Ramachandran Sreeraj ; Rattani Ajita

Abstract

Published studies have suggested the bias of automated face-based genderclassification algorithms across gender-race groups. Specifically, unequalaccuracy rates were obtained for women and dark-skinned people. To mitigate thebias of gender classifiers, the vision community has developed severalstrategies. However, the efficacy of these mitigation strategies isdemonstrated for a limited number of races mostly, Caucasian andAfrican-American. Further, these strategies often offer a trade-off betweenbias and classification accuracy. To further advance the state-of-the-art, weleverage the power of generative views, structured learning, and evidentiallearning towards mitigating gender classification bias. We demonstrate thesuperiority of our bias mitigation strategy in improving classificationaccuracy and reducing bias across gender-racial groups through extensiveexperimental validation, resulting in state-of-the-art performance in intra-and cross dataset evaluations.