OpenEDS2020 Challenge on Gaze Tracking for VR: Dataset and Results.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Biomedical Engineering

Background:

• Eye-tracking technology is crucial for understanding user behavior and developing assistive technologies.
• Existing datasets often lack the high frame rates and controlled conditions necessary for robust gaze analysis.
• The OpenEDS2020 dataset addresses these limitations with synchronized, high-frequency eye-image sequences.

Purpose of the Study:

• To introduce and release the OpenEDS2020 dataset for eye-tracking research.
• To establish baseline performance for gaze prediction and semantic eye segmentation tasks.
• To benchmark the performance of deep learning models on these challenging tasks.

Main Methods:

• Development of two competitions: Gaze Prediction Challenge and Sparse Temporal Semantic Segmentation Challenge.
• Utilizing the OpenEDS2020 dataset, comprising 87 subjects' eye-image sequences at 100 Hz.
• Employing deep learning approaches for both baseline models and challenge solutions.

Main Results:

• Baselines achieved 5.37 degrees average angular error for gaze prediction and 84.1% mIoU for semantic segmentation.
• Winning solutions significantly outperformed baselines, reaching 3.17 degrees for gaze prediction and 95.2% mIoU for segmentation.
• The dataset was made publicly available to foster further research.

Conclusions:

• The OpenEDS2020 dataset provides a valuable resource for advancing eye-tracking research.
• Deep learning models demonstrate significant potential for accurate gaze prediction and semantic eye segmentation.
• The challenge highlighted the feasibility of high-performance eye-tracking in controlled virtual reality environments.