MaskCL: Semantic Mask-Driven Contrastive Learning for Unsupervised Person Re-Identification with Clothes Change
๐ MaskCL is dedicated to an innovative task: unsupervised clothes changing person re-identification.
โญ Within the realm of clothing changing person re-identification, MaskCL proudly stands as the inaugural unsupervised methodology to attain commendable outcomes across a multitude of datasets!!!
โค๏ธ We warmly welcome and encourage fellow researchers to engage in enlightening discussions and exchanges on this topic!!!
๐Paper Link
https://arxiv.org/abs/2305.13600
๐Dataset๏ผ
We evaluate MaskCL on Six datasets:
Dataset | Link |
---|---|
PRCC | https://www.isee-ai.cn/~yangqize/clothing.html |
LTCC | https://naiq.github.io/LTCC_Perosn_ReID.html |
Celeb-ReID | https://github.com/Huang-3/Celeb-reID |
Celeb-ReID-Light | https://github.com/Huang-3/Celeb-reID |
VC-Clothes | https://wanfb.github.io/dataset.html |
DeepChange | https://github.com/PengBoXiangShang/deepchange |
๐ฌ Remark:
In MaskCL, during the training process, it needs to prepare the original dataset along with the corresponding dataset of person masks.
The mask images used in the MaskCL are generated based on human parsing networks.
In this study, we employed SCHP for pedestrian silhouette extraction.
After generating the corresponding mask image dataset, you should change the dataset path in CMC.py and /utils/dataset/data/preprocessor.py.
โค๏ธ **We extend a warm invitation to researchers to venture into exploring diverse approaches in generating pedestrian semantic information and conducting experiments on MaskCL.
โค๏ธWe are greatly anticipating the researchers' invaluable contributions in terms of sharing and providing feedback on their experimental outcomes!**
๐กTrain:
sh run_code.sh
๐ข We will release the model weight soon!