EXTRACTING SEMANTICS FROM 3D AERIAL MAPS – SAAB DYNAMICS

The challenge in this use case is to derive symbolic information from raw 3D maps. A combination of Deep Learning methods for classification and logic-based methods for reasoning about symbolic spatial relations was developed and demonstrated on satellite image data provided by Saab Dynamics (maps of Boden and Stockholm).

This work culminated with the achievement of semantic maps of both cities containing automatically generated symbolic information regarding buildings and other structures (with a classification accuracy of 90%), as well as symbolic spatial relations among these structures. An overall software package integrating the overall work was developed, called SemMap.