Multi-view photometric stereo using a normal consistency approach

Scene reconstruction is one of the important problems in computer vision. One approach to scene reconstruction is Photometric Stereo. There the main idea is that we can glean the surface orientation from changes in pixel intensity values. These changes are induced by varying the illumination while keeping the viewpoint fixed. If we want to reconstruct a complete scene with photometric stereo we have to drop the fixed viewpoint assumption and the pixel correspondences are not trivial anymore. In this work we present a novel normal consistency metric for points in 3D space which enables us to find points on the surface without having to explicitly model the pixel correspondences - the problem of pixel correspondences is solved implicitly. We obtain a set of oriented points in a volumetric grid from which a surface can be easily reconstructed. The proposed algorithm thus combines the advantages of classic photometric stereo and multi-view reconstruction methods. It automatically reconstructs a triangle mesh from input images with known viewpoints and illumination directions. If the scene is sampled densely enough the proposed approach is robust against self-occlusion, shadowing and isolated specular highlights.