Evaluation of Gender Classification Methods with Automatically Detected and Aligned Faces

Abstract

We present a systematic study on gender classification with automatically detected and aligned faces. We experimented with 120 combinations of automatic face detection, face alignment and gender classification. One of the findings was that the automatic face alignment methods did not increase the gender classification rates. However, manual alignment increased classification rates a little, which suggests that automatic alignment would be useful when the alignment methods are further improved. We also found that the gender classification methods performed almost equally well with different input image sizes. In any case, the best classification rate was achieved with a support vector machine. A neural network and Adaboost achieved almost as good classification rates as the support vector machine and could be used in applications where classification speed is considered more important than the best possible classification accuracy.

Robust gender recognition by exploiting facial attributes dependencies

“…To make gender classifiers robust to face pose, images are aligned to some canonical pose or configuration (Mäkinen and Raisamo, 2008;Gallagher and Chen, 2009). This procedure is computationally very intensive, prone to errors, and, as reported by Mäkinen and Raisamo (Mäkinen and Raisamo, 2008), it does not increase significantly the performance of gender classifiers. Our proposal in this section is to transfer the alignment problem to the learning phase, avoiding the need for on-line alignment.…”

“…Gender is perhaps the most widely studied facial demographic attribute in the Computer Vision field (Moghaddam and Yang, 2002;Baluja and Rowley, 2007;Mäkinen and Raisamo, 2008;Bekios-Calfa et al, 2011). The state-ofthe-art recognition rate for the Color FERET database (Phillips et al, 2000) involving frontal faces with frontal illumination and 5 fold cross-validation is around 93% using either a Support Vector Machine with Radial Basis function (Moghaddam and Yang, 2002), pair-wise comparison of pixel values within a boosting framework (Baluja and Rowley, 2007) or linear discriminant techniques (Bekios-Calfa et al, 2011).…”

“…Previous gender classification in real-world settings align the face image to a canonical pose (Mäkinen and Raisamo, 2008;Gallagher and Chen, 2009). Face alignment requires the automated or manual detection of fiducial points (Mäkinen and Raisamo, 2008;Gallagher and Chen, 2009) or a congealling previous step (aligning all the images in a set by reducing entropy (Learned-Miller, 2006)).…”

“…Face alignment requires the automated or manual detection of fiducial points (Mäkinen and Raisamo, 2008;Gallagher and Chen, 2009) or a congealling previous step (aligning all the images in a set by reducing entropy (Learned-Miller, 2006)). In their work, (Mäkinen and Raisamo, 2008) report that, although manual face alignment does increase gender classification rates, the performance improvements achieved by automatic methods are not significant. This is caused by the lack of reliability of automated approaches.…”

De-genderization by body contours reshaping

“…The goal of such privacy filters is to make the identity of people unrecognizable. However, images of people offer additional information, and soft biometric attributes such as age [5], gender [10] and ethnicity [9] can be extracted from face/body images. In some cases, it is possible to deduce identity from a bag of soft biometric traits.…”