Speaker: Elias Kristmann
Abstract
Large LiDAR point clouds are commonly stored in the lossless LAZ format, but the sequential prediction and adaptive arithmetic coding used by LASzip make conventional GPU parallelization difficult. We present CuLAZ, a GPU decoder that operates directly on standard LAZ files without requiring changes to the compressed representation. CuLAZ exploits the existing chunk structure of LASzip by decoding independent chunks across GPU warps. Our current implementation supports LAS point data record format 2 encoded with the LASzip version-2 POINT10 and RGB12 codecs and produces output that is bit-for-bit identical to the official LASzip decoder. Across five real-world LiDAR datasets and three GPU systems, CuLAZ decodes up to 540 million points per second and is up to 7x faster than a parallel CPU implementation.