HumanPoseGAN: Generating and Discriminating Human Poses

Welcome to the HumanPoseGAN GitHub repository. This repository hosts the code for a GAN (Generative Adversarial Network) designed to generate and discriminate human poses for the purpose of my master's thesis. The project leverages the power of PyTorch, a leading deep-learning library, to create realistic human poses.

Overview

HumanPoseGAN is a machine learning project focusing on the generation and discrimination of human poses using GANs. The project consists of two main components: the HumanPoseGenerator and the HumanPoseDiscriminator, both implemented as neural network models using PyTorch.

Features

• Human Pose Generation: Using the HumanPoseGenerator model, generate realistic human pose data.
• Human Pose Discrimination: The HumanPoseDiscriminator model distinguishes between real and generated human poses.
• 3D Visualization: Visualize generated poses in 3D using Plotly.
• Dataset Handling: Includes HPFrameDataset for efficient handling of pose data.

Installation

To set up this project, follow these steps:

1. Clone the repository.
2. Install the required packages: pip install -r requirements.py
3. Ensure you have a suitable Python environment.

Usage

To use the HumanPoseGAN, run the main.py script. This will initiate the training process for the GAN. The script includes the creation of both the generator and discriminator models, training on a dataset of human poses, and periodically visualizing the generated poses in 3D.

Code Structure

• HumanPoseDiscriminator: Defines the discriminator model.
• HumanPoseGenerator: Defines the generator model.
• HPFrameDataset: Custom dataset class for loading and transforming pose data.
• visualize_frame: Function to visualize a generated pose in 3D.

Training

• The models are trained using a dataset of human poses.
• Training involves both the generator and discriminator.
• Visualization of generated poses occurs every 100 epochs.

Contributions

Contributions to the project are welcome. Please ensure to follow the existing code structure and maintain the readability and quality of the code.

License

This project is licensed under the MIT License - see the LICENSE file for details.