Understanding and Evaluating Racial Biases in Image Captioning

Abstract

Image captioning is an important task for benchmarking visual reasoning and for enabling accessibility for people with vision impairments. However, as in many machine learning settings, social biases can influence image captioning in undesirable ways. In this work, we study bias propagation pathways within image captioning, focusing specifically on the COCO dataset. Prior work has analyzed gender bias in captions; here we focus specifically on racial biases. Our first contribution is in annotating the skin color of 28,315 of the depicted people after obtaining IRB approval. Using these annotations, we compare racial biases present in both manual and automatically-generated captions, focusing specifically on differences between images depicting people with lighter versus darker skin. We observe differences in caption accuracy, sentiment, and word choice. Most importantly, we find that modern captioning systems exhibit stronger biases than older models — thus, social bias concerns are likely to become increasingly prevalent in image captioning.