A top-down saliency model with goal relevance.

Area of Science:

• Cognitive Science
• Computer Vision
• Neuroscience

Background:

• Current visual saliency models often use machine learning for top-down factors but lack a general theory for task-attention interaction.
• Existing methods require extensive data and do not explain why certain information attracts gaze based on task relevance.

Purpose of the Study:

• To construct a combined saliency model integrating bottom-up, learned top-down, and goal relevance features.
• To test the impact of goal relevance on predicting human eye movements in a task-based scenario.

Main Methods:

• Recorded eye movements of 80 participants playing variants of the Mario video game.
• Developed a computational model incorporating bottom-up, learned top-down, and goal relevance features.
• Evaluated model performance using the Normalized Scanpath Saliency (NSS) score.

Main Results:

• The combined saliency model significantly improved prediction accuracy compared to models without goal relevance.
• The addition of goal relevance increased the NSS score from 4.35 to 5.82 (p < 1 × 10-100).

Conclusions:

• Goal relevance is a crucial factor in predicting task-directed visual attention and eye movements.
• The findings support Tanner and Itti's theory of goal relevance and advance computational models of visual saliency.