Analyzing the Differences of Female/ Male Faces on Recognition Accuracy

This paper provides a comprehensive analysis of the differences between men and women on recognition accuracy. Proposed methodology depends on standard (ArcFace) face matcher , which was trained with Residual Network (ResNet-100) depending on MS1MV2 dataset images. Differences between male/female faces are analyzed to inspect the effects on recognition results. It is shown that recognition accuracy is lower for female faces due to set of factors. Yet, the deviation in the genuine distribution for female face tends toward lower scores of similarity. Accuracy difference is also obvious over face subsets that have minimum possible pitch angle (close to zero). Even by excluding images with partial forehead obstruction by hat or hair, accuracy differences persist for the juggler/genuine distributions. It is also witnessed that the genuine distribution was enhanced where only women's images without makeup were used; however, the female juggler distribution is also degraded as well. Finally, the results show that differences in recognition accuracy are witnessed even if (standard adopted) the deep learning techniques are trained on a face dataset with an explicitly balanced number of images and subjects for female and male faces.