Abstract
Perfectly mirrored objects distort the image of the scenery according to the surface curvature of the object. It is theoretically impossible to completely recover the shape of a mirrored object (e.g. from a photograph) because there is an infinite number of possible combinations of illumination and surface properties which lead to the same appearance. Despite this, the human visual system seems to be remarkably adept at constraining the possible interpretations. Based on previous work (Fleming, Torralba, and Adelson. Journal of Vision, 4(9), 2004), this contribution presents two different methods for analysing images of mirrored objects to recover certain properties of 3D shape. We constrain the problem by assuming isotropic contrast information to be present in the surrounding scene. Our first method is a mathematical approach, based on the structure tensor. In this context, the eigenvectors of the tensor tell us the orientation of curvature and the eigenvalues of the tensor give us information about the anisotropy of curvature. Our second method is a biological motivated approach, based on Gabor filters. We apply an iterative refinement in a simple model of cortical feedforward/feedback processing (Neumann and Sepp, Biol. Cyb., 81, 1999). Context information is collected by cells with long-range lateral connections (bipole cells). This information is fed back to enhance regions where local information matches the context. Our approach shows that under the assumption of isotropic oriented contrast information in the reflected world, it is possible to recover two characteristic curvature properties of mirrored objects: (i) the direction of maximal and minimal curvature and (ii) the anisotropy of curvature (ratio of maximal and minimal curvatures). We further demonstrate that the model performs well even if the assumption is violated to a certain degree.