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

77– Pearson’s correlation coefficient:– Spearman’s rank correlation coefficient:– Mean standard error:r̂MOS,MOSr ŜMOS,MOS= 0.7794. (6.7)= 0.7860. (6.8)MSE = 1.158. (6.9)The high Pearson’s coefficient implies a robust prediction accuracy of the facial aestheticsmetric. The Spearman’s coefficient gives an indication about the correlation between estimatedand real MOS, but without the restriction of linear dependence. It considers each monotonicfunction connecting the two vectors. In our case this coefficient is relatively high as well. TheMSE on the other hand gives an idea about the absolute error between the predicted and actualvalues. It is interesting to observe that the testing set provides even higher correlation coefficientsthan the calibration set, but the MSE reveals that the absolute error increases for the testing set,and thus that the actual performance decreases.We proceed with three experiments using the validated MOS–prediction metric.6.6 Experiments with ̂MOSThe above designed MOS prediction metric is in this section employed towards (partial) quantificationof the general concept of beauty. We are specifically interested in addressing questionssuch as:– Are famous females known for their beauty more beautiful than average females?– What is the influence of age on beauty?– How much does facial surgery change the beauty score?Towards addressing the above, we proceed to apply our metric on images drawn from the internetand from official databases such as the FG–NET and the Plastic surgery database.6.6.1 Metric verification on highly ranked femalesTowards verification of its usefulness, we applied the above designedMOS–prediction metricon images of females who have been highly ranked by the popular media. Specifically we consideredimages of females leading the lists of People’s magazine as the ‘most beautiful people’ from1991 to 2011, as well as the top 10 entries from the same list for the year 2010. The considered imagesincluded, among others, those of Jennifer Lopez (winner 2011), Julia Roberts and AngelinaJolie. After annotation and calculation of the related beauty indices, we contrasted the results fromthe above lists, to those we obtained when we considered images from the HOTorNOT database(see Figure 6.5). The test validated our choice of metrics, with the entries from the above ‘beautifulpeople’ lists, consistently scoring significantly higher scores, as well as exhibiting a lowervariance. We displayed, for both image sets, the average MOS values, as well as those within aconfidence interval of 95%.6.6.2 Dependence between beauty and age: FG–NET aging databaseTowards investigating the dependence between beauty and age, we considered images fromthe FG–NET database [fgn11], as this database provides us with multiple images of subjects as