Reformulating Constraint Models Using Input Data

Consider the problem of mapping postal addresses to buildings in satellite imagery using publicly available information, defined as the Building Identification (BID) problem in [1]. This problem takes as input a bounding box that defines the area of a satellite image, buildings identified in the image, vector information that specifies streets in the image, and a set of phone-book entries for the area. The task is to find the set of possible address assignments for each building. In [1], we showed how the task can be framed as a Constraint Satisfaction Problem (CSP), which we solved with an existing solver in [1] and a custom solver in [2]. The CSP is given by P = (V,D, C) where V is the set of buildings, D the set of their respective potential addresses, and C a set of constraints that describe the physical layout of the buildings on the map and address numbering strategies. In the context of a web application, a typical BID scenario is as follows. A user, presented with a map such as a Google map, either selects a specific building in an area of interest and requests the address of the building, or he/she provides an address and requests the buildings that could have this address. This process is repeated for millions of areas throughout the United States. To answer the entire spectrum of user queries, this application needs to contend with the slight addressing variations found in cities across the US. For example, some cities adhere to a block numbering scheme where addresses increment by a fixed factor (i.e., 100 or 1000) across street blocks while others do not. The direction in which addresses increase also varies, in some cities this occurs to the east while in others it is to the west. In other cities, addresses along East-West running streets increase to the West in one part of town but to the East in the other. Finally, expanding this application to support the rest of the world would require the set of constraints to model new addressing characteristics not seen in the US. The globalization of addressing across continents ensures that some general guidelines are followed, but this standardization is typically met with regional/cultural customization such as the districting in Venice or the historical numbering seen in Japan. The creation of individual models, for each city in the