Optimal Stereo Reconstruction and 3D Visualization.

摘要

In this thesis different aspects of stereo-based 3D reconstruction and 3D visualization and manipulation are investigated. The focus of this research is specifically on sampled object representations (SOR) in which the object(s) is (are) shown using a collection of samples such as discrete pixels, voxels, points, or surface elements. We study different techniques of optimal sampling, and optimal representation and visualization of SORs for conventional and stereo-based 3D applications. In this regard, we first analyze the process of capturing and viewing stereo content on stereo capture and 3D display devices, and propose (unified) optimal sampling models for both the stereo capturing and viewing ends. Specifically, we theoretically show that for a given total resolution, a finer horizontal sampling rate, compared to the usual horizontally-vertically similar sample (pixel) distribution, results in a more accurate 3D estimation and enhanced 3D visual experience. We validate our theoretical results through subjective studies and show that human observers indeed have a better 3D viewing experience with an optimized vs. a non-optimized representation of stereo 3D content. We next study different techniques for compact representation and interactive rendering of sampled-based (point-based) 3D object representations. In particular, we introduce a tree structure, called a multi-section tree, and show how this structure can be used in creating a fully balanced, multi-resolution, hierarchical structure over space (and time) to support interactive rendering of spatial (or spatio-temporal samples) of still (or moving) 3D objects. Using an implicit representation of the multi-section tree and improved, dense hierarchical encoding techniques we can achieve highly compact representations for both still and animated point-based 3D models. We can also achieve an interactive frame rate, and quality rendering of large models, on commodity desktop or mobile devices. Finally, we study 3D interaction and manipulation within a virtual stereoscopic 3D space. In this regard, based on a stereo-based 3D cursor, we have developed some simple 3D tools for manipulating point-based 3D objects. We discuss and through user studies show how this stereo-based cursor can be considered as a generic alternative to its 2D counterpart in virtual stereoscopic 3D space. We also discuss the pointing accuracy of the stereo-based 3D cursor with regard to the optimal sampling model proposed and discussed in the earlier chapters.