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

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Measuring Depth Inference in the Ames Window Illusion

Emil Stroecker¹, Fan L. Cheng¹, Nikolaus Kriegeskorte¹; ¹Columbia University

Presenter: Emil Stroecker

Under the absence of binocular depth information, humans must resolve ambiguous monocular depth cues to determine one likely scene geometry from an infinite set of possible geometries. The brain's ability to heuristically choose the most plausible interpretation of a scene can be studied using synthetic stimuli such as the rotating Ames Window — a rotating flat object giving humans a false impression of depth. While past work found that deep neural networks (DNNs) share human biases in depth estimation, that finding is confined to naturalistic scenes, rather than ambiguous objects. To determine the strategy DNNs use to process depth in such stimuli, we presented three DNNs with Ames Window stimuli to examine if they are equipped with similar biases to resolve geometric ambiguity. We find no meaningful similarities between models, indicating that a DNN's ability to estimate depth in naturalistic scenes does not imply the ability to make depth judgments in ambiguous and artificial conditions as it would for humans.

Topic Area: Computational Models of Vision & Visual Cortex