Enforcing Spatially Coherent Structures for Accurate 3D Shape Recovery

Usman Ali, Muhammad Tariq Mahmood

Abstract: Most of the depth optimization schemes in shape from focus (SFF) consider neighborhood information which, unfortunately, over-smooths depth edges. Further, usually, no additional information about the scene is incorporated while estimating depth. In this work, we tackle the first issue by employing nonconvex penalty that preserves depth edges effectively. For the second issue, we design a novel structure-based guidance map that takes mean of image intensities along optical axis in image sequence. The proposed framework fuses information of guidance map, iteratively updated depth map and the structural similarity between them. Nonconvex objective function has been solved through majorize-minimization algorithm. Experiments conducted on synthetic and real image sequences provide excellent results which reveal effectiveness of the proposed method.