Mid-level feature contributions to category-specific gaze guidance.

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Visual Perception

Background:

• Scene category prediction from eye movements is established.
• Category prediction dynamics reveal bottom-up and top-down information influences.

Purpose of the Study:

• To quantify the contribution of different image feature levels to gaze guidance in scene perception.
• To differentiate the roles of low-level and mid-level features in bottom-up vs. top-down visual attention.

Main Methods:

• Gaze tracking of participants viewing grayscale photographs and line drawings of scenes.
• Analysis of prediction accuracy over a 2-second time course.
• Computation and comparison of low-level (luminance, orientation) and mid-level (symmetry, junctions) image features.

Main Results:

• Scene categories were predictable from fixation density throughout the 2-second viewing period for both image types.
• Prediction accuracy initially decreased (bottom-up) then increased (top-down), consistent with prior findings.
• Low-level salience primarily drove initial bottom-up gaze guidance.
• Mid-level features, particularly junctions, contributed significantly to top-down, knowledge-driven gaze guidance, while symmetry influenced both.

Conclusions:

• Image features at varying representational levels differentially guide gaze during scene perception.
• Mid-level structural features play a crucial role in integrating bottom-up and top-down visual processing for scene understanding.