DAGER: Deep Age, Gender and Emotion Recognition

Problem

Commercial products needed age, gender, and emotion estimation accurate enough to make decisions on, across the range of faces and conditions a real deployment sees rather than a curated benchmark.

Approach

A multi-task convolutional network sharing one backbone across all three attributes, so the marginal cost of each additional prediction stays near zero. Trained and evaluated against the commercial APIs of the day rather than only against academic baselines.

Result

State-of-the-art age, gender, and emotion recognition, outperforming competitors including Microsoft by a large margin. Published as arXiv:1702.04280.