Feature Adaptive3D Point-Based Model Simplification

3D point-based model or point cloud model, as one form of natural representation of3D geometry, takes points as primitives.3D point-based model has a simple structure, and is compact in terms of storage space. Since the point-based representation gets rid of the need for maintaining and processing complex topological information, it is particularly suitable for representing highly detailed and densely sampled3D surfaces.With the advances in3D laser scanning technology and hardware processing power, the use of point primitives for3D surface representation and rendering has been gaining increasing popularity in the past decade. And with the current3D laser scanning technology, a point-based3D model with hundreds of millions of point primitives can be easily acquired. It is still not feasible to render and manipulate those huge point-based models at an interactive speed using commodity computers. Therefore, simplification techniques have been investigated by researchers.In this thesis, we propose a novel bottom-up collapsing approach, a top-down divisive approach and improve a hybrid clustering algorithm for point-based3D model simplification. Unlike most of the previous pure-geometry-driven algorithm, the proposed and improved ones take both the geometrical and texture information into account during the simplification process in order to preserve the most prominent geometrical or textural features at a reduced data budget.In order to preserve as many textural features as possible even after simplification, we extend the color image sharpening scheme to save the textural characteristics of the simplified3D point-based models. This method combines the Laplacian of Gaussian operator with spatial filters that approximate the contrast sensitivity functions of human visual systems. It is sensitive to variations of viewing conditions, and because of the smoothing Gaussian functions, it does not tend to increase noise.Furthermore, we propose a method for out-of-core simplification of gigantic point-based models that cannot be held in memory as a whole. It is based on the discrete Lagrangian optimization technique and yields optimal model quality for any given data reduction ratio.