A facade being the representative face of any building, we can gauge its real estate features by just looking at it.  We naturally deduce from its structure, components, size or condition, diverse properties of the whole edifice without ever going inside it. Being able to model this skill could lead to multiple housing-related applications: automated property valuation, custom insurance contract suggestion, or as a market studying tool.  Deep learning methods have evolved to mimic human perception and they now outperform us on object classification/recognition, artistic style imitation, and even on some video or board games. However, these performances are often at the expense of understandability: depicted as black boxes, their predictions are obtained in an unexplicit or inexplicable way. Recently, great progress in explainable AI is trying to solve this issue.

From here, can we use neural networks to craft an understandable expert of a given city housing market solely from building facades? We propose to use open source assessment data and facade photographs obtained via Google Street View to constitute a dataset. We then train a convolutional neural network at characterizing a property given its facade. Different network architectures are explored to, for instance, predict its construction year within 4 years and its property value with less than 27 kCAD average error. Then, we leverage the learned representation to better understand the phenomena at use in housing assessment: our proposed method reveals different groups within a particular city real estate and their corresponding quantitative, geographic, and visual properties.