Closed yxlao closed 5 years ago
@syncle The majority of the improvements comes from I/O: decoupled I/O with processing (before, I/O and processing are interleaved for each point, which is very slow). We are still using the original interpolation method. The improvement cuts the run time from minutes to ~10s (~3s compute, ~7s IO). Open3D was used for point cloud reading. Mainly in this commit: https://github.com/IntelVCL/Open3D-PointNet-Semantic/pull/6/files#diff-229b614579b2ac8792b260103c762ca1
Great! How did you make it much faster?