Visual Utility – A Framework for Focusing Computer Vision Algorithms

Abstract

The real world is a rich environment, fraught with complexity. To be robust in this complex environment, computer Vision algorithms that operate in Unstructured Environments (VUE) tend to use large amounts of data or complex modeling. Unfortunately, these algorithms also require signicant computational resources. In this thesis, we examine a visual utility framework that we show is used regularly in an ad hoc manner. This framework uses visual utility estimators to speed up VUE algorithms with minimal performance degradation by focusing those higher level algorithms on the most relevant imagery. We formally dene this framework, show that it has a submodular structure and discover under what conditions using it is valuable in practice. We nd that visual utility approaches are most eective when using fast, task specic visual utility estimators and the VUE task is computationally expensive, We also introduce SCATAT, a cascade building algorithm. SCATAT takes advantage of the submodular structure of the visual utility framework in order to build a near optimal cascade that trades o task performance and processing requirements explicitly. We then validate this algorithm in an experimental case study on object detection. Finally, in theoretical case studies, we prove that two seminal cascade algorithms are special cases of our visual utility framework. We show they optimize a submodular visual utility function, explaining their high observed performance in practice.