CLaP: Contrast Label Predict for Human Pose Estimation

This repository contains the implementation of CLaP, a framework that leverages contrastive learning for human pose estimation in data-constrained domains.

Overview

CLaP combines contrastive pre-training with supervised pose estimation to improve performance when limited training data is available. The framework consists of:

1. Contrastive Pre-training: Multiple state-of-the-art contrastive learning approaches including:

• SimCLR
• SimSiam
• MoCo

2. Pose Estimation: Supervised training for pose estimation using the pre-trained representations

Project Structure

├── contrastive_training/    # Contrastive pre-training implementations
│   ├── MoCo/               # Momentum Contrast implementation
│   ├── simclr/             # SimCLR implementation
│   ├── simsiam/            # SimSiam implementation 
│   ├── LASCon/             # LASCon implementation (not included in the paper)
│   ├── visualize_data.py   # Tools for visualizing embeddings
│   └── train.py           
├── dataloaders/            # Dataset loading utilities
├── experiments/            # Training configuration files
├── pose_estimation/        # Pose estimation model & training
├── contrastive_HPE.py      # Main training script
└── README.md

Installation

# Clone repository
git clone https://github.com/username/CLaP.git
cd CLaP

# Install dependencies
pip install -r requirements.txt

Usage

Train a model using the provided configuration files:

python contrastive_HPE.py --experiment experiments/resnet18.json

Configuration files in experiments contain hyperparameters and model settings:

• resnet18.json: ResNet18 backbone configurations
• resnet50.json: ResNet50 backbone configurations
• moco.json: MoCo-specific configurations

Datasets

The SkiPose dataset is not publicly available, but can be requested from the authors.

Key Features

Multiple contrastive learning approaches (SimCLR, SimSiam, MoCo, LASCon) Flexible backbone architectures (ResNet18, ResNet50) Multi-view and data augmentation strategies TensorBoard integration for training visualization Configurable training parameters via JSON files

Citation

If you find this work useful, consider citing the paper

Authors

• Davide Cavicchini
• Alessia Pivotto
• Sofia Lorengo
• Andrea Rosani
• Nicola Garau