Hybrid neuro-bayesian spatial contextual reasoning for scene content understanding

Geospatial scene content understanding facilitates a large number of increasingly important applications. These range from tools to help intelligence analysts perform rapid, high-precision identification of urban scene content to other civilian and military security applications such as geospatial queries, functional object level change detection, and mission planning. In this paper, we present initial research results from a multi-faceted approach for determining and understanding scene content. Our approach performs context-dependent probabilistic reasoning on a set of object hypotheses obtained from a suite of individual object detection algorithms. This neurally-inspired reasoning approach improves the quality of object detections within a given scene and enhances scene content understanding by fusing low-level features, identified objects, high-level context, and spatial constraints to more accurately determine the nature of specific scene level targets. We present results from application of our hybrid spatial contextual reasoning approach to a set of objects automatically obtained from an urban scene by a suite of state-of-the art detection algorithms. We demonstrate that reasoning on the individual detector outputs produces improved precision-recall performance over using the detector outputs alone.