Scalable Exploration of 3D Massive Models

Abstract

This thesis introduces scalable techniques that advance the state-of-the-art in massive model creation and exploration. Concerning model creation, we present methods for improving reality-based scene acquisition and processing, introducing an efficient implementation of scalable out-of-core point clouds and a data-fusion approach for creating detailed colored models from cluttered scene acquisitions. The core of this thesis concerns enabling technology for the exploration of general large datasets. Two novel solutions are introduced. The first is an adaptive out-of-core technique exploiting the GPU rasterization pipeline and hardware occlusion queries in order to create coherent batches of work for localized shader-based ray tracing kernels, opening the door to out-of-core ray tracing with shadowing and global illumination. The second is an aggressive compression method that exploits redundancy in large models to compress data so that it fits, in fully renderable format, in GPU memory. The method is targeted to voxelized representations of 3D scenes, which are widely used to accelerate visibility queries on the GPU. Compression is achieved by merging subtrees that are identical through a similarity transform and by exploiting the skewed distribution of references to shared nodes to store child pointers using a variable bit-rate encoding. The capability and performance of all methods are evaluated on many very massive real-world scenes from several domains, including cultural heritage, engineering, and gaming.