FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
75A direct way to find a relationship between the MOS and each <strong>of</strong> the 37 traits is using Pearson’scorrelation coefficient. We remind the reader that for two vectors, X = x 1 ,x 2 ,...,x n andY = y 1 ,y 2 ,...,y n , the Pearson’s correlation coefficient is given byr X,Y = cov(X,Y)σ X σ Y= E[(X −µ X)(Y −µ Y )]σ X σ Y, (6.1)where σ X and σ Y are being the standard deviations for X and Y , respectively. The coefficientranges between −1 and 1, with the two extreme points being obtained when the variables aremaximally linearly related.Pearson’s correlation coefficients are calculated for all 37 vectors, each vector correspondingto a feature. Per feature, a 260-values X vector describes each feature for each one <strong>of</strong> the 260training images 1 . The 260–values vector Y describes each image related MOS rating. Table 6.1itemizes these coefficients in decreasing order <strong>of</strong> importance with respect to the absolute Pearson’scorrelation coefficient.6.4.2 Insight provided from empirical dataThe first notable result reveals the strong correlation between the best ranked traits and theMOS, which even exceeds a Pearson’s correlation coefficient <strong>of</strong> 0.5 for the trait ’ratio eyeheight/face-height’.Particularly in regard to an automatic MOS prediction image processing toolthese results are very encouraging. Further we observe that photo-quality features play a less significantrole than facial aesthetics, as expected, but they are not to be neglected, since they achieveanr 14,MOS = 0.168. Moreover we note that the high ranked traitsx 1 ,x 2 andx 4 , which representthe ratios (eye-height/face-height) and (head-width/head-height), and furthermore face shape, seeTable 6.1 are features corresponding strongly to person’s weight. This outcome brings to the forethe strong importance <strong>of</strong> low human weight for aesthetics. Furthermore it is worth noting thatTable 6.1 reveals the surprising fact among others, that non permanent traits place a pivotal rolein raising the MOS rating. Eye make-up, lipstick, glasses and hair-style are all among the top 11<strong>of</strong> the obtained ranking. These results hint the high modifiability <strong>of</strong> facial aesthetics perceptionby simple means like make-up or hair styling. The relevance <strong>of</strong> eye make-up had been previouslyobserved in [GKYG10]. Together with the different conclusions that one may draw from Table6.1, it also becomes apparent that different questions are raised, on the interconnectedness <strong>of</strong> thedifferent traits. This is addressed in Section 6.4.3. Finally we note that traits, such as x 1 , x 7 ,x 12 and x 13 directly comply with the well known babyfaceness hypothesis (see [bea11]), whichdescribes that childlike facial features in females increase attractiveness, such features include bigeyes, e.g. x 1 and a relative low location <strong>of</strong> facial elements, e.g. x 7 , x 12 and x 13 . One measureknown for increasing attractiveness, if equal to the golden ratio φ = 1.618, is x 16 .6.4.3 Interconnectedness <strong>of</strong> different traitsTo get a better understanding <strong>of</strong> the role <strong>of</strong> the different traits in raising theMOS, it is helpfulto study the inter-relationship between these traits. This is addressed in Table C.3 in the Appendix,which describes the correlation between selected traits. Due to lack <strong>of</strong> space we limit the correlationmatrix to just a group <strong>of</strong> the first six traits. Table C.3 can answer different questions such asfor example the validity <strong>of</strong> the conclusion in Table 6.1 on the importance <strong>of</strong> the make-up feature.In this case, the question arises whether it is truly the make-up that affects the MOS or whether1. For information on denotation <strong>of</strong> features and according X–values, please refer to the Appendix, Table C.2