CSCL-CC/Code-for-CSE-Module

Code for Establishing Continuous 2D-3D Surface Correspondences for Cloth-Changing ReID

To help you better understand the CSCL framework presented in our paper, we provide codes for establishing reliable continuous 2D-to-3D mapping for pedestrian images in both general and cloth-changing ReID datasets, which corresponds to the key CSE module in our paper.

Setup

• Python 3.8
• Pytorch 1.11.0 Run python setup.py install to compile this project.

Datasets

General ReID datasets:

• Market1501
• DukeMTMC-ReID

Cloth-Changing ReID datasets:

• LTCC
• PRCC
• VC-clothes
• DP3D

Download the dataset to your local folder. For example, './data/LTCC'

Models

ImageNet Classification Models

• ResNet
• DenseNet
• Inception-ResNet-V2
• Inception-V4
• Xception

Image Segmentation Models

• UNet
• Graphonomy

ReID-Specific models

• MuDeep
• ResNet-mid
• HACNN
• PCB
• MLFN

Train

CSE training codes are implemented in

• trainer/cse_trainer.py: train CSE module for ReID models

To train CSE module, you can do

python engine/cse_trainer.py \
--dataset './data/DP3D' \
--epoch 40