UNC-CH Technical Report TR02-013 GigaWalk: Interactive Walkthrough of Complex Environments William V. Baxter III Avneesh Sud Naga K. Govindaraju Dinesh Manocha University of North Carolina at Chapel Hill {baxter,sud,naga,dm}@cs.unc.edu http://gamma.cs.unc.edu/GigaWalk Figure 1: DoubleEagle Tanker: This 4 gigabyte environment consists of more than 82 million triangles and 127 thousand objects. Our algorithm can render it 11-50 frames per second on an SGI system with two IR2 graphics pipelines and three 300MHz R12000 CPUs. Abstract We present a new parallel algorithm for interactive walkthrough of complex, gigabyte-sized environments. Our approach combines occlusion culling and levels- of-detail and uses two graphics pipelines with one or more processors. We use a unified scene graph repre- sentation for multiple acceleration techniques, and we present novel algorithms for clustering geometry spa- tially, computing a scene graph hierarchy, performing conservative occlusion culling, and performing load- balancing between graphics pipelines and processors. The resulting system, GigaWalk, has been used to ren- der CAD environments composed of tens of millions of polygons at interactive rates on an SGI Onyx system with two Infinite Reality rendering pipelines. Overall, our system’s combination of levels-of-detail and occlu- sion culling techniques results in significant improve- ments in frame-rate over view-frustum culling or either single technique alone. Keywords: Interactive display systems, parallel ren- dering, occlusion culling, levels-of-detail, Engineering Visualization. 1 Introduction Users of computer-aided design and virtual reality ap- plications often create and employ geometric models of large, complex 3D environments. It is not uncom- mon to generate gigabyte-sized datasets representing power plants, ships, airplanes, submarines and urban scenes. Simulation-based design and design review of such datasets benefits significantly from the ability to generate user-steered interactive displays or walk- throughs of these environments. Yet, rendering these environments at the required interactive rates and with high fidelity has been a major challenge. Many acceleration techniques for interactive display of complex datasets have been developed. These in- clude visibility culling, object simplification and the use of image-based or sampled representations. They have been successfully combined to render certain specific types of datasets at interactive rates, includ- ing architectural models [FKST96], terrain datasets [Hop98], scanned models [RL00] and urban environ- ments [WWS01]. However, there has been less success in displaying more general complex datasets due to sev- eral challenges facing existing techniques: Occlusion Culling: While possible for certain envi- ronments, performing exact visibility computations on large, general datasets is difficult to achieve in real time on current graphics systems [ESSS01]. Furthermore, occlusion culling alone will not sufficiently reduce the load on the graphics pipeline when many primitives are actually visible. GigaWalk: Interactive Walkthrough of Complex Environments Page 1 of 15