Webcam based Gaze Estimation

Estimate gaze on computer screen

Installation Guidelines

1. Set Up Python Virtual Environment

Create a virtual Python environment (Python 3.8-3.11 recommended):

# Using conda (recommended)
conda create --name gaze_estimation python=3.10
conda activate gaze_estimation

# OR using venv
python3 -m venv venv
source venv/bin/activate  # On macOS/Linux
# venv\Scripts\activate  # On Windows

2. Install Dependencies

All dependencies have been consolidated into a single requirements.txt file.

macOS ARM (M1/M2/M3/M4) - Special Setup Required

OpenVINO pip wheels are corrupted on Apple Silicon. Use conda-forge instead:

# Install OpenVINO from conda-forge first
conda install -c conda-forge openvino

# Then install remaining dependencies
pip install --upgrade pip
pip install -r requirements.txt

Windows/Linux/Intel Mac

pip install --upgrade pip
pip install -r requirements.txt

Note: Installation may take several minutes as it downloads PyTorch, OpenVINO, and other large packages.

For Windows users: If you need H.264 video codec support, download openh264-1.8.0-win64.dll from here and copy it to your environment's root folder (e.g., C:\Anaconda3\envs\gaze_estimation).

3. Grant Camera Permission (macOS only)

On macOS, the Terminal needs camera access. When you first run the application, you'll be prompted to grant permission:

• Go to System Settings → Privacy & Security → Camera
• Enable checkbox for Terminal (or iTerm2 if you use that)
• Restart your terminal after granting permission

4. Run the Application

Option A: OpenVINO-based model (recommended for speed)

python src/main.py

Option B: PyTorch pl_gaze model

python src/main_pl.py

This model is based on pytorch_mpiigaze_demo. Training code available at pl_gaze_estimation.

Option C: Web interface (Flask app)

python src/app.py

Then open http://localhost:5000 in your browser.

Option D: Compare with Tobii Eye Tracker

python src/main_compareWithTobii.py

Requires Tobii Eye Tracker 5 SDK and a custom executable file.

Modifications

This repository is based on the original work by Lucas Falch and Katrin Solveig Lohan. Key modifications include:

• macOS ARM (Apple Silicon) compatibility: Fixed OpenVINO integration to work with M1/M2/M3/M4 chips
• Screen size estimation: Added fallback for physical screen dimensions on macOS where screeninfo returns None
• OpenVINO 2025 support: Updated code to support both OpenVINO 2024 and 2025 API versions
• Removed deprecated imports: Fixed turtle module import issue
• Updated installation guide: Added specific instructions for macOS ARM setup

Credits & Original Work

This project is based on:

• Original Repository: WebCamGazeEstimation by Lucas Falch
• Research Paper: Lucas Falch and Katrin Solveig Lohan, "Webcam-based gaze estimation for computer screen interaction", Frontiers in Robotics and AI, Volume 11 - 2024 | https://doi.org/10.3389/frobt.2024.1369566

If you use this code in an academic context, please cite the original paper above.

Additional Credits

The PyTorch pl_gaze model is based on:

• pytorch_mpiigaze_demo
• pl_gaze_estimation