
conda env create -f reid.yml
conda activate reid
pip install -r reid.txt
Train datasets and evaluation datasets should be download in the directory datasets. The download way is shown in the MetaBIN.
Download pretrained model into the directory pretrain.
Download trained models into the directory model_weights.
# Train or evaluate in the C2-C3-D-M-CS setting
sh train1.sh
sh eval1.sh
# Train or evaluate in the C3-D-M-MT setting
sh train2.sh
sh eval2.sh
The implement of proposed stochastic sampler in ./fastreid/data/samplers/triplet_sampler.py/DomainSplitBalancedSampler
The implement of proposed gradient dropout in ./fastreid/engine/hooks.py/DropoutSGDHook and ./fastreid/solver/optim/dropout_sgd.py/DropoutSGD
@InProceedings{Zhao_2022_ACMMM,
author = {Zhao, Jiajian and Zhao, Yifan and Chen, Xiaowu and Li, Jia},
title = {Revisiting Stochastic Learning for Generalizable Person Re-identification},
booktitle = {Proceedings of the 30th ACM International Conference on Multimedia (ACM MM)},
month = {October},
year = {2022},
pages = {1758-1768}
}
This repository is based on the implementation of fast-reid.