Temporal 3D Shape Modeling for Video-based Cloth-Changing Person Re-Identification (SEMI)

This repository contains official implementation for the paper: Temporal 3D Shape Modeling for Video-based Cloth-changing Person Re-Identification (SEMI) at WACV'24 - 4th Real-World Surveillance Workshop.

1. Features

Supported CNN backbones

• c2dres50: C2DResNet50
• i3dres50: I3DResNet50
• ap3dres50: AP3DResNet50
• nlres50: NLResNet50
• ap3dnlres50: AP3DNLResNet50

Summary of VCCRe-ID datasets

This baseline currently supports two public VCCRe-ID datasets: VCCR and CCVID.

Dataset	Paper	Num.IDs	Num.Tracklets	Num.Clothes/ID	Public	Download
Motion-ReID	link	30	240	-	X	-
CVID-reID	link	90	2980	-	X	-
SCCVRe-ID	link	333	9620	2~37	X	-
RCCVRe-ID	link	34	6948	2~10	X	-
CCPG	link	200	~16k	-	Per Request	project link
CCVID	link	226	2856	2~5	Yes	link
VCCR	link	392	4384	2~10	Yes	link

2. Running instructions

2.1. Getting started

Create virtual environment

First, create a virtual environment for the repository

conda create -n semi python=3.8

then activate the environment

conda activate semi

Clone the repository

git clone https://github.com/dustin-nguyen-qil/Video-based-Cloth-Changing-ReID-Baseline.git

Next, install the dependencies by running ...

pip install -r requirements.txt

2.2. Data Preparation

1. Download the datasets VCCR and CCVID following download links above
2. You need the pickle files containing the paths to sequence images, clothes id, identity and camera id of the sequences. To do this:
   • Create a folder named data inside the repository
   • Run the following command line (Note: replace the path to the folder storing the datasets and the dataset name)

python datasets/prepare.py --root "/media/dustin/DATA/Research/Video-based ReID" --dataset_name vccr

2.3. Run evaluation only to reproduce results presented in the paper

If you want to see the evaluation results with our pretrained model on VCCR, follow these steps:

• Download our pretrained model from here (password: dustinqil), put it in work_space/save.
• Replace the path to the pretrained model in test.py
• Run

python test.py

• Evaluation results will be saved to work_space/output

2.4. Run training and testing

Configuration options

Go to ./config.py to modify configurations if needed: Dataset (VCCR or CCVID), number of epochs, batch size learning rate, CNN backbone (according to model names above), etc.

Preparation

Create a folder named work_space as below.

Download the pretrained SPIN model and the SMPL mean parameters needed to train the 3D regressor from here (password: dustinqil). Put it inside work_space.

data
work_space
|--- save
|--- output
|--- tsm
main.sh

Run

bash main.sh

• Checkpoints will be automatically saved to work_space/lightning_logs.
• Trained model will be automatically saved to work_space/save.
• Testing results will be automatically saved to work_space/output.

If you want to train from checkpoint, add checkpoint path to RESUME in config.py.

Citation

If you find this repo helpful, please cite:

@InProceedings{Nguyen_2024_WACV,
    author    = {Nguyen, Vuong D. and Mantini, Pranav and Shah, Shishir K.},
    title     = {Temporal 3D Shape Modeling for Video-Based Cloth-Changing Person Re-Identification},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops},
    month     = {January},
    year      = {2024},
    pages     = {173-182}
}

Acknowledgement

Related repos:

• Simple-CCReID.