Ethically sourced image data set encourages fairness in AI research

Researchers have created a human-image data set called the Fair Human-Centric Image Benchmark, consisting of 10,318 images of 1,981 people from 81 geographical areas around the world, and collected in a way that reflects diversity, mitigates bias, protects intellectual-property rights and includes consent. Subjects themselves described features shown in the images — researchers could then compare these descriptions with the outputs of models designed, for example, to recognize faces or parse facial features, in order to judge the fairness of each system.