Machine Learning Bias in Computer Vision: Why do I have to care?

Abstract

Machine Learning bias is an issue with two main disadvantages. It compromises the quantitative performance of a system, and depending on the application, it may have a strong impact on society from an ethical viewpoint. In this work we inspect the literature on Computer Vision focusing on human-centered applications such as computer-aided diagnosis and face recognition to outline several forms of bias, bringing study cases for a more thorough inspection of how this issue takes form in the field of machine learning applied to images. We conclude with proposals from the literature on how to solve, or at least minimize, the impacts of bias.