One Label, One Billion Faces: Usage and Consistency of Racial Categories in Computer Vision (2102.02320v1)

Abstract: Computer vision is widely deployed, has highly visible, society altering applications, and documented problems with bias and representation. Datasets are critical for benchmarking progress in fair computer vision, and often employ broad racial categories as population groups for measuring group fairness. Similarly, diversity is often measured in computer vision datasets by ascribing and counting categorical race labels. However, racial categories are ill-defined, unstable temporally and geographically, and have a problematic history of scientific use. Although the racial categories used across datasets are superficially similar, the complexity of human race perception suggests the racial system encoded by one dataset may be substantially inconsistent with another. Using the insight that a classifier can learn the racial system encoded by a dataset, we conduct an empirical study of computer vision datasets supplying categorical race labels for face images to determine the cross-dataset consistency and generalization of racial categories. We find that each dataset encodes a substantially unique racial system, despite nominally equivalent racial categories, and some racial categories are systemically less consistent than others across datasets. We find evidence that racial categories encode stereotypes, and exclude ethnic groups from categories on the basis of nonconformity to stereotypes. Representing a billion humans under one racial category may obscure disparities and create new ones by encoding stereotypes of racial systems. The difficulty of adequately converting the abstract concept of race into a tool for measuring fairness underscores the need for a method more flexible and culturally aware than racial categories.

An Empirical Study of Racial Categories in Computer Vision Datasets

The paper "One Label, One Billion Faces: Usage and Consistency of Racial Categories in Computer Vision" presents a critical examination of the representation and usage of racial categories within computer vision datasets. The authors tackle the complexities and inconsistencies that arise when racial categories are used to annotate large-scale datasets, which are fundamental for training and evaluating facial analysis systems, including those aimed at ensuring fairness.

Key Findings and Methodology

The paper reveals that racial categories, although nominally similar across various datasets, are often inconsistent and reflect diverse racial systems that may encode stereotypes. A powerful methodological approach underpins the paper—training an ensemble of classifiers on different datasets to model their inherent racial systems. This approach allows for an empirical comparison of how racial categories generalize and transfer across datasets.

The authors conducted several experiments which provide significant insights:

• Cross-Dataset Generalization: Classifiers trained on each dataset were evaluated to see how well they predict racial categories on datasets they were not trained on. The findings indicate significant variation in how racial categories are interpreted, with individual datasets encoding unique racial systems despite nominal equivalence in category names.
• Self-Consistency: By splitting datasets into subsets for training classifiers, the authors found that the self-consistency of racial category assignment within each dataset is higher than consistency across datasets, indicating substantial intra-dataset coherence but inter-dataset divergence.
• Systematic Differences in Racial Classification: Through experiments involving biometric datasets, the researchers assessed how systemic the disagreements are regarding racial category assignments. They discovered that certain categories, notably 'Black', were more consistently assigned across datasets than 'White' or 'South Asian'.

Critical Evaluation of Racial Categories

The paper critiques the validity and utility of using broad racial categories in computer vision for several reasons:

• Poorly Defined Categories: Racial categories often represent loose and arbitrary geographical demarcations without standardized definitions, leading to inconsistent assignments and interpretations across different datasets.
• Stereotyping and Exclusion: The research highlights a tendency for datasets to reinforce stereotypes, revealing through experiments with ethnic groups that datasets may systematically exclude ethnicities that do not conform to prevalent stereotypes of racial categories.

Implications and Future Directions

This paper underscores the inadequacies of current practices in labeling racial categories for fairness evaluations in computer vision. The research suggests the need for more nuanced approaches to representing human diversity in facial datasets. Given that AI systems increasingly rely on these datasets, ensuring that they reflect human variety without reinforcing stereotypes or biases is crucial for both ethical and performance reasons.

Moving forward, the paper advocates for methodologies that are culturally aware and flexible, going beyond simplistic racial categorizations. This could involve adopting systems that recognize finer ethnic distinctions and multifaceted representations of identity, potentially incorporating self-reported data as well as environmental and phenotypic annotations.

Conclusion

Overall, the paper presents a compelling case for revisiting the approach to racial categories in computer vision datasets. It provides a robust empirical framework for understanding and improving how race is represented and interpreted, not only to enhance fairness and ethical considerations but also to foster more accurate and inclusive AI models. As datasets increasingly guide the development of AI technologies, ensuring they are representative of the complexity and diversity of human identities should be a priority in computational research and practice.