Abstract
Estimating the three-dimensional structure of surfaces in natural scenes is a fundamental visual task. Previously, we reported that human tilt estimation in natural scenes is tightly predicted by an image-computable Bayes-optimal model grounded in the statistics of natural scenes (Kim & Burge, VSS2017). However, the previous model was limited in two respects: i) it predicted only unsigned tilt estimation (i.e., tilt modulo 180°), and ii) it used only local tilt cues without considering global context. Here, we extend the model to produce signed tilt estimates and to utilize spatial pooling, and we test the predictions of the extended model against newly collected psychophysical data. Each human observer estimated 7200 stereo-image patches that were randomly sampled from natural scenes. Given an image patch, the extended model estimates the Bayes-optimal tilt by computing the mean of the posterior distribution over signed tilt conditioned on the image cue values. The model and human estimates have a similar pattern of bias and variance, and the distributions of model estimates nicely predict the distributions of human estimates for each signed tilt. Next, we found that a simple spatial pooling (i.e., a straight average) of the model's local tilt estimates (i.e., 'global' estimates) provide a better account of the human data than the model's local tilt estimates alone. This result is expected given the spatial correlation of tilt in natural scenes. Taken together, the findings suggest that the human visual system makes the best possible use of image information to estimate local signed tilt and spatially integrates local tilts to estimate 3D surface orientation in natural scenes. Future work will develop a Bayes-optimal model of spatial pooling that incorporates the statistics of tilt correlation in natural scenes.
Meeting abstract presented at VSS 2018