Ideally, each region is only loaded, processed and unloaded once. Ordering of points: For out-of-core algorithms, a certain locality between points is beneficial since it reduces the need to swap data frequently from external storage.Overlaps between input files: Non-overlapping files can be converted in parallel, and the results can be merged afterwards.In our experience, the following properties of the input data have a significant impact on the LOD generation: Figure 2 illustrates individual octree nodes and the subsets of points that are stored inside. The advantage of the replacement approach is that it would allow us to compute representative subsets with baked-in anti-aliasing, similar to mip maps, however we chose the additive approach at this time because it is faster to generate and render, and requires less memory. Alternatively, one could use a replacement scheme where higher levels of detail replace lower levels. Additive scheme means that during rendering, higher levels of detail contain additional points that are rendered together with points in lower levels of detail. These octree structures populate each node with a subset of the full data set, and the combination of all nodes yields the original data set without duplicates. The LOD structure that we target with our method is a variation of a layered point cloud that uses an octree with an additive scheme, as used by Potree and Entwine. Processing pipeline: An out-of-core hierarchical counting sort quickly generates chunks of suitable size which can then be indexed in parallel, and eventually merged into a single octree.
0 kommentar(er)
