Large-Scale Data Management for PRT-Based Real-Time Rendering of Dynamically Skinned Models

Abstract

Computer games and real-time applications frequently adopt mesh skinning as a deformation technique for virtual characters and articulated objects. Rendering skinned models with global shading effects, such as interreflection and subsurface scattering, using precomputed radiance transfer enables high-quality real-time display of dynamically deformed objects. In this approach, we need to precompute radiance transfer for many sampled poses. Resulting datasets reach hundreds of gigabytes, and are orders of magnitude larger than those for a static object. This paper presents simple but effective large-scale data management techniques so that runtime data communication, decompression and interpolation can be performed efficiently and accurately. Specifically, we have developed a mesh clustering technique based on spectral graph partitioning to facilitate interpolation from nearest neighbors and an incremental clustering method for transfer matrix compression. By exploiting additional data redundancies among different sampled poses, we can achieve higher compression ratios with the same fidelity. Our incremental clustering can make the runtime cost of per-frame data decompression and interpolation satisfy a prescribed upper bound. As a result, we can achieve real-time performance using the massive precomputed data and an efficient runtime algorithm.