High Quality Texture Mapping for Multi-view Reconstruction

In this paper, a new method for mapping textures onto a 3D model produced by a multi-view reconstruction pipeline is proposed. A Markov Random Field (MRF) is constructed so as to define each triangle face as a node. An optimal labeling is estimated by globally minimizing an energy function defined on the MRF using graph cuts algorithms. This labeling assigns exactly one view to each triangle face. Textures are extracted by back-projecting each assigned view to the corresponding face. Color discontinuities among adjacent faces are minimized using a global correction method followed by a local Poisson editing procedure. The proposed method can cope with both image noises and out-of-focus blur in a simple way, which prefers views with clear details of the scene. Comparing with two state-of-the-art methods, this new method can generate textures with higher quality. A simple but efficient strategy is also proposed to further improve the quality of textures: use different image sets for reconstruction and texture mapping. Experimental results show that by applying this strategy, the quality of textures could be significantly improved, while the total computation time only increased slightly.