Abstract
Accretion/deletion is traditionally believed to unambiguously specify depth order, by assigning ground status to the accreting/deleting side. However, accretion/deletion can also result from self-occlusion of a 3D object rotating in depth. Thus an accreting/deleting region can also be perceived as being in front (Froyen et al, 2013, JOV). This phenomenon has been a focus of study recently (Tanrikulu et al., 2013, 2014, 2015, VSS), where it was shown that the perception of 3D columns rotating in front can occur in regions that should be unambiguously in back according to traditional accounts of accretion/deletion. Specifically, 3D rotation was perceived in one set of regions despite the constant-speed motion of the textures (which is clearly inconsistent with 3D rotation). It was also found that increasing the degree of convexity of the boundary on one side makes the accreting/deleting region on that side more likely to be interpreted as a 3D object rotating in front. Current computational models cannot accommodate these findings. Here, we present a computational model of relative depth and motion interpretation that accounts for the rich interaction described above. It is a Bayesian model that estimates the relative probability of interpreting accreting/deleting textures either as a flat surface translating behind or a curved 3D object rotating in front. The model takes into account the speed profiles of the accreting/deleting textures and the geometry of the occluding contour in terms of its convexity (treated as a quantitative factor). The effect of accretion/deletion is incorporated into the model through qualitative constraints which are then combined with the quantitative information about contour convexity and the speed profile of the texture. In this way, the model can account for variety of phenomena described in the literature that cannot be explained by existing computational models of accretion/deletion and relative-depth judgments
Meeting abstract presented at VSS 2016