Conceptualization, C.B., L.C. and F.N.; methodology, C.B., L.C. and F.N.; software, C.B., L.C. and F.N.; validation, C.B., L.C. and F.N.; formal analysis, C.B., L.C. and F.N.; investigation, C.B., L.C. and F.N.; resources, C.B., L.C. and F.N.; data curation, C.B., L.C. and F.N.; writing—original draft preparation, C.B., L.C. and F.N.; writing—review and editing, C.B., L.C. and F.N.; visualization, C.B., L.C. and F.N.; supervision, C.B., L.C. and F.N.; project administration, C.B., L.C. and F.N. All authors have read and agreed to the published version of the manuscript.

Figure 1. The workflow of the proposed fusion strategy.

Figure 1. The workflow of the proposed fusion strategy.

Figure 2. Some examples of images shown to participants during the data acquisition process in the GANT [36]. The first column shows images of two landscapes. The last two columns, on the other hand, show images of women and men: in the first row there are images of unknown people while in the second row there are images of two famous actors. Figure 2. Some examples of images shown to participants during the data acquisition process in the GANT [36]. The first column shows images of two landscapes. The last two columns, on the other hand, show images of women and men: in the first row there are images of unknown people while in the second row there are images of two famous actors.

Figure 3. Distribution of data samples with respect to the two classification tasks.

Figure 3. Distribution of data samples with respect to the two classification tasks.

Figure 4. The results of the age and gender classification using different classifier. The accuracies over the red line are taken in considerations.

Figure 4. The results of the age and gender classification using different classifier. The accuracies over the red line are taken in considerations.

Figure 5. The results of combining 2, 3, 4, 5, and 6 classifiers in age recognition using sum, product and Bayes rule as fusion strategies.

Figure 5. The results of combining 2, 3, 4, 5, and 6 classifiers in age recognition using sum, product and Bayes rule as fusion strategies.

Figure 6. The results of combining 2, 3, 4, and 5 classifiers in gender recognition using sum, product and Bayes rule as fusion strategies.

Figure 6. The results of combining 2, 3, 4, and 5 classifiers in gender recognition using sum, product and Bayes rule as fusion strategies.

Figure 7. The results of the classifiers are obtained using the combination of the best two scores of SVM and Random Forest for age classification, and of Gradient Boosting and SVM for gender one. In red are reported the best accuracies for age and gender, respectively.

Figure 7. The results of the classifiers are obtained using the combination of the best two scores of SVM and Random Forest for age classification, and of Gradient Boosting and SVM for gender one. In red are reported the best accuracies for age and gender, respectively.

Figure 8. The results of the best classifiers on single biometric traits. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. Exploiting only the blink for gender classification there are two classifiers that report the same higher accuracy.

Figure 8. The results of the best classifiers on single biometric traits. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. Exploiting only the blink for gender classification there are two classifiers that report the same higher accuracy.

Table 1. Spearman’s correlation coefficients with respect to the pairs of the three modalities.

Table 1. Spearman’s correlation coefficients with respect to the pairs of the three modalities.

Spearman’s Correlation CoefficientsBlink-pupil0.1439Blink-fixation−0.0992Fixation-pupil0.0644

Table 2. The results of the combination of the classifiers in gender recognition using different fusion strategies. n is the number of classifiers involved in the fusion process. The numbers in bold are the best results.

Table 2. The results of the combination of the classifiers in gender recognition using different fusion strategies. n is the number of classifiers involved in the fusion process. The numbers in bold are the best results.

Transformation-Based Score Fusion for Gender nCombination of ClassifiersAcc.Sum2GB&SVM
SVM&KNN0.84393GB&SVM&RF
GB&SVM&BG0.84624GB&SVM&RF&BG0.84625GB&SVM&RF&BG&KNN0.8439Prod2GB&SVM0.84623GB&SVM&BG0.84624GB&SVM&BG&RF0.84625GB&SVM&RF&BG&KNN0.8439Bayes
Rule 2GB&SVM
SVM&KNN0.8439

Table 3. The results of the combination of the classifiers in age recognition using different fusion strategies. n is the number of classifiers involved in the fusion process. The numbers in bold are the best results.

Table 3. The results of the combination of the classifiers in age recognition using different fusion strategies. n is the number of classifiers involved in the fusion process. The numbers in bold are the best results.

Transformation-Based Score Fusion for Age nCombination of ClassifiersAcc.Sum2RF&SVM
GB&SVM0.83493GB&SVM&DT0.84454SVM&KNN&DT&RF0.84455GB&SVM&DT&KNN&RF
GB&SVM&DT&KNN&BG0.84456GB&KNN&DT&BG&SVM&RF0.8445Prod2--3RF&SVM&DT
GB&SVM&DT
SVM&DT&BG0.84174GB&SVM&KNN&DT
SVM&KNN&DT&RF
SVM&BG&DT&RF0.84315SVM&BG&DT&RF&KNN0.84456GB&SVM&BG&DT&RF&KNN0.8417 Bayes
Rule 2RF&SVM0.8349

Table 4. The results of the age classification are obtained from combination of the best scores of the three biometric traits. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. The numbers in bold are the best results. X indicates the features selected for experimentation.

