Architectural scene reconstruction from multiple images is one of hotly pursued topics in computer vision community in recent years, and finds its application in a variety of fields such as architecture heritage preservation, digital city-scale modeling, and virtual reality. However, due to various adverse factors,such as illumination variation, perspective distortion, poorly textured regions,most existing stereo methods usually could only recover some incomplete scene structure. In addition, the computational efficiency is also a major concern when the high resolution images are used.This thesis focuses primarily on the reconstruction completeness and computational efficiency, two key issues in architectural scene reconstruction. The main results include: 1. We proposed an effective method to obtain complete depth maps through pixel-level and region-level match propagations. The pixel-level match propagation can rapidly obtain accurate depths for those regions with abundant textures and provide effective constraints for the subsequent region-level match propagation, which is a follow-up of the former one to infer depths for poorly textured regions and highly slanted surfaces. The two propagations are both formulated and worked out under the energy minimization framework, which provides ample flexibility to tackle the problems in a unified manner. 2. We proposed a new energy-based plane-fitting method for architectural scene reconstruction from multiple images. Based on reconstructed initial sparse spatial points, the method can effectively and efficiently extract more reliable spatial planes to explain the scene structures through two steps: generating reliable candidate planes using spatial point match propagation and globally optimizing spatial planes by incorporating plane priors. Our proposed method is shown able to adequately overcome the two drawbacks of the traditional multi-model fitting methods, incomplete reconstruction and low computational efficiency under high resolution images. 3. We proposed an effective multi-view piecewise planar stereo method based on multi-level energy optimization. By incorporating the prior knowledge of scenes and geometric constraints in terms of carefully designed energy potentials,the method can not only recover complete depths for low textured regions or slanted surface, it can also handle the inevitable inaccuracy problem of initially segmented superpixels, a severe headache for the traditional piecewise plan...
修改评论