How to build a building classifier: A data-driven approach to characterize building functions from streetview images with computer vision and machine learning in the city of Amsterdam

Organised by Laboratory of Geo-information Science and Remote Sensing

Mon 26 November 2018 09:30 to 10:00

Venue Lumen, gebouwnummer 100
Room 1

By David Swinkels

More and more information about cities is being captured by satellites, planes, drones, smartphones and recently from cameras on cars. Pictures from cameras on cars can help to solve automatic urban land use mapping, i.e. to detect shops, offices, industrial or residential buildings. Streetview photography platforms, such as Google Street View and Mapillary, have large datasets of ground-level images available. Advances in machine learning and computer vision, i.e. convolutional neural networks, make it possible to classify building functions from streetview images. The research in this thesis showed that building functions can accurately be predicted with streetview images. Also, it was observed that if buildings were more recently built and streetview images had a smaller distance from the camera to the building, the prediction accuracy of the building functions was significantly higher.