PVEye: A Large Posture-Variant Eye Tracking Dataset for Head-Mounted AR Devices

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Virtual and Augmented Reality

Background:

• Eye tracking is crucial for enhancing user experience in virtual reality (VR) and augmented reality (AR) devices.
• Existing eye tracking datasets lack diversity in wearing postures and suffer from low resolution and camera skew.
• Advanced head-mounted devices increasingly integrate eye tracking as a standard feature.

Purpose of the Study:

• To introduce the Posture-Variant Eye Tracking Dataset (PVEye) to address the scarcity of diverse eye tracking data.
• To facilitate the development and application of appearance-based eye tracking methods.
• To analyze the impact of wearing postures on eye tracking performance.

Main Methods:

• Collected a large-scale dataset (PVEye) with over 11 million high-resolution near-eye images from 104 participants across various wearing postures.
• Evaluated appearance-based eye tracking models (e.g., NVGaze) against traditional feature-based methods using the PVEye dataset.
• Quantified the contribution of posture variations to overall eye tracking error.

Main Results:

• The appearance-based NVGaze model demonstrated superior accuracy and robustness compared to feature-based methods on the PVEye dataset.
• Wearing posture variations were found to significantly impact eye tracking performance.
• Posture-related errors accounted for approximately 45% of the total error variance in eye tracking.

Conclusions:

• The PVEye dataset is a valuable resource for advancing appearance-based eye tracking research, especially considering diverse user postures.
• Posture variation is a critical factor affecting eye tracking accuracy in VR/AR devices.
• Further research should focus on optimizing eye tracking performance by accounting for calibration and posture variations.