Point Clouds for Machine Learning

Principal Investigator: Dr Ioannis Brilakis 
Trimble contact: Robert Banfield 

Project Description:

This mini project aims to explore ways and create a set of basic tools for preparing Point Cloud Datasets (PCD) that are “machine learning friendly”.

PCD are seeing a lot of use in the last decade within machine learning systems. Lately, new tools have emerged (e.g. PointNET) that exploit the advantages of deep learning for PCD and are used to detect spatial components and their relationships. However, researchers need substantial manually pre-process the datasets often in billions of points before they can be used even in high performance computing systems (e.g. via Google’s TensorFlow). Existing PCD processing tools allow users to apply generic filters to evenly down-sample the point clouds, remove noise without considering the context, etc. This renders the resulting down-sampled datasets often lose information along the way, and/or do not compress the point clouds nearly enough when dealing with massive datasets. Ms Ruodan Lu (Trimble sponsored student) will investigate and implement a) a smart PCD compression algorithm for point cloud sub-sampling, taking into account the local context information, and b) a novel 3D feature detection algorithm that can be used to address non-contextual occlusions. This will allow users to automatically down-sampled the point clouds to generate “machine learning friendly” PCD and will also improve the efficiency of the object detection task.