IntPhys 2019: A Benchmark for Visual Intuitive Physics Understanding.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computer Vision

Background:

• Artificial systems require world knowledge for complex visual tasks.
• Infant intuitive physics research inspires a new AI evaluation method.
• Current AI lacks robust understanding of physical dynamics.

Purpose of the Study:

• To propose an evaluation benchmark for intuitive physics in AI systems.
• To diagnose AI's understanding of physical plausibility.
• To compare AI performance against human judgment.

Main Methods:

• Developed a benchmark using a game engine to create videos of possible and impossible physical events.
• Utilized pixel-matched video quadruplets to mitigate perceptual biases.
• Trained two Deep Neural Networks (DNNs) in an unsupervised manner using physically possible videos.
• Employed a future semantic mask prediction objective for training.

Main Results:

• The proposed benchmark effectively tests AI's ability to discriminate physical plausibility.
• Trained DNNs demonstrated varying degrees of success in identifying impossible events.
• Analysis revealed insights into the potentials and limitations of next-frame prediction architectures.
• AI performance was compared against human data for context.

Conclusions:

• The benchmark provides a novel method for assessing intuitive physics in AI.
• Unsupervised learning with future prediction shows potential for intuitive physics acquisition.
• Further research is needed to address limitations in current AI models' physical reasoning.