How We've Taught Algorithms to See Identity: Constructing Race and Gender in Image Databases for Facial Analysis

Abstract

Race and gender have long sociopolitical histories of classification in technical infrastructures-from the passport to social media. Facial analysis technologies are particularly pertinent to understanding how identity is operationalized in new technical systems. What facial analysis technologies can do is determined by the data available to train and evaluate them with. In this study, we specifically focus on this data by examining how race and gender are defined and annotated in image databases used for facial analysis. We found that the majority of image databases rarely contain underlying source material for how those identities are defined. Further, when they are annotated with race and gender information, database authors rarely describe the process of annotation. Instead, classifications of race and gender are portrayed as insignificant, indisputable, and apolitical. We discuss the limitations of these approaches given the sociohistorical nature of race and gender. We posit that the lack of critical engagement with this nature renders databases opaque and less trustworthy. We conclude by encouraging database authors to address both the histories of classification inherently embedded into race and gender, as well as their positionality in embedding such classifications.

One Label, One Billion Faces

“…Our approach and exposition is heavily inspired by [47], who used an ensemble of classifiers to study dataset bias and cross-dataset generalization in commonly used datasets for object detection and classification, noting substantial differences in the representation of nominally similar categories across datasets. We draw on the survey of computer vision datasets in [42], who enumerate the diverse and ill-defined array of racial classification schemes and definitions in computer vision datasets. Geiger et.al [16] find significant variability in the quality of human annotations in a case study of papers on tweet classification.…”

“…Racial categories are used without being defined or only loosely and nebulously defined [42] -researchers either take the existence and definitions of racial categories as a given that does not have to be justified, or adopt a "I know it when I see it" approach to racial categories. Given that the categories are allusions to both geographic origin and physical characteristics, it is understandable that deeper discussion of these categories is avoided because it veers into unpleasant territory.…”

“…The construction of large-scale, fair computer vision datasets typically involves collecting a large number of face images, either from an established dataset or web scraping, and then ascribing a racial category to each of the collected faces. Scheurman et.al., [42] detail the myriad racial categories and classification schemes used in computer vision, but we identify and focus on a commonly used set of racial categories: East Asian, South Asian, White, and Black. Occasionally, these categories are referred to by different names -(e.g., Indian instead of South Asian).…”

“…Crucially, the annotation processes for computer vision datasets largely involve a "moment of identification" by the annotators [42], in which the annotators ascribe a racial label onto an individual based on their physical characteristics. This is different from data collection regimes in other domains, where racial labels may be provided by individuals themselves.…”

“…Our work is related in spirit to Hanna et.al, [18] as well as Benthall and Haynes [5], who both critique the usage of racial categories from the perspective of critical race theory. We draw on Scheurman et.al's [42] survey of identity classification schemes, definitions, and annotation methods in computer vision. Our work is heavily inspired by Torrallba and Efros [47], who similarly use an ensemble of classifiers to empirically study category bias and generalization Figure 1: Classifiers trained on fair computer vision datasets agree strongly about the race of some individuals, while disagreeing about others.…”