Physics knowledge aids object perception in dynamic scenes
J. Hamrick, Peter Battaglia, Joshua B. Tenenbaum · Journal of Vision · 2012
People naturally infer unobservable physical properties such as mass, elasticity, and friction in many uncertain and complex scenarios. For example, when picking up a cup of coffee it is useful to determine the weight; similarly, when walking around in socks, it is useful to infer how slippery the floor is. We hypothesize that humans reason about such situations using an 'intuitive mechanics'' that allows physical knowledge to inform visual 'beliefs'' about a dynamic scene as it unfolds over time. To test people's ability to make these inferences, we showed participants simulated 2D movies of arrangements of red and yellow building blocks colliding, collapsing, or remaining upright. They then estimated properties such as the mass ratio or elasticity between differently colored blocks. In different trials we presented static 'previews'' of the scene for varying durations in order to modulate participants' initial visual precision at estimating the blocks' positions. People's judgments were sensitive to variations in mass ratio, elasticity, and preview length, making more consistent and accurate judgments when allowed a longer preview. We compared these results with a probabilistic time-series observer model which makes its judgments using knowledge of collision dynamics and Newtonian mechanics in a simulation-based inference procedure. Visual precision in the model is represented through uncertainty in block positions; both human judgments and model predictions vary proportionally with the level of uncertainty, implying an important role for visual precision in physical reasoning. Our model is additionally designed to handle arbitrarily complex situations and can predict people's judgments in scenarios ranging from simple two-body collisions to complex stacks of objects. We conclude that human judgments of latent physical properties depend both on the precision of visual scene interpretation and knowledge of how physical laws influence the scene's dynamics. Meeting abstract presented at VSS 2012