The role of surface-based representations of shape in visual object recognition

Area of Science:

• Cognitive Psychology
• Computer Vision
• Neuroscience

Background:

• Three-dimensional (3D) object recognition is crucial for human and artificial intelligence.
• Understanding how shape primitives (contours, surfaces, volumes) contribute to recognition is key.
• Existing models often emphasize volumetric representations, but empirical evidence is mixed.

Purpose of the Study:

• To contrast the roles of surfaces versus volumetric shape primitives in 3D object recognition.
• To investigate how viewpoint, occlusion, and part similarity affect whole-part matching performance.
• To challenge volumetric part-based models and support surface-based representations.

Main Methods:

• Fifty observers performed a whole-part matching task with novel 3D objects.
• Subsets of closed contour fragments, surfaces, or volumetric parts were matched to whole objects.
• Manipulated factors included part viewpoint, surface occlusion, and target-distractor similarity (nonaccidental and metric properties).

Main Results:

• A whole-part matching advantage was observed for surface-based parts and volumes over contour fragments.
• No significant benefit of volumetric parts over surfaces was found.
• Performance decreased when matching volumetric parts to wholes with occluded surfaces, irrespective of viewpoint or similarity.

Conclusions:

• Findings challenge models relying solely on volumetric part-based shape representations for recognition.
• Results support a surface-based model for high-level shape representation in 3D object recognition.
• Surface information appears more critical than volumetric primitives for robust object recognition.