Using Computer Vision for the Automatic Classification of Building Facades

The growth in data availability has led to an increase in the number of studies tackling different urban problems, including accessibility, walkability, and the impacts of climate change on communities. Despite this growth, however, certain studies are still limited by a lack of data that accurately describes the built environment. Such a data scarcity scenario creates opportunities for developing new computational frameworks that leverage and combine already collected data to extract new urban features. This thesis then presents an innovative framework called BuildingSurfaces that employs multi-scale training and semantic segmentation techniques to accurately identify building elements and classify their primary exterior material. We use labeled data from three major cities, combined with street-level imagery, to iteratively train a segmentation model that can achieve a classification accuracy of 92%. The principal, steps of our work can be summed up as follows: 1. We a detailed survey on the availability of building data information across the US 2. We propose a computational framework for the integration of building data and streetlevel imagery 3. We present a detailed experimental evaluation of our segmentation model 4. We make our data available so that researchers can build on top of our efforts