Computational perception of physical object properties

Abstract

We study the problem of learning physical object properties from visual data. Inspired by findings in cognitive science that even infants are able to perceive a physical world full of dynamic content at a early age, we aim to build models to characterize object properties from synthetic and real-world scenes. We build a novel dataset containing over 17, 000 videos with 101 objects in a set of visually simple but physically rich scenarios. We further propose two novel models for learning physical object properties by incorporating physics simulators, either a symbolic interpreter or a mature physics engine, with deep neural nets. Our extensive evaluations demonstrate that these models can learn physical object properties well and, with a physic engine, the responses of the model positively correlate with human responses. Future research directions include incorporating the knowledge of physical object properties into the understanding of interactions among objects, scenes, and agents.