Project
UTKFace & Face Analysis
The widely used UTKFace dataset and the conditional adversarial autoencoder behind it — age progression and regression learned from 20,000+ faces.
How will a face age? AICIP’s conditional adversarial autoencoder (CAAE, CVPR 2017) answered by learning a smooth face manifold where age becomes a traversable direction — generating plausible progressions and regressions of a face across decades from a single photograph. It became one of the lab’s most-cited papers.
The work’s most durable artifact is UTKFace: a dataset of over 20,000 face images annotated with age (0–116), gender, and ethnicity, spanning pose, expression, illumination, and resolution in the wild. Released alongside the paper for non-commercial research, UTKFace became a standard benchmark far beyond its origin — used across research on age estimation, face generation, and, notably, algorithmic fairness, where its demographic annotations made it a common testbed for studying bias in face analysis systems.
The face-analysis line continued through 2021 with work on decoupled generative models and face aging, before the lab’s generative-modeling energy shifted toward representation learning.
Publications
Age Progression/Regression by Conditional Adversarial Autoencoder
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2017
Origin of the UTKFace dataset; the lab's most-cited paper.