An Individual-Difference-Aware Model for Cross-Person Gaze Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate gaze prediction is crucial for understanding human attention and interaction.
• Existing methods often struggle with person-specific variations in eye and face images.
• Cross-person gaze prediction remains a challenging task due to individual differences.

Purpose of the Study:

• To propose a novel method for refining cross-person gaze prediction using only eye and face images.
• To explicitly model and compensate for person-specific differences in gaze prediction.
• To improve the accuracy and robustness of gaze prediction systems.

Main Methods:

• Introduced a novel method incorporating three modules: Validity Module (VM), Self-Calibration (SC), and Person-specific Transform (PT).
• VM identifies and mitigates the impact of invalid eye/face samples (e.g., blinks).
• SC and PT modules learn to compensate for valid samples by bridging initial prediction gaps and learning person-specific transformations.

Main Results:

• The proposed method achieved significant performance improvements over state-of-the-art (SOTA) methods on EVE, XGaze, and MPIIGaze datasets.
• Relative performance gains of 21.7%, 36.0%, and 32.9% were observed on the respective datasets.
• The method was validated on three publicly available datasets and demonstrated superior results.

Conclusions:

• The novel method effectively refines cross-person gaze prediction by modeling person-specific differences.
• The proposed approach offers substantial improvements in gaze prediction accuracy and robustness.
• This work represents a significant advancement in gaze estimation technology, evidenced by winning the GAZE 2021 EVE Challenge.