Table 4. The results of the age classification are obtained from combination of the best scores of the three biometric traits. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. The numbers in bold are the best results. X indicates the features selected for experimentation.

Age Classification without ConcatenationFixationPupilBlinkClassifierskAcc.XXXKNN20.8 XXKNN50.8091X XAD30.6182 XXDT20.7818BG10SVM3

Table 5. The results of the gender classification are obtained from combination of the best scores of the three biometric traits. For blinks, as maximum accuracy is achieved with two different classifiers, both scores are taken into account. For “Blink_1” we refer to the scores related to the DT classifier, while for “Blink_2” to those related to the BG classifier. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. The numbers in bold are the best results. X indicates the features selected for experimentation.

Table 5. The results of the gender classification are obtained from combination of the best scores of the three biometric traits. For blinks, as maximum accuracy is achieved with two different classifiers, both scores are taken into account. For “Blink_1” we refer to the scores related to the DT classifier, while for “Blink_2” to those related to the BG classifier. k is the value of the implemented k-fold cross-validation strategy relating to the best accuracy achieved. The numbers in bold are the best results. X indicates the features selected for experimentation.

Gender Classification without ConcatenationFixationPupilBlink_1Blink_2ClassifierskAcc.XXX SGD40.8346XX XSVM30.8421X X KNN60.7894X XKNN90.7669 XX SVM30.7970 X XSVM20.8045KNN3AD2XX KNN30.7970

Table 6. The results of the combination of the classifiers in gender recognition using different transformation-based score techniques without a a preliminary feature level fusion. The numbers in bold are the best results.

Table 6. The results of the combination of the classifiers in gender recognition using different transformation-based score techniques without a a preliminary feature level fusion. The numbers in bold are the best results.

Transformation-Based Score Fusion for Gender without Concatenation Combination of ClassifiersAcc. FixationPupilBlink SumRFSVMBG0.8054RFSVMDT0.8167ProdRFSVMBG0.7443RFSVMDT0.7511 Bayes
Rule RFSVM 0.7579RF BG0.6923RF DT0.7059 SVMBG0.7624 SVMDT0.7851

Table 7. The results of the combination of the classifiers in age recognition using different transformation-based score techniques without a a preliminary feature level fusion. The numbers in bold are the best results.

Table 7. The results of the combination of the classifiers in age recognition using different transformation-based score techniques without a a preliminary feature level fusion. The numbers in bold are the best results.

Transformation-Based Score Fusion for Age without Concatenation Combination of ClassifiersAcc. FixationPupilBlink SumSVMBGKNN0.7913ProdSVMBGKNN0.7763 Bayes
Rule SVMBG 0.7804SVM KNN0.5921 BGKNN0.7844

Table 8. For both classification tasks, the results of the single biometric traits (blink, fixation, and pupil) obtained with the same protocol are reported in the first three lines. The next line shows the best results obtained with our fusion strategy. In the last line there is a comparison with a paper that uses the same dataset with the same purpose. The numbers in bold are the best results.

Table 8. For both classification tasks, the results of the single biometric traits (blink, fixation, and pupil) obtained with the same protocol are reported in the first three lines. The next line shows the best results obtained with our fusion strategy. In the last line there is a comparison with a paper that uses the same dataset with the same purpose. The numbers in bold are the best results.

Summary Table: Best Results StrategyFeaturesClassifierskAcc.Age ClassificationFirst
classificationBlinkKNN100.5921FixationSVM30.5853PupilBG40.7872AllRF60.8336FusionSumwith
conc.AllGB&SVM&DT
SVM&KNN&DT&RF
GB&SVM&DT&KNN&BG
GB&SVM&DT&KNN&RF-0.8445without
conc.Fixation
Pupil
BlinkSVM&BG&KNN-0.7913Prod.with
conc.AllSVM&DT&KNN&BG&RF-0.8445without
conc.Fixation
Pupil
BlinkSVM&BG&KNN 0.7763Bayeswith
conc.AllRF&SVM-0.8349without
conc.Pupil
BlinkBG&KNN-0.7844Classifierswith
conc.AllBG70.8409without
conc.Pupil
BlinkKNN50.8091[8]PupilMultilayer perceptron-0.8369Gender ClassificationFirst
classificationBlinkBG
DT8
60.6357FixationRF60.6244PupilSVM20.7579AllGB20.8326FusionSumwith
conc.AllGB&SVM&BG
GB&SVM&RF
GB&SVM&RF&BG-0.8462without
conc.Fixation
Pupil
BlinkRF&SVM&DT-0.81674Prod.with
conc.AllGB&SVM
GB&SVM&BG
GB&SVM&RF&BG-0.8462without
conc.Fixation
Pupil
BlinkRF&SVM&DT-0.7511Bayeswith
conc.AllGB&SVM
SVM&KNN-0.8439without
conc.Pupil
BlinkSVM&DT-0.7851Classifierswith
conc.AllKNN40.8421without
conc.Fixation
Pupil
BlinkSVM30.8421[8]PupilSGD-0.5848