Geometrically Motivated Fuzzy Measures and Choquet Integral for Multi-Look Fusion

The fuzzy integral (FI), which is defined and ultimately driven by the fuzzy measure (FM), is a versatile aggregation operator. Data driven approaches tend to learn fixed FIs that may not perform well in new contexts. Herein, we introduce a new method for generating a set of FMs based on the relative camera orientation and field of view of an object's surface in the context of computer vision. This method is used to fuse the detection output from one or more classifiers across multiple views of an object by considering the geometry of the scenario. Furthermore, by incorporating historical data, a detectors' strengths and weaknesses may be incorporated to improve the result. Synthetic examples are provided and worked out for important extreme operating conditions to demonstrate our fusion process and its desirable behavior.