Understanding the buildings we insure and the associated risk factors for different peril types is key for natcat modelling and risk assessment. While some building attributes are readily available from external data sources such as OpenStreetMap, the coverage is very dependent on the geography and attribute type. It is therefore far from usable in all situations. How can we overcome this problem? A solution is leveraging high resolution aerial imagery and using Computer Vision models to predict the building attributes we are interested in when it is unavailable by other means.

In this presentation, I will show you the results of a prototype we created last year on two different building attributes. While the first one is straightforward to detect, requiring only vertical images and standard CNN models, the second one requires multiple input oblique images. This is why we developed an attention-based Computer Vision architecture that I will be presenting.