Real estate building detail enrichment with street view image prediction

Abstract

Within the domain of urban analytics, fundamental aspects defining a real estate property include its construction year, number of floors, and building use. However, it is essential to acknowledge that obtaining accurate and up-to-date information regarding those three attributes can be a challenging and costly endeavor due to diverse data sources, privacy concerns, or proprietary data ownership. This research addresses these challenges by exploring an innovative approach leveraging street view image data to derive key attributes of real estate properties. The inherent visual cues present in images can be- analyzed and processed using advanced computer vision techniques such as Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP), offering a cost-effective and scalable solution for obtaining essential attributes that might be challenging to acquire through traditional means. The scope of this research covers the entire machine learning pipeline, from data acquisition and processing, model training and evaluation, to service deployment. In the pursuit of an effective Minimum Viable Product (MVP), simplicity, computational efficiency, scalability, and modularity are crucial. Results suggest that while determining floor count and building use based on visual cues from images emerges as a visually intuitive task, estimating the construction year poses a more intricate challenge. These insights underscore the potential of image-based data and machine learning in urban data acquisition.