Perception of 3D shape integrates intuitive physics and analysis-by-synthesis

Area of Science:

• Cognitive Science
• Computer Vision
• Psychology

Background:

• Traditional 3D shape perception relies on surface cues.
• Humans can perceive shape even when surface cues are absent, such as with draped objects.

Purpose of the Study:

• To propose a novel framework for 3D shape perception.
• To explain shape perception in both typical and atypical scenarios using an analysis-by-synthesis approach.
• To integrate intuitive physics into a generative model for inferring shape from object deformations.

Main Methods:

• Developed a generative model for image formation incorporating intuitive physics.
• Conducted behavioral and computational studies comparing the proposed model with alternatives.
• Evaluated model performance against human observers (n=174) on accuracy and response times.

Main Results:

• The proposed analysis-by-synthesis model significantly outperformed alternative models.
• The model showed strong correlation with human performance, especially in difficult discriminations.
• Intuitive physics integration effectively explained shape inference from deformations, like cloth draping.

Conclusions:

• The proposed framework provides a robust account of human 3D shape perception.
• Current bottom-up deep neural network models may be insufficient for fully explaining human shape perception.
• The findings suggest pathways for developing more human-like machine vision systems.