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.

Citation

    @inproceedings{zhao2021captionbias,
       author = {Dora Zhao and Angelina Wang and Olga Russakovsky},
       title = {Understanding and Evaluating Racial Biases in Image Captioning},
       booktitle = {International Conference on Computer Vision (ICCV)},
       year = {2021}
    }
  

Demographic annotations on COCO

We crowdsource skin color and gender annotations on the COCO 2014 validation set using Amazon Mechanical Turk (AMT). We use the Fitzpatrick Skin Type Scale, which ranges from 1 (lightest) to 6 (darkest) to measure skin color, and the binary perceived gender expression. In total, we collect annotations for 15,762 images and 28,315 person instances.





The annotations will be available for download upon request.

Analysis

Using the crowdsourced demographic annotations, we consider both the ground-truth images and manual captions as well as the automatically generated captions.


Our analysis shows that not only does bias exist in the ground-truth data, beyond just the underepresentation of certain skin tone groups, but also that this bias is propagating into the generated captions. Furthermore, we find that newer and more advanced image captioning models tend to exhibit more bias.

We provide an in-depth analysis of these bias propagation pathways in our paper.

Acknowledgements

This work is supported by the National Science Foundation under Grant No. 1763642 and the Friedland Independent Work Fund from Princeton University's School of Engineering and Applied Sciences. We thank Arvind Narayanan, Sunnie S. Y. Kim, Vikram V. Ramaswamy, and Zeyu Wang for their helpful comments and suggestions, as well as the Amazon Mechanical Turk workers for the annotations.

The webpage template was adapted from this project page.

Contact