Efficient End-to-End Convolutional Architecture for Point-of-Gaze Estimation

Area of Science:

• Computer Vision
• Human-Computer Interaction

Background:

• Point-of-gaze estimation is crucial for enhancing user experience and enabling new interaction methods.
• The COVID-19 pandemic accelerated the need for advanced e-meeting platforms, highlighting limitations in current gaze estimation techniques.
• Existing research often involves complex data collection, creating a barrier to wider adoption.

Purpose of the Study:

• To develop a non-restrictive and efficient methodology for acquiring diverse and high-quality gaze data.
• To introduce a novel convolutional neural network (CNN) for calibration-free point-of-gaze estimation.
• To establish a new state-of-the-art baseline for gaze estimation accuracy.

Main Methods:

• A novel data acquisition methodology designed for ease of use and increased data yield.
• Development of a specialized convolutional neural network (CNN) architecture for accurate gaze estimation without calibration.
• Performance evaluation on the MPIIFaceGaze dataset and a newly collected dataset.

Main Results:

• The proposed data acquisition method significantly increases data yield without sacrificing quality or diversity.
• The novel CNN architecture achieves superior performance compared to existing state-of-the-art methods on the MPIIFaceGaze dataset.
• The CNN establishes a strong performance baseline on the newly collected dataset, demonstrating its generalizability.

Conclusions:

• The presented approach offers a practical and effective solution for calibration-free point-of-gaze estimation.
• This work addresses the limitations of current methods, paving the way for more accessible and advanced gaze tracking applications.
• The findings have significant implications for improving virtual collaboration tools and human-device interaction.