• TraND
• • Requirements
  • • Installation
  • Data Preparation
  • • Data Pretreatment
    • Data Structrue
  • Train
  • • Stage I: Supervised Prior Knowledge Learning on Source Domain
    • Stage II: Transferable Neighbor Discovery on Target Domain
  • Test
  • Cross domain generation models
  • • 1. Cross virtual-to-real human style based on diffusion model
    • 2. Cross indoor-to-outdoor human walking style based on diffusion model
    • 3. Generation model of dynamic human body based on human nerf
    • 4. Generation model of moving human body based on neural body
  • Citation
  • Acknowledgement

TraND

This is the code for the paper "Jinkai Zheng, Xinchen Liu, Chenggang Yan, Jiyong Zhang, Wu Liu, Xiaoping Zhang and Tao Mei: TraND: Transferable Neighborhood Discovery for
Unsupervised Cross-domain Gait Recognition. ISCAS 2021" (MSA-TC “Best Paper Award - Honorable Mention”)

Requirements

• Conda
• GPUs
• Python 3.7
• PyTorch 1.1.0

Installation

You can replace the second command from the bottom to install
pytorch
based on your CUDA version.

git clone https://github.com/JinkaiZheng/TraND.git
cd TraND
conda create --name py37torch110 python=3.7
conda activate py37torch110
conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=10.0 -c pytorch
pip install -r requirements

Data Preparation

Download CASIA-B and OU-LP

Data Pretreatment

pretreatment_casia.py and pretreatment_oulp.py use the alignment method in
this paper.
In the case of CASIA-B dataset, you need to run the command:

python GaitSet/pretreatment_casia.py --input_path='root_path_of_raw_dataset' --output_path='./data/CASIA-B'

Data Structrue

After the pretreatment, the data structure under the directory should like this

./data
├── CASIA-B
│  ├── 001
│     ├── bg-01
│        ├── 000
│           └── 001-bg-01-000-001.png
├── OULP
│  ├── 0000024
│     ├── Seq00
│        ├── 55
            └── 00000061.png

Train

Stage I: Supervised Prior Knowledge Learning on Source Domain

Training the GaitSet model in the source domain, run this command:

 python GaitSet/train.py --data "casia-b"

Our models: CASIA_best_model and OULP_best_model

Stage II: Transferable Neighbor Discovery on Target Domain

Fine-tuning the GaitSet model in the target domain with TraND method, run this command:

sh Experiment.sh

Our models: CASIA2OULP_best_model and OULP2CASIA_best_model

Test

Testing the model in self domain, such as CASIA-B dataset, run this command:

python GaitSet/test.py --data "casia-b"

Testing the model in cross domain, such as CASIA-B -> OU-LP dataset, run this command:

python GaitSet/test_cross.py --source "casia-b" --target "oulp"

Cross domain generation models

1. Cross virtual-to-real human style based on diffusion model

Download the model at here (code: 6vd1h8).

2. Cross indoor-to-outdoor human walking style based on diffusion model

Download the model at here (code: 3pgcwl).

3. Generation model of dynamic human body based on human nerf

Download the model at here (code: f4c71h).

4. Generation model of moving human body based on neural body

Download the model at here (code: 9qbd5t).

Citation

Please cite this paper in your publications if it helps your research:

@article{DBLP:journals/corr/abs-2102-04621,
  author    = {Jinkai Zheng and
               Xinchen Liu and
               Chenggang Yan and
               Jiyong Zhang and
               Wu Liu and
               Xiaoping Zhang and
               Tao Mei},
  title     = {TraND: Transferable Neighborhood Discovery for Unsupervised Cross-domain
               Gait Recognition},
  journal   = {ISCAS},
  year      = {2021}
}

Acknowledgement

• GaitSet
• AND