Racial Artifacts in the COCO Dataset

Abstract

My paper, extending previous work on gender bias, investigates racial artifacts in the Common Objects in Context (COCO) dataset, which are visual cues in the images of the dataset that are correlated with race and learnable by a machine learning model. Using recently crowd-sourced Fitzpatrick skin type annotations on the dataset, I train a racial artifacts model to classify images as lighter-skinned or darker-skinned, and evaluate its performance after the image is transformed in various ways to remove or obscure different aspects of it, such as the color of the image or the faces of the people, to see if the model can still learn some correlation between the skin types. I quantitatively analyze performance and qualitatively analyze example images for these experiments to determine what these correlations might be, and discuss findings and implications. I then investigate possible downstream effects of these racial artifacts by training an object classification model and analyzing how its performance, and performance disparity between skin types, change after these image transformations are applied to the model's input. This work aims to facilitate understanding of racial bias in visual datasets and guide bias-aware fairness interventions and dataset construction in the future.