Interactive Rendering Of Massive 3D Models

With the growing ability of obtaining information and computing, there have been significant advances in computer-aided-design (CAD), and simulation technologies. Today, we could gener-ate many kinds of complex digital models, including CAD models, scanned models and various scientific simulation data. These massive models play an important role in game, film and indus-trial design. The massive data typically require giga byte size and even larger. Rendering such massive models interactively presents an important challenge to visualization software and system developers, with their inherent focus on interactive, low-latency, and real-time processing.Interactive rendering of massive models is a research direction of computer graphics. With the computing power of CPU and GPU in computer system, this technology could generate dynamic images for massive models interactively with limited memory。Interactive rendering of massive models consists of two parts:preprocessing and interactive rendering. In preprocessing, the input massive models are divided into spatial structure index, and then LOD models are generated with simplification algorithm. In interactive rendering, based on the spatial index and LOD structure, dynamic images are generated with visibility culling, LOD refinement and external data manage-ment.We present a new efficient approach, Adaptive Voxels, for out-of-core construction and view-dependent rendering of very large models on single commodity pc platform. It contributes to mas-sive model rendering in the following aspects:1. We present a multi-level scene-graph structure and file format that could be employed in LOD generating, rendering and out-of-core strategy. Update Traditional scene graph with single BSP tree is costly due to the deep BSP tree. We propose a dual-BSP spatial scene-graph, which can keep the bsp structure of every single model not changed, and adaptively update the entire bsp scene-graph. It guarantees that spatial data structure can still be eciently utilized while interaction.2. Then we present a GPU based adaptive sampling strategy for LOD hierarchy generating which reduce the time cost for preprocessing. The color, normal and depth texture are used to generate the voxelized LOD models. Next, we encode the sample data into a GPU friendly data format which is suitable for realtime rendering.3. We present a new external data based interactive rendering algorithm. We join the LOD model into hardware occlusion culling. And we use LRU based memory strategy for mem-ory/VRAM control. At rendering time, we combine various accelerating techniques together: hardware occlusion culling, out-of-core algorithm and RAM/VRAM management. These techniques make it possible to interactive render the massive models.As a result, our system could render GB size massive model interactively on common PC platform. Contrasted with other render systems, the preprocessing time cost of Adaptive Voxels is much shorter which could be very efficient for deployment.