Accurate multi-view stereo by selective expansion

Abstract

We present a multi-view stereo method for robust and efficient dense modeling based on selective expansion. The core is that the proposed method progressively refines pixel depths by expanding pixels that are selected with small photo-consistency costs to their neighbors, which is more efficient than conventional PatchMatch methods. Besides, a depth refinement process including mapping and filtering against neighboring frames is used to further improve the accuracy of estimated depths. In contrast to previous methods, our method has low computational complexity and doesn't need any optimization algorithm. The accuracy of proposed method is evaluated quantitatively and qualitatively on benchmark data.