Multi-View Depth Estimation with Color-Aware Propagation and Texture-Aware Triangulation

PatchMatch-based multi-view stereo relies on photometric consistency to estimate the depth map for each view. However, due to the matching ambiguity introduced by photo-consistency metrics, it often lacks completeness on textureless areas. In this paper, we propose a novel Multi-View Stereo algorithm for the multi-view depth map estimation problem. First, assuming local areas with similar colors are on the same piecewise plane, depth hypotheses are propagated with color-aware weights during PatchMatch propagation. Second, we propose a texture-aware triangulation to generate reliable planes to approximate the textureless regions, which helps to decrease the matching ambiguities in textureless areas. Moreover, a probabilistic filter is proposed aiming to optimize the estimated depth map with small speckles. The effectiveness of our method has been validated on the ETH3D benchmark. Experiments of our method show competing performance against the state-of-the-art.