서강대학교

초록/요약

A three-dimensional scene can be separated into two regions: foreground and background. The foreground and background can be defined as the regions in front of and behind the focused plane, respectively. From the four-dimensional (4-D) light field, this thesis proposes a depth estimation method by accumulating binary maps, which are computed by the separation of the foreground and background with the light field re-parameterization. In the proposed foreground and background separation, an optical phenomenon is used: the bundles of rays from the background are flipped on their conjugate planes. With the Lambertian assumption and the gradient constraint, the foreground and background of a scene can be distinguished as a binary map by voting the gradient signs in every angular patch. With the light field re-parameterization, the disparity map can be obtained by simply accumulating the separated binary maps. Finding the extremum index in the existing methods corresponds to finding the zero crossing index in the proposed method. By simply accumulating the separated binary maps, the proposed method can achieve not only high-quality disparity map estimation but also the computational efficiency in terms of the memory usage. Experimental results with synthetic and real images show that the proposed method performs better than existing methods in terms of general error metrics.