Post-recognition Score Analysis

Abstract

Thus far, we’fve mostly discussed EVT in the context of distributions with abstract relationships to visual data. In order to make the transition from theory to practice, we will assume that scores are available as samples drawn from some distribution reflecting the output of a measurable recognition function. The distance or similarity score produced by a recognition function (e.g., a distance calculation between two vectors or a machine learning-based model) is a primary artifact of the pattern recognition process itself, and can tell us much about specific instances of visual perception. Most typically in automatic visual recognition, little, if anything, is currently done with these scores beyond checking the sign to assign class membership, or perhaps a quick comparison against an ad hoc threshold to accept or reject a sample. At a fundamental level, we can ask what exactly a recognition score is and why it is important for decision making. Further, we can model distributions of scores to determine if they were generated by a matching or non-matching process. Once we understand this basic model, we can then extend it to other modes such as score normalization and calibration (see Chapters 4 and 5). In this chapter, we will examine the problem of failure prediction as a case study embodying these concepts.