Uncovering Bias in the Face Processing Pipeline: An Analysis of Popular and State-of-the-Art Algorithms Across Demographic Groups

Numerous algorithms process face images to perform tasks such as person identification and estimation of attributes such as the race and gender. While previous work has focused on biases in face recognition systems, relatively limited work has considered the full face processing pipeline to determine if other components also exhibit any biases related to a person’s demographic attributes. An evaluation of popular and state-of-the-art methods in the face processing pipeline reveals that, although the overall performance may appear satisfactory, numerous differences are uncovered when digging deeper to consider the performance not just within a single demographic group, but also across different types of groups. Several avenues of future work are also provided.