A Differential Approach for Gaze Estimation.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Non-invasive gaze estimation often uses single images, leading to accuracy issues due to individual eye variations and biases.
• Calibration is typically required to map generalized gaze predictions to a specific user's actual gaze.

Purpose of the Study:

• To develop a novel approach for gaze estimation that enhances accuracy by directly learning gaze differences between paired eye images.
• To reduce the impact of common image perturbations like alignment and illumination variations inherent in single-image methods.

Main Methods:

• A differential convolutional neural network was trained to predict gaze differences between two eye images from the same subject.
• The model leverages subject-specific calibration images to infer gaze direction for novel eye samples.
• Fine-tuning of the differential network allows adaptation to user-specific reference pairs for consistent predictions.

Main Results:

• The proposed differential network approach consistently outperformed state-of-the-art methods across three public datasets.
• The method achieved high accuracy even with a single calibration sample.
• It demonstrated superior performance compared to methods relying on subject-specific gaze adaptation.

Conclusions:

• Training a differential network to predict gaze differences offers a more robust and accurate gaze estimation solution.
• This approach effectively mitigates common challenges in single-image gaze estimation, improving overall prediction quality.
• The method shows significant potential for improving non-invasive gaze tracking systems.