Rethinking Common Assumptions to Mitigate Racial Bias in Face Recognition Datasets

Many existing works have made great strides towards reducing racial bias in\nface recognition. However, most of these methods attempt to rectify bias that\nmanifests in models during training instead of directly addressing a major\nsource of the bias, the dataset itself. Exceptions to this are\nBUPT-Balancedface/RFW and Fairface, but these works assume that primarily\ntraining on a single race or not racially balancing the dataset are inherently\ndisadvantageous. We demonstrate that these assumptions are not necessarily\nvalid. In our experiments, training on only African faces induced less bias\nthan training on a balanced distribution of faces and distributions skewed to\ninclude more African faces produced more equitable models. We additionally\nnotice that adding more images of existing identities to a dataset in place of\nadding new identities can lead to accuracy boosts across racial categories. Our\ncode is available at\nhttps://github.com/j-alex-hanson/rethinking-race-face-datasets.\n