A Space-Carving Approach to Surface Estimation

In this project, the problem of detecting the surface of an unknown arbitrarily-shaped scene from a set of points is considered. This problem is usually formulated as the estimation of a function whose zero set represents the surface. Most existing approaches solve this problem by employing implicit surface representations to recover the surface from the given data points. The proposed approach differs from existing methods as it not only uses data points but also utilizes the additional geometric ray-based information often ignored by earlier formulations. By posing this problem as an infinite-constraint minimization task, a maximummargin based formulation is developed and solved using a kernel-based online subgradient method. Experimental results validate the proposed approach.