2021/08/14 by Puspita Majumdar, Surbhi Mittal, Majumdar, Puspita +5
Computer Science · #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Face and Expression Recognition #Face recognition and analysis
paper · pdf · doi:10.48550/arxiv.2108.06581
openalex publication_date 2021/08/14 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Identifying and mitigating bias in deep learning algorithms has gained\nsignificant popularity in the past few years due to its impact on the society.\nResearchers argue that models trained on balanced datasets with good\nrepresentation provide equal and unbiased performance across subgroups.\nHowever, \can seemingly unbiased pre-trained model become biased when\ninput data undergoes certain distortions? For the first time, we attempt to\nanswer this question in the context of face recognition. We provide a\nsystematic analysis to evaluate the performance of four state-of-the-art deep\nface recognition models in the presence of image distortions across different\n\gender and \race subgroups. We have observed that image\ndistortions have a relationship with the performance gap of the model across\ndifferent subgroups.\n