Is DFR for Soft Biometrics Prediction in Unconstrained Images Fair and Effective?

Udaybhan Rathore, Akshay Agarwal

Keywords: Face Recognition, Fairness, Unconstrained Dataset, Gender and Age Identification

Abstract: Face being a nonintrusive recognition modality, is an ideal candidate for identifying criminals. The modality is not only related to identity but can also extract several other important features such as age, race, and gender. In this preliminary research, we have collected a novel unconstrained face recognition dataset using mobile phones. On the collected dataset, we have evaluated the current state-of-the-art (SOTA) deep face recognition (DFR) algorithms for face attribute identification. This research aims to identify whether current algorithms are effective for the task or biased towards their training set. The results suggest that the current technology is not effective in identifying face attributes when the images are captured in unconstrained environments. For example, deep face networks yield the best F-1 score of $\mathbf{0.43}$ when asked to predict gender on the collected dataset.