Robust Fusion: Extreme Value Theory

Recognition problems in computer vision often benefit from 5 5 a fusion of different algorithms and/or sensors, with score level fusion be6 6 ing among the most widely used fusion approaches. Score level fusion re7 7 quires the different data to be normalized before combining. Choosing an 8 8 appropriate score normalization technique before fusion is a fundamen9 9 tally difficult problem because of the disparate nature of the underlying 10 10 distributions of scores for different sources of data. Further complica11 11 tions are introduced when one or more fusion inputs outright fail or have 12 12 adversarial inputs, which we find in the fields of biometrics and forgery 13 13 detection. Ideally a score normalization should be robust to model as14 14 sumptions, modeling errors, and parameter estimation errors, as well as 15 15 robust to algorithm failure. In this paper, we introduce the w-score, a 16 16 new technique for robust recognition score normalization. We do not as17 17 sume a match or non-match distribution, but instead suggest that the top 18 18 scores of a recognition system’s non-match scores follow the statistical 19 19 Extreme Value Theory, and show how to use that to provide consis20 20 tent robust normalization with a strong statistical basis. We cover the 21 21 background theory, use this theory to develop the w-score, and present 22 22 fusion results for a variety of biometric recognition algorithms and for 23 23 content-based image retrieval descriptors. 24 24