A Parallel Splatting Algorithm With Occlusion Culling

Splatting is a volume rendering technique that projects and accumulates voxels onto the screen. It is able to incorporate a variety of reconstruction kernels without extra computational overhead, as well as reduce computational and storage costs using a sparse volume representation. Previous splatting algorithms suffered from artifacts because they incorrectly separate volume reconstruction and volume integration. The IASB (Image-Aligned Sheet-Based) splatting overcomes these problems by accumulating voxels onto sheets aligned to be parallel with the image plane. In addition, it introduces a novel approach for splatting to cull occluded voxels using an opacity map, called an occlusion map, that provides a substantial speedup in serial implementations. Parallel approaches to volume rendering are able to overcome the enormous amount of computation required. We present a parallel solution to the IASB splatting approach that leverages its image-aligned volume partition construction. The occlusion culling technique poses difficulties in developing a parallel solution due to its inherent serial nature. Our approach assigns processors to render data closest to the screen first and propagates an opacity map to avoid rendering occluded data. Data communication is substantially reduced making use of occlusion culling, as well as view coherence. We utilize a dynamic load balancing scheme where tasks are partitioned in image space. Our implementation runs on both NOW (network-of-workstations) and MPP platforms. We present results and analysis on a NOW cluster of Pentium II’s.