Abstract
Liquids like water, syrup and tar are highly mutable. Intrinsic properties (e.g. viscosity) and external forces (e.g. object interactions, gravity) cause liquids to adopt a wide range of different shapes depending on the context. Previous work has shown we can predict perceived viscosity in a given scene (e.g., pouring onto a plane) very well through a weighted combination of mid-level shape features (e.g. clumping, spiralling, spread) describing the behaviour of the liquid in that scene. However, one problem such models face is that they do not generalise well to novel scenes (e.g. a liquid being stirred). Because the available shape features are highly context dependent, the visual system must somehow select the most relevant and discriminative features in each specific situation. Here, we investigated the selection process through which the visual system determines which features to use for inferring viscosity in each scene. We simulated liquids with seven different viscosities, interacting with eight different test scenes (e.g. stirring, smearing, raining). Each animated scene (10 seconds) had its own unique interactions with the liquid, yielding a wide range of shapes. In experiment 1, observers matched the viscosity in each test scene by adjusting the liquid in a standard match scene (64 viscosity steps, pouring liquids). In experiment 2, observers rated viscosity in the different scenes. In experiment 3, observers rated 20 different shape features for each scene and viscosity. We find that observers show good overall constancy across scenes, with small but systematic over- and underestimations for certain scenes. By comparing these results to those from the shape feature ratings, we identified which features were visually most influential for each individual scene. The results indicate the visual system flexibly re-weights mid-level cues on a scene-by-scene basis to achieve high levels of constancy.
Meeting abstract presented at VSS 2017