Instant Visualization and Editing of Arbitrarily Large 3D Data Sets

The size of 3D datasets continues to grow at a rate that far outpaces improvements in hardware and algorithms. At the macro-scale, 3D Scanners can now capture terabytes of data from country-wide scans at centimeter accuracy; on the micro-scale, they produce petabytes for a one-cubic-millimeter brain scan, with nanometer accuracy. Unfortunately, preparing the raw data for inspecting, processing, and editing still takes days to weeks, which are spent constructing acceleration structures for viewing or for repeated construction and refinement of derivatives like maps, statistics, and segmentations.

This project proposes novel visualization methods for handling arbitrarily large data sets directly, enabling near-instant inspection, processing, and editing—without laborious precomputation of levels-of-detail (LODs) or similar slow preprocessing. The key idea is that a tiny, randomly sampled subset of the data—quickly extracted using modern, consumer-grade storage devices—suffices to initialize an overview. As users navigate the 3D model, we prioritize loading more data and constructing LODs on the fly based on the user's viewpoint, leveraging correlations between 3D regions of interest and their memory locations.

This research aims to replace workflows that currently demand days of processing with a near-instant feedback loop, allowing, e.g., medical researchers, geographers, and surveyors to rapidly develop and refine their insight-and-knowledge-gain-driven processes.