Keywords
line drawing interpretation, 3D shape inference, Bayesian modeling, depth perception, shape perception
Abstract
Human depth comparisons in line drawings reflect the underlying uncertainty of perceived 3D shape. We propose a Bayesian model that estimates the 3D shape from line drawings based on the local and non-local contour cues. This model estimates the posterior distribution over depth differences at two points on a line drawing. The likelihood is numerically computed by assuming a generative model, which generates random 3D surfaces and, via projection, random line drawings. The 3D surfaces are inflated from random skeletons and projected into line drawings. Given a novel line drawing, the model samples probable local surfaces based on the relations between local 3D surface patches and corresponding 2D contour segments. Then, the likelihood function of depth differences is estimated from the distribution of probable surface orientations. The prior is modeled as a Gaussian using known human biases in depth perception, such as slant-underestimation and closer lower-region in figure/ground organization. This model predicts the probabilities assigned to depth differences between two points on line drawings from the posterior on depth differences. These probabilities were consistent with human responses, showing that the model accounts for human interpretation of line drawings. This model encodes the uncertainty in 3D shape interpretation from line drawings, simulates the propagation of depth information from local and global contours, and provides a tool for testing the scope of cues in 3D shape inference.
Start Date
13-5-2015 3:15 PM
End Date
13-5-2015 3:40 PM
Session Number
02
Session Title
Shape and Form
Bayesian modeling of 3D shape inference from line drawings
Human depth comparisons in line drawings reflect the underlying uncertainty of perceived 3D shape. We propose a Bayesian model that estimates the 3D shape from line drawings based on the local and non-local contour cues. This model estimates the posterior distribution over depth differences at two points on a line drawing. The likelihood is numerically computed by assuming a generative model, which generates random 3D surfaces and, via projection, random line drawings. The 3D surfaces are inflated from random skeletons and projected into line drawings. Given a novel line drawing, the model samples probable local surfaces based on the relations between local 3D surface patches and corresponding 2D contour segments. Then, the likelihood function of depth differences is estimated from the distribution of probable surface orientations. The prior is modeled as a Gaussian using known human biases in depth perception, such as slant-underestimation and closer lower-region in figure/ground organization. This model predicts the probabilities assigned to depth differences between two points on line drawings from the posterior on depth differences. These probabilities were consistent with human responses, showing that the model accounts for human interpretation of line drawings. This model encodes the uncertainty in 3D shape interpretation from line drawings, simulates the propagation of depth information from local and global contours, and provides a tool for testing the scope of cues in 3D shape inference.