FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP

40 3. BAG OF FACIAL SOFT BIOMETRICS FOR HUMAN IDENTIFICATION– Quantification error: the discrete values of soft biometric traits are mapped onto a limitedamount of bins and cause such an error. A lower amount of bins corresponds to less misclassificationsat the cost though of decreasing distinctiveness of this trait, as elaboratedabove.– The nomenclature of traits varies and is ambiguous. For example a hair color that might bedenoted with "red" can be labeled by different subjects as a variety of synonyms: auburn,orange, copper, reddish; but also as a completely different trait e.g. brown. A related studyestablishing labels for soft biometric traits with the Mechanical Turk was recently conductedby the authors in [CO11].– Different people perceive different traits (e.g. colors) differently. Specifically if the witnessof a crime has a different color understanding than the SBS performing the search it canlead to an erroneous search. This aspect though can be minimized if the witness is asked topoint at reference colors than just human labeling.– The awareness of people can be bad or wrong in how they remember traits.– Often occurring mixed categories like red-brown for hair color can be challenging for all,human perception, the SBS - training and - classification step.To visualize just the quantification error introduced by a human understandable SBS we havethe following simulation. We display in Figure 3.8 on the one hand purely the collision probabilityof subjects with 8 quantification bins(=traits instances) of hair color (light blond, dark blond, red,brown, black, grey, white and bald). On the other hand we have the re–identification of non–quantified and discrete computer–to–computer search. It is of interest, that even in the presence ofan estimation error the over all error probability is decreased in the absence of quantification error.A full computer–to–computer is presented in Chapter 5 and specifically in Figure 5.4, where theperformance of a SBS employing AdaBoost [FHT98] boosted algorithms for hair, skin and clothscolor, their textures and patch histograms is illustrated. The system is used for frontal–to–sidere–identification.10.90.80.7P err0.60.50.40.3Error probability of non−quantification, estimation error prone re−identificationCollision probability of 8−bin−quantified re−identification0.22 4 6 8 10 12 14 16 18 20Subjects NFigure 3.8: Re–identification error for hair color in ann-sized authentication group.3.6 SummaryIn this chapter we explored the use of multi-trait SBSs for human identification, studying analyticallythe relationship between an authentication group v, its size n, the featured categories ρ,and the effective categories F . Then we proceeded to show that in the interference limited setting,