Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Perceiving Object Motion with Shadow Cues

Reuben McNair¹, Max Siegel², Kartik Chandra², Joshua B. Tenenbaum², Katherine Rebecca Storrs¹; ¹University of Auckland, ²Massachusetts Institute of Technology

Presenter: Reuben McNair

In biological vision, discriminative models give the best current accounts of psychophysical and neural data, but we lack good tasks which reveal the advantages of generative (or analysis-by- synthesis; AbS) models. A promising candidate is scenes containing shadows: shadows are rarely directly task-relevant and have often been treated as nuisance variables. We present a benchmark that quantifies how shadows are used to perceptually infer object positions and trajectories. Both humans and an AbS model used cues from a shadow to disambiguate an object’s motion (Experiment 1). However, changes in shadow contrast altered the outputs of the AbS model, but not human perception (Experiment 2). Replacing the shadow with a non- shadow-like alternative or breaking the correspondence between ball and shadow yielded stimuli which the AbS model could not represent, while humans continued to coherently respond (Experiment 3). Our data provide systematic, quantitative measurements of human shadow understanding, and a challenging benchmark for machine vision.

Topic Area: Computational Models of Vision & Visual Cortex