Related Experiment Video
Updated: Jul 4, 2026
10:43
Eye-tracking Technology and Data-mining Techniques used for a Behavioral Analysis of Adults engaged in Learning Processes
Published on: June 10, 2021
Learning Gaze Synthesizer via 3D-Eye Controlled Diffusion and Cross-Domain Feature Alignment
Summary
This study introduces a new method for gaze estimation using generative AI to create realistic eye images. The approach enhances accuracy and robustness in gaze tracking by synthesizing high-quality, diverse eye data.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Human-Computer Interaction
Background:
- Appearance-based gaze estimation methods using full-face images have advanced significantly.
- Current methods require extensive human annotation, limiting industrial-level accuracy and robustness.
- Generative AI can synthesize eye images but often produces low-quality, monotonous data with inaccurate gaze directions.
Purpose of the Study:
- To develop a novel gaze data synthesizer framework for high-quality eye image synthesis.
- To improve the accuracy and robustness of gaze estimation by overcoming limitations of existing generative AI methods.
- To create domain-invariant gaze representations by minimizing feature discrepancies between real and synthetic data.
Main Methods:
- A 3D-eye model is integrated with a stable diffusion large generative model for controlled synthesis of eye images with arbitrary gaze angles.
- A cross-domain feature alignment module is proposed to minimize feature distribution discrepancies between real and synthetic samples.
- The framework aims to generate high-quality synthetic eye images and train a gaze feature extractor for domain-invariant gaze representation.
Main Results:
- The proposed scheme successfully generates high-quality synthetic eye images with diverse gaze angles.
- Qualitative and quantitative experiments demonstrate superior gaze estimation performance compared to state-of-the-art methods.
- The cross-domain feature alignment module effectively minimizes feature distribution discrepancies, leading to domain-invariant gaze representations.
Conclusions:
- The novel gaze data synthesizer framework effectively addresses the limitations of current generative AI methods in gaze estimation.
- The integration of a 3D-eye model and stable diffusion model enables the synthesis of high-quality, arbitrary-gaze eye images.
- The proposed approach achieves state-of-the-art performance in gaze estimation, offering a robust and accurate solution.
