GazeHRNet: Head-Centric Spatial Encoding and Gaze-Aware Feature Interaction for Gaze Target Detection

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Artificial Intelligence

Background:

• Gaze target detection is crucial for understanding human attention.
• Current methods often neglect the head's role in interpreting gaze direction.
• This limitation hinders accurate modeling of head-scene dependencies.

Purpose of the Study:

• To introduce GazeHRNet, a novel head-centric framework for RGB-based gaze target detection.
• To improve gaze prediction by integrating head-centric reasoning and scene context.
• To enhance robustness and generalization in diverse visual environments.

Main Methods:

• GazeHRNet utilizes Head-Centric Polar Encoding to represent the scene from the gazer's viewpoint.
• Head-Aware Attention Routing organizes visual features based on spatial relevance to the head.
• Combines coarse spatial reasoning with fine-grained anisotropic heatmap prediction.

Main Results:

• GazeHRNet achieved high AUC scores (0.952 and 0.929) on GazeFollow and VideoAttentionTarget datasets.
• Demonstrated low L2 distances (0.102 and 0.103) using only RGB input.
• Showcased improved robustness and generalization through cross-dataset evaluations.

Conclusions:

• GazeHRNet offers a more effective approach to gaze target detection by centering on head-relative scene interpretation.
• The framework successfully addresses limitations of generic localization methods.
• Results indicate significant advancements in gaze prediction accuracy and reliability.