Learning non-local features for 3D reconstruction of buildings

The goal of this project is to help automate the process of scanning buildings with consumer digital cameras. Currently, fully automated scanning with a commercial camera produces inaccurate scans, while accurate scans require significant manual effort on each individual photograph (of which there are many) of the building to be scanned. We plan to use modern machine learning techniques to reduce the human labor required to create very accurate 3D scans of buildings. The partner organisation, Butterwick Projects Ltd., will use the solution we develop to scan buildings with poor insulation, then manufacture insulated panels offsite that attach to the outside of the old buildings’ walls and roofs. Since this will be done in a factory, it will be very efficient, and won’t significantly disturb the current occupants. This is also why high accuracy is important – the panels need to fit the building very accurately despite being built offsite.

Faculty Supervisor:

Eleni Stroulia

Student:

Logan Gilmour

Partner:

Butterwick Projects Ltd.

Discipline:

Computer science

Sector:

Construction and infrastructure

University:

University of Alberta