ABSTRACT
In this paper, a new multi-view stereo algorithm is proposed that improves 3-dimensional (3D) reconstruction quality of dense point clouds for weakly textured surfaces. In the proposed algorithm, two contrast-enhancement methods (i.e., Wallis filter and complete histogram equalization) are applied to the original image set, and three depth/normal maps are generated via stereo-matching from the original and two contrast-enhanced sets. Finally, depth in the high-quality depth map is determined by adaptively selecting the depth-per-pixel having the best photometric consistency from the three maps to generate better 3D dense point clouds. For the adaptive selection, the least-cost and least-invalid-depth methods are introduced. Reconstruction completeness and accuracy of the proposed algorithm are examined for the original low-textured and publicly available datasets. Our integrated method generates larger reconstructed regions than those generated only from original or single-contrast-enhanced image sets.
