How can we apply machine learning to large pointclouds? This is important for working with for drones, self-driving cars, and other places where we use lasers and LiDAR.
Traditional convolution layers are designed to exploit a fixed and regular grid. However, unstructured data like 3D point clouds don’t fit into a grid so it can be hard to apply methods that normally work well on 3D images.
So how can we deal with huge datasets of sparse data? I'll go over some of the clever approaches that have been published, these include graph based, clustering, kd-trees, nearest neighbour convolutions and more. Some of these techniques seemed really clever to me so I aim to show you why they are interesting.
This is a medium to advanced level talk: I won't describe what convolutions are but I'll try to go slow enough that you get the problem and get "ah hah" moments when you understand the main approaches.
A lot of the time I've spent on this topic has been part of a project designed and funded by Paul Wighton from 3dimageautomation.com.au. And it's been fun so thanks Paul!
If you like to read up on a topic beforehand check out the introduction of this paper for a good overview of the state of the art: https://arxiv.org/abs/1803.07289
Please arrive before 6pm to ensure entrance