Abstract

With the popularity of points as graphics primitives, it is
important to handle large-scale point sets that exceed available
in-core (main) memory. In particular, high-performance
level-of-details (LODs) visualization from
out-of-core is a challenging problem. In this context we
present a novel point-splatting approach, short XSplat, that
breaks the main memory barrier. It is based on a paginated
multiresolution point hierarchy and virtual memory mapping.
The main contributions are a novel block-based
sequential multiresolution point hierarchy, an efficient
LOD-block paging mechanism and dynamic mapping into
video-cache. XSplat is scalable by using sequentialized
data structures, and it seamlessly bridges the disk-, mainand
video-memory sub-systems. Experiments demonstrate
the quality and efficiency that is achieved by XSplat.