Applications of Computer Vision in Evaluating the Effects of New Housing Projects

Abstract

Cities are laden with visual clues. Tapping on the large volume of street view imagery (SVI) made available in the last decade, we investigate how modern computer vision tools can characterize the visual quality and linguistic diversity of cities and leverage on these novel metrics to study the impact of new housing projects.

Streets form a public space and how they look plays an important role in shaping how walkable they are, how safe people perceive them to be, and the general quality of living in the urban environment. To provide useful metrics to quantify the quality of streets, we construct a scalable process with state-of-the-art machine learning models to generate second-order metrics which capture both physical and perceptual features in an urban environment. Recognizing that the abundance of linguistic features littered across streetscapes gives us clues about underlying individual and social preferences in streetscapes, we also seek to quantify the linguistic diversity in cities. To that end, we construct a language detection model supporting English, Swedish, Arabic and Chinese that outperforms existing optical character recognition (OCR) tools. We evaluate visual interpretability with gradient-weighted class activation maps (Grad-CAM) and find that our model is both accurate and interpretable. We apply these tools to our case study of Stockholm and find intuitive spatiotemporal characterizations of the city. We also advance the application of these metrics by using them in a difference-in-difference (DID) setting to study the effects of newly completed housing projects on the built environment. We find that these projects generate spillover effects, as evident in the increase in enclosure and linguistic diversity in their immediate surroundings.