This is the official implementation of the paper "Identity-Sensitive Knowledge Propagation for Cloth-Changing Person Re-identification"
git clone https://github.com/KimbingNg/DeskPro && cd DeskPro
Download the prepared datasets and the pretrained teacher models from this link, and runs
tar -zxvf dataset.tar.gz
pip3 install -r ./requirements.txt
Train on Celeb:
dataset=celeb
CUDA_VISIBLE_DEVICES=0 python3 main.py --config ./config_files/celeb_config.yaml \
data_test $dataset"_lr_hr_mask" data_train $dataset"_lr_hr_mask" \
loss '1*CrossEntropy+1*Triplet' \
tag 'exp_version' \
batchid 5 \
batchimage 6 \
kd_loss.enable True \
kd_loss.T 5. \
kd_loss.alpha 0.7 \
mse.mse_weight 7.0 \
mse.spatial_attn_lr 1.0 \
forward_mode all \
pre_train "$dataset"_teacher.pth
Train on Celeb-light:
dataset=celeb-light
CUDA_VISIBLE_DEVICES=0 python3 main.py --config ./config_files/celeb_config.yaml \
data_test $dataset"_lr_hr_mask" data_train $dataset"_lr_hr_mask" \
loss '1*CrossEntropy+1*Triplet' \
tag 'exp_version' \
batchid 5 \
batchimage 6 \
kd_loss.enable True \
kd_loss.T 5. \
kd_loss.alpha 0.7 \
mse.mse_weight 7.0 \
mse.spatial_attn_lr 1.0 \
forward_mode all \
pre_train "$dataset"_teacher.pth
Train on PRCC:
dataset=prcc
CUDA_VISIBLE_DEVICES=0 python3 main.py --config ./config_files/prcc_config.yaml \
data_test $dataset"_lr_hr_mask" data_train $dataset"_lr_hr_mask" \
loss '1*CrossEntropy+1*Triplet' \
tag 'exp_version' \
batchid 5 \
batchimage 6 \
kd_loss.enable True \
kd_loss.T 1. \
kd_loss.alpha 0.8 \
mse.mse_weight 7.0 \
mse.spatial_attn_lr 1.0 \
forward_mode all \
pre_train "$dataset"_teacher.pth
If you find this work useful in your research, please consider citing:
@inproceedings{wuIdentitySensitiveKnowledgePropagation2022,
title = {Identity-{{Sensitive Knowledge Propagation}} for {{Cloth-Changing Person Re-identification}}},
booktitle = {2022 {{IEEE International Conference}} on {{Image Processing}}},
author = {Wu, Jianbing and Liu, Hong and Shi, Wei and Tang, Hao and Guo, Jingwen},
year = {2022},
publisher = {{IEEE}}
}
The codes was built on top of deep-person-reid, reid-strong-baseline, MGN-pytorch, and LightMBN, we thank the authors for sharing their code publicly.