Abstract
Imagine watching water rush through a ramp dotted with planks. We may readily recognize the viscosity of the liquid and its non-rigid boundaries, but here we ask: Does the visual system temporally segment the complex dynamics of liquids into discrete events, reflecting a prediction hierarchy? To address this, we simulated nine brief naturalistic animations of liquids interacting with 25 planks in a 5-by-5 maze formation. The liquid flowed from the top of the planks, starting either from the left, middle, or right. Studying event segmentation in this setting poses a unique challenge: Unlike a typical physical scene with rigid bodies (e.g., colliding balls), there is no obvious pattern for temporally demarcating the complex branching and splashing flow of the liquid through the planks. Instead, we take a data-driven approach using a recent performance-based task for discovering event boundaries. We transiently (120 ms) slowed down each animation at one of 18 different time points and recruited participants (N=30 per scene and time point combination) to press a key when they noticed a pause, with the idea that detection accuracy would be decreased at event boundaries. Chi-squared tests revealed a significant effect of time points on the number of detections in five of the nine animations; three others were marginally significant. Fisher's exact tests comparing participants’ detection counts in each time point and the original videos (without slow-downs) indicated that 1-2 time windows in a typical video were as undetectable as the original video. Finally, we found that pixel-distances resulting from slow-downs significantly correlated with detection rates (r = .55). Future work should systematically decouple the effects of pixel change and potential event boundaries in perception of liquid dynamics. Spontaneous event segmentation suggests that the visual system may build predictive “event models” of near-future dynamics of otherwise non-linear, complex liquid flow.