For video footage from past you can visit the individual event pages, or go to our YouTube Channel

To filter by event category, click on the event category link in the table below or use the menu on the right.

List of Past Events

Probabilistic inference of 3D shape from line drawings

Seha Kim

Monday, September 15, 2014, 12:00pm - 07:00pm

Graduate Student, Rutgers University, Department of Psychology

Copy to My Calendar (iCal) Download as iCal file
 

Human vision has the ability to perceive 3D shape from line drawings although most of the area in a line drawing image except the contours has no information about depth at all. This study looks for the mechanisms underlying inferences about 3D shape from line drawings. First, we conducted and analyzed a psychophysical experiment to understand the nature of human interpretation of line drawings. We showed subjects line drawings with two ''probe dots'' superimposed at locations in the drawing and asked them which appeared closer. We found that the depth judgments propagated spatially with related to the distance of probe pairs from the T-junction in line drawings. This result supports that the belief in relative depth from the perceived surface is not deterministic as in traditional “junction catalog” methods but rather probabilistic based on the interaction of local T-junction structure and non-local contour structure. To explain such results, we suggest a computational model that interprets line drawings into probable 3D surfaces in a Bayesian framework. This model numerically estimates a likelihood function, assuming a generative model which generates random line drawings by inflating 3D surfaces of biological shapes from skeletons and projecting the surfaces from random viewpoints. The relation between the contour segments in the random line drawings, which are tabulated via various local and non-local features, and  local 3D surfaces of the inflated shapes are kept in a database, in order to search for the similar contour segments and to sample the probable local surfaces when a novel line drawing is observed. The samples of surface orientations allow us to estimate the posterior distribution of depth throughout the line drawing. Finally this posterior can be integrated to give an estimate of the relative depth of the probe points, which can be compared directly to subjects' judgments.Such 3D shape inference from this model, the predicted certainty in depth difference for the probe points on the stimuli in the experiment, was perceptually correct with the human responses, confirming that 3D shape perception from line drawing is probabilistic and influenced by both local and non-local cues together.

Seha Kim