Fast Scalable k-NN Computation for Very Large Point Clouds

dc.contributor.authorSpina, Sandroen_US
dc.contributor.authorDebattista, Kurten_US
dc.contributor.authorBugeja, Keithen_US
dc.contributor.authorChalmers, Alanen_US
dc.contributor.editorHamish Carr and Silvester Czanneren_US
dc.date.accessioned2013-11-08T10:32:02Z
dc.date.available2013-11-08T10:32:02Z
dc.date.issued2012en_US
dc.description.abstractThe process of reconstructing virtual representations of large real-world sites is traditionally carried out through the use of laser scanning technology. Recent advances in these technologies led to improvements in precision and accuracy and higher sampling rates. State of the art laser scanners are capable of acquiring around a million points per second, generating enormous point cloud data sets. These data sets are usually cleaned through the application of numerous post-processing algorithms, like normal determination, clustering and noise removal. A common factor in these algorithms is the recurring need for the computation of point neighborhoods, usually by applying algorithms to compute the k-nearest neighbours of each point. The majority of these algorithms work under the assumption that the data sets operated on can fit in main memory, while others take into account the size of the data sets and are thus designed to keep data on disk. We present a hybrid approach which exploits the spatial locality of point clusters in the point cloud and loads them in system memory on demand by taking advantage of paged virtual memory in modern operating systems. In this way, we maximize processor utilization while keeping I/O overheads to a minimum. We evaluate our approach on point cloud sizes ranging from 50K to 333M points on machines with 1GB, 2GB, 4GB and 8GB of system memory.en_US
dc.description.seriesinformationTheory and Practice of Computer Graphicsen_US
dc.identifier.isbn978-3-905673-93-7en_US
dc.identifier.urihttps://doi.org/10.2312/LocalChapterEvents/TPCG/TPCG12/085-092en_US
dc.publisherThe Eurographics Associationen_US
dc.subjectI.3.6 [Computer Graphics]en_US
dc.subjectMethodology and Techniquesen_US
dc.subjectGraphics data structures and data typesen_US
dc.titleFast Scalable k-NN Computation for Very Large Point Cloudsen_US
Files
Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
085-092.pdf
Size:
794.07 KB
Format:
Adobe Portable Document Format