Nonaccidental properties underlie human categorization of complex natural scenes

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Visual Perception

Background:

• Humans excel at rapid and accurate natural scene categorization.
• Understanding the visual properties enabling this ability is crucial for visual neuroscience and AI.
• Previous research explored various scene properties, but their specific role in human categorization remains debated.

Purpose of the Study:

• To identify the critical structural properties of natural scenes that facilitate human categorization.
• To compare human categorization performance with computational models based on scene properties.
• To elucidate the visual features processed in early visual areas, such as V2, that support scene recognition.

Main Methods:

• Extracted structural properties (orientation, length, curvature, junction types/angles) from line drawings of natural scenes.
• Developed computational models for scene categorization using these properties.
• Conducted a six-alternative forced-choice human scene categorization experiment.
• Perturbed junction properties in line drawings to assess their impact on human accuracy.

Main Results:

• Both computational analysis and human experiments revealed that scene properties contain category-relevant information.
• Human categorization accuracy was significantly influenced by contour junctions and curvature.
• Orientation and contour length showed less contribution to human performance compared to junctions and curvature.
• Perturbing junctions led to a significant decrease in human scene categorization accuracy.

Conclusions:

• Human scene categorization relies on specific nonaccidental properties, particularly contour curvature and junction configurations.
• These identified properties align with those used in object recognition and are represented in visual area V2.
• The findings suggest a shared neural basis for object and scene recognition in early visual processing.