On Fairness in Face Albedo Estimation
Haiwen Feng, Timo Bolkart, Joachim Tesch, Michael J. Black, Victoria Fernández Abrevaya · 2022
Digital avatars will be crucial components for immersive telecommunication, gaming, and the coming metaverse. Unfortunately, current methods for estimating the facial appearance (albedo) are biased to estimate light skin tones. This talk raises awareness of the problem with an analysis of (1) dataset biases and (2) the light/albedo ambiguity. We show how these problems can be ameliorated by recent advances, improving fairness in albedo estimation.