Researchers: Avital Steinitz, Ram Vasudevan

Many variants of object recognition consist of optimizing an objective that maps features to some desired output. For example, face recognition may be accomplished by a mapping from eigen-faces to identities. In our work, we wish to establish a generic framework for describing such objectives and for obtaining their (locally) optimal value. We believe that such clean and generic formalization would be applicable in various domains, ranging from object recognition to optimal control.

To start, we work with one standard visibility objective function, in which we break down the computation of the objective to two components; determining which portions of the object are visible, and then integrating a visibility objective over those portions.

The figure below shows the portions of a 2D object that are visible from one particular point.