Dual Regression-Enhanced Gaze Target Detection in the Wild

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Behavioral Science

Background:

• Gaze analysis is crucial for understanding human behavior and social interaction.
• Current gaze target detection methods often require complex models or depth data, limiting practical applications.
• There is a need for efficient and accurate gaze detection models for unconstrained environments.

Purpose of the Study:

• To propose a simple and effective gaze target detection model with low complexity.
• To improve detection accuracy using a dual regression approach.
• To validate the model's performance on diverse datasets, including clinical data for autism screening.

Main Methods:

• Developed a novel dual regression model for gaze target detection.
• Employed coordinate and Gaussian-smoothed heatmap labels for supervised training.
• Inference phase directly outputs gaze target coordinates, avoiding heatmap generation.

Main Results:

• Achieved high accuracy in gaze target detection across various datasets.
• Demonstrated fast inference speed compared to existing methods.
• Showcased strong generalization capabilities on public and clinical autism screening data.

Conclusions:

• The proposed dual regression model offers a simple yet effective solution for gaze target detection.
• The model maintains high accuracy and inference speed with excellent generalization.
• This approach has potential applications in behavior analysis and clinical diagnostics, such as autism screening.