Visual attention graph.

Area of Science:

• Cognitive Science
• Neuroscience
• Computer Vision

Background:

• Visual attention is crucial for active visual tasks and understanding object relationships.
• Current methods for predicting visual attention often rely on individual fixations or scanpaths from raw gaze data, which can have high variability.
• Existing approaches may not adequately capture common attention patterns due to a lack of semantic scene information.

Purpose of the Study:

• To propose a novel graph-based representation, the visual attention graph (VAG), for encoding both visual saliency and scanpaths.
• To better reveal common attention behaviors by integrating semantic information.
• To establish a robust benchmark for evaluating attention prediction methods and explore VAG's potential in cognitive state assessment.

Main Methods:

• Developed a visual attention graph (VAG) to represent visual saliency and scanpaths simultaneously.
• Defined semantic-based scanpaths as paths on the graph.
• Calculated object saliency by computing fixation density on graph nodes.

Main Results:

• The proposed VAG, along with new evaluation metrics, provides a superior benchmark for attention prediction.
• Experiments demonstrated the VAG's effectiveness in capturing common attention patterns.
• The VAG shows promising potential for applications in assessing human cognitive states, including autism spectrum disorder screening and age classification.

Conclusions:

• The visual attention graph (VAG) offers a novel and effective method for representing and analyzing visual attention patterns.
• VAG enhances the understanding of common attention behaviors by integrating semantic scene information.
• The VAG framework has significant potential for future research in cognitive neuroscience and clinical applications.