End to End Analysis of Large 3D Geospatial Datasets with GPUs

22^nd January 2019, 13:00 Ashton Lecture Theater
John Murray
Fusion Data Science and University of Liverpool

Abstract

Location intelligence is key to understanding many aspects of modern life including property insights, environmental monitoring, disaster management and prevention, traffic flows, customer behaviour etc. In order to build deep learning models, large volumes of data from satellite images, 3D sensors, GPS enabled devices, social media and many other sources need to be merged using computationally intensive coordinate conversion and matching.

In Europe, the property insurance sector has been disrupted by growing use of comparison websites which requires real time quotations to be provided. Fusion Data Science has developed a solution, using 3D deep learning CNNs, to estimate risk factors from colour 3D virtual models of individual properties.

The talk will cover the entire process, using GPUs, from processing raw data, merging source, generating & labelling voxel cubes for training, building the models and inference.

The session will conclude with a demonstration with live data.