AA Point Cloud

Point cloud supervoxels

Railway point cloud object detection

Immersive visualisation of railway point clouds

Semantic segmentation of point cloud catenary arches

Boosting PointNet++ performance

PointNet++ has been a major milestone in the area of applying deep learning to point clouds. Though impressive results can be obtained, the computational load, especially on the CPU side is high. Luckily, after a quick internet search an elegant solution was found which can speed things up by a factor two. Credits for this solution go to the blog post by Open3D. During the interpolation stage PointNet++ uses a nearest neighbout search, the original implementation uses a brute force approach to finding nearest neighbours. The solution put forward by Open3D is by using a k-d tree to speed up the nearest neighbour search. This post will show how this can be implemented using the nanoflann library. This library is a C++ header only library for building k-d trees.