Data Challenges with 3D Computer Vision

Abstract

A wide range of end user and industrial applications rely on accurate 3D scene representation. Automated 3D modelling from optical sensors such as LIDAR scanners, stereo or RGBD cameras became inevitable since this enables to replace the time consuming manual 3D modelling process. However, optical sensors deliver extremely high amount of data which is required to be processed to information. An 8 bit stereo system with 9MPx colour images delivers 4GB of data per second. Being able to make sense out of this enormous flow of data is a huge and ongoing research and development task. This talk will give an overview of the challenges which naturally involve efficient search, storage and scalable algorithms for information extraction from 3D data. Also an overview of practical implications on 3D reconstruction of cities, autonomous driving in public and industrial facilities, or inspection and monitoring as part of security strategies will be presented.