UCLA

In this dissertation, we used computational models to answer two questions about human perception. First, what is the underlying computational mechanism of the stereokinetic effect in human structure from motion perception? Second, what is the functional nature of the boundary extension effects in human natural scene perception?

To answer the first question, we extended the motion coherence theory in two-dimensional (2D) space (Yuille and Grzywacz, 1988) and the minimal total motion theory in 3D space (Rokers, Yuille, and Liu, 2006). We framed the underlying computational mechanism as an optimization problem. We proposed that among all the 2D and 3D structure interpretations, the one that gives rise to the minimal and spatially smoothest motion is preferred by the visual system, and such structure interpretation is perceived by human observers. We also found that it is important to take into account the higher order motion spatial smoothness. The computational model we proposed was able to predict human 2D and 3D structure from motion perception in varies scenarios. We concluded that the perceptual ambiguity of structure and related motion can be resolved using the minimal total motion and spatially smooth motion principle alone, and any additional assumptions are not necessary.

To answer the second question, we designed two visual memory experiments that made use of a modified test procedure that allowed us to explore boundary extension in terms of signal detection theory. We asked questions about the perceived viewing distance change between study and test scenes from two different psychological dimensions: in terms of close or wide or in terms of change or no change. We found that a criterion bias could explain the boundary extension effects when we asked the perceived viewing distance change in terms of close or wide. In contrast, both discrimination sensitivity and bias contributed to the boundary extension effects when we asked the perceived viewing distance change in terms of change or no change. Remarkably, these results could be explained in a straightforward manner by the multisource model (Intraub, 2010; 2012), with a simple assumption that the view-angle of a memorized scene widened.