Virtual Clothin TryOn

Dataset Generation

Overview

This guide provides details on generating the dataset structure required for AI Clothing Try-On. The dataset consists of train/ and test/ directories, each containing cloth/ and image/ subdirectories. Additional processing generates cloth_mask/, image-parse-v3/, openpose_img/, and openpose_json/ directories.

Requirements

Hardware

• GPU: NVIDIA GPU with CUDA support (Recommended: NVIDIA RTX series)
• Memory: At least 8GB RAM

Software

Operating System

• Windows 10/11 (Recommended)
• Ubuntu (Optional)

Dependencies

• Python 3.8+
• OpenCV (cv2)
• numpy
• Pillow (PIL)
• rembg
• Docker
• CUDA
• MS Visual Studio (For Windows)

Installation

Step 1: Set Up Virtual Environment

python -m venv dataset_env
source dataset_env/bin/activate  # On Windows use: dataset_env\Scripts\activate

Step 2: Install Dependencies

pip install numpy pillow opencv-python rembg

Step 3: Install CUDA and cuDNN

For Windows:

1. Download CUDA Toolkit:

   • Go to CUDA Toolkit Downloads
   • Select your operating system and download the latest stable version.
   • Install it and ensure to add CUDA to your system’s PATH.

2. Download cuDNN:

   • Visit NVIDIA cuDNN and log in to download.
   • Extract and copy the contents to your CUDA installation directory (e.g., C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.x\)

For Linux:

sudo apt update
sudo apt install -y nvidia-cuda-toolkit

Step 4: Install Microsoft Visual Studio (Windows Users Only)

1. Download Visual Studio:
   • Visit Microsoft Visual Studio
   • Download and install the Community Edition.
2. Enable C++ Development Tools:
   • During installation, select Desktop Development with C++.
3. Set Up Environment Variables:
   • Open System Properties > Advanced > Environment Variables
   • Add C:\Program Files (x86)\Microsoft Visual Studio\<version>\Community\VC\Auxiliary\Build to PATH.

Dataset Structure

The dataset should be organized as follows:

dataset/
│-- train/
│   │-- cloth/  # Original cloth images
│   │-- cloth_mask/  # Generated cloth masks
│   │-- image/  # Original images
│   │-- image-parse-v3/  # Processed images (output of SCHP)
│   │-- openpose_img/  # OpenPose processed images
│   │-- openpose_json/  # OpenPose keypoint JSON files
│
│-- test/
│   │-- cloth/  # Original cloth images
│   │-- cloth_mask/  # Generated cloth masks
│   │-- image/  # Original images
│   │-- image-parse-v3/  # Processed images (output of SCHP)
│   │-- openpose_img/  # OpenPose processed images
│   │-- openpose_json/  # OpenPose keypoint JSON files

1. Generating Cloth Masks

To generate the cloth masks from the cloth/ directory, run the following command:

python cloth_mask.py

Cloth Mask Generation Process

The script cloth_mask.py follows these steps:

1. Reads images from the cloth/ directory.
2. Removes the background using rembg.
3. Converts the foreground to white and background to black.
4. Saves the processed mask images in the cloth_mask/ directory.

2. OpenPose Integration

To generate OpenPose images and keypoints:

python openpose.py

OpenPose Processing Steps

1. Uses OpenPose to extract keypoints from images in image/.
2. Saves the skeleton images in openpose_img/.
3. Saves the keypoints JSON files in openpose_json/.

Docker Image Setup

The OpenPose processing runs inside a Docker container. The package virtual_clothing_tryon openpose_cmu contains the necessary setup. Ensure you have CUDA and MS Visual Studio installed before using OpenPose.

To build and run the OpenPose container:

docker build -t openpose_cmu .
docker run --gpus all -v /path/to/dataset:/opt/openpose -it openpose_cmu

3. Human Parsing with SCHP

To perform semantic human parsing, run the following command:

python image_parse.py

SCHP Processing Steps

1. Loads pre-trained SCHP model.
2. Parses human images from image/.
3. Saves parsed outputs in image-parse-v3/.

SCHP Repository:

This implementation is based on the Self-Correction-Human-Parsing (SCHP) repository.

Troubleshooting

• Missing Dependencies: Ensure all required libraries are installed.
• Incorrect Image Paths: Check if the dataset is structured correctly before running scripts.
• CUDA Not Recognized: Ensure CUDA is installed and correctly added to the system PATH.
• Docker Issues: Verify that Docker, CUDA, and MSVS are properly installed.

Acknowledgements

This project is part of the AI Clothing Try-On dataset preparation pipeline.