Robust Fusion: Extreme Value Theory for Recognition Score Normalization

Abstract

Abstract. Recognition problems in computer vision often benefit from a fusion of different algorithms and/or sensors, with score level fusion being among the most widely used fusion approaches. Choosing an appropriate score normalization technique before fusion is a fundamentally difficult problem because of the disparate nature of the underlying distributions of scores for different sources of data. Further complications are introduced when one or more fusion inputs outright fail or have adversarial inputs, which we find in the fields of biometrics and forgery detection. Ideally a score normalization should be robust to model assumptions, modeling errors, and parameter estimation errors, as well as robust to algorithm failure. In this paper, we introduce the w-score, a new technique for robust recognition score normalization. We do not assume a match or non-match distribution, but instead suggest that the top scores of a recognition system's non-match scores follow the statistical Extreme Value Theory, and show how to use that to provide consistent robust normalization with a strong statistical basis.

Data fusion for multi-lesion Diabetic Retinopathy detection

“…In order to improve this result, we are investigating possible alternative fusion solutions such as [16] which maps the distances to a normalized probability space. Furthermore, we are investigating the effects of poor quality retinal images in a DR screening.…”

Multi-class Open Set Recognition Using Probability of Inclusion

“…For recognition problems in computer vision, EVT has been demonstrated to be a powerful explanatory theory [43] and an effective tool for statistical modeling [7,22,21], including fitting probability estimators [44,42,41]. The most relevant work in EVT modeling is the multi-attribute spaces approach of Scheirer et al [42], which applies EVT calibration over binary classifiers for visual attribute assignment.…”

“…And since that boundary is defined by the training samples that are effectively extremes, we conclude that proper models for efficient SVM calibration should be based on extreme value theory [25]. Different from previous calibration work for visual recognition [44,42,41] that has applied EVT via rejection of a hypothesis, we use EVT to directly model probability of inclusion P I for a class of interest.…”

“…Prior work in EVT models for visual recognition [44,43,42,22,21] simply recommends choosing a tail size as an arbitrary percentage (not exceeding 50%) of the available data. How much data can be considered a proper tail: 1%, 5%, 10%, 20%?…”

Reduced Analytical Dependency Modeling for Classifier Fusion

“…To solve this problem, many algorithms have been developed to extract local or global discriminative features such as Local Binary Patterns [1], Laplacianfaces [2], etc. Instead of extracting a high discriminative feature, classifier fusion has been proposed and the results are encouraging [3] [4] [5] [6] [7] [8] [9] [10]. While many classifier combination techniques [3] have been studied and developed in the last decade, it is a general assumption that classification scores are conditionally independent distributed.…”