Abstract

We assess the state of thinking about gender identities in computer vision through an analysis of how research papers in gender and facial recognition are designed, what claims they make about trans and non-binary people, what values they espouse, and what they describe as ongoing challenges for the field. In our corpus of 50 research papers, the seven papers that consider trans and non-binary identities use questionable assumptions about medicalization as a measure of transness, about gender transition as a linear and bounded process, and about the concept of gender deception. Otherwise, non-normative gender identities are absent and their consideration is in fact hindered by prevailing research values, particularly deeply embedded ones such as performance and accuracy. We point out how the use of shared datasets calcifies binary conceptions of gender. In the way that the field of computer vision conceives of ongoing challenges for its research, it does not yet face questions that trans and non-binary user experiences pose and often falls back on biologically essentialist notions of sex classification. We make two recommendations: that computer vision researchers undertake interdisciplinary work with researchers who study gender as a socio-cultural phenomenon, and that journal editors and conference organizers do the same in peer review and conference acceptance processes.