This dissertation presents a complete pipeline of algorithms for the challenging task of visualizing point clouds generated by multi-view stereo (MVS) algorithms from uncalibrated images, which often have relatively inferior quality characterized by the sparsity and high irregularity in the spatial distribution of points, as well as location errors and noise associated with low-precision 3D coordinates.;To deal with these problems, we first apply statistical methods for the nearest neighbor searching, outlier removal and coordinate filtering to process the raw point clouds for improving data quality. Next, for object surface reconstruction, we have developed multiple solutions to match surface complexities of the target objects. For the case of simple and smooth objects, we extract the surface using a robust Poisson surface reconstruction algorithm and polygonize the implicit representation into triangles. For the case of complex scenes, we introduce a novel meshless method to approximate the surface using adaptively-sized elliptical surfel discs.;To visualize the objects with coherent and photorealistic texture appearance, we stitch the images photographed from multiple views onto the primitives of the 3D geometric model (i.e., triangles or surfel discs) using a multi-view texture mapping strategy. By optimizing a Markov random field (MRF) energy, this strategy selects the best viewing directions for each primitive while preserving the coherence between neighboring primitives and thus reducing the color disparity between them. To further eliminate the color discrepancy along the boundary of two adjacent regions mapped to different textures, we apply a feature-based color alignment method for all images pairwisely by selecting the optimal root reference image and the shortest color transfer paths.;We evaluate related methods in the literature and compare our results with the state-of-the-art. We show that our methods can handle the low-quality 3D point clouds data better than the previous ones.
展开▼
