Improving Machine Vision Using Human Perceptual Representations: The Case of Planar Reflection Symmetry for Object...

Area of Science:

• Computer Vision
• Cognitive Science
• Machine Learning

Background:

• Human-like visual perception is a key goal for machine vision.
• Understanding how human vision can inform machine vision development remains a challenge.

Purpose of the Study:

• To identify systematic differences between machine vision models and human object perception.
• To investigate whether addressing these differences can enhance machine vision performance.

Main Methods:

• Collected a large dataset of human perceptual distances for isolated objects.
• Evaluated common machine vision algorithms against human perceptual data.
• Augmented a state-of-the-art convolutional neural network with symmetry scores.

Main Results:

• Machine vision algorithms systematically underestimate distances for symmetric objects compared to human perception.
• The best algorithms explained approximately 70% of the variance in human perceptual data.
• Augmenting a convolutional neural network with symmetry information improved classification accuracy by 1-10%.

Conclusions:

• Machine vision models exhibit systematic biases compared to human object perception.
• Incorporating human visual principles, such as symmetry, can significantly improve machine vision classification.
• Discovering and rectifying systematic differences between human and machine vision offers a path to more capable AI systems.