Exploiting automatically inferred constraint-models for building identification in satellite imagery

The building identification (BID) problem is based on a process that uses publicly available information to automatically assign addresses to buildings in satellite imagery. In previous work, we have shown the advantages of casting the BID problem as a Constraint Satisfaction Problem (CSP) using the same generic constraint-model to represent all problem instances. However, a generic model is unable to represent with the necessary precision the addressing variations throughout the world, limiting the applicability of our previous approach. In this paper, we describe the end-to-end process used to solve the BID with a new model-generation technique that uses instance-specific information to automatically infer a representative constraint model of the BID. This inferred model is used by our custom constraint solver to identify buildings in satellite imagery more efficiently and with higher precision than using a single model. We evaluate our approach on El Segundo California, and empirically demonstrate its effectiveness for geographic areas larger than previously tested. We conclude with a discussion of the generality of our approach, and present directions for future work.