Abstract
The computation of visual surface perception is not well documented, mainly because of the complex interactions between its phenomenological components, such as illusory contour, transparency, occlusion, and stereopsis. Here it is hypothesized that these phenomenological elements for visual surface perception are compound phenomena that consist of three computational elements. 1. Contour delineation: If a set of features satisfy a relation that is invariant or preserved under projection, the features appear grouped, as they are likely to constitute the image of an object. Shape is a view-invariant and therefore intrinsic property of an object. Thus, recognizable shape features and elements should be preserved under projection. 2. Optics estimation: The optics estimation process is a scission or splitting of the intensity of a pixel into components. Components are surfaces, cast shadows, illumination, and cast lights. While shape is a view-invariant property in the contour delineation process, reflectance and transmittance are view-invariant properties in the optics estimation process. 3. Depth stratification: By detecting a multiple match or a binocularly-unpaired region, one identifies a location where multiple depth values must be assigned to one direction of sight. If the identified depths or local luminance contrasts are consistent along a contour, they lead to the perception of a visual surface. An advantage in the computational decomposition is simplification of inter-module interactions. The adequacy of the hypothesis can be tested empirically through variations of the depth propagation effect (Perception, 21, 177), and theoretically in ecological optics.
Supported by MEXT and JST.