Semantically Derived Geometric Constraints for MVS Reconstruction of Textureless Areas

Conventional multi-view stereo (MVS) approaches based on photo-consistency measures are generally robust, yet often fail in calculating valid depth pixel estimates in low textured areas of the scene. In this study, a novel approach is proposed to tackle this challenge by leveraging semantic priors into a PatchMatch-based MVS in order to increase confidence and support depth and normal map estimation. Semantic class labels on image pixels are used to impose class-specific geometric constraints during multiview stereo, optimising the depth estimation on weakly supported, textureless areas, commonly present in urban scenarios of building facades, indoor scenes, or aerial datasets. Detecting dominant shapes, e.g., planes, with RANSAC, an adjusted cost function is introduced that combines and weighs both photometric and semantic scores propagating, thus, more accurate depth estimates. Being adaptive, it fills in apparent information gaps and smoothing local roughness in problematic regions while at the same time preserves important details. Experiments on benchmark and custom datasets demonstrate the effectiveness of the presented approach.