Uncertainty-Aware Face Recognition

In this chapter, we introduced the motivation of data uncertainty estimation in deep face recognition systems as well as its applications. From a probabilistic perspective, traditional deep face embeddings can be viewed as deterministic face embeddings, which do not take intrinsic data uncertainty of image samples into account. And therefore they will inevitably fail on ambiguous samples that are hardly recognizable. To solve the issue, Probabilistic Face Embedding (PFE) is introduced to represent each face image/template as a Gaussian distribution in the feature space. The variances of these distributions are then used as uncertainty estimation for feature comparison, feature pooling, and quality assessment. Data Uncertainty Learning (DUL) further extends the uncertainty estimation into the learning stage of backbone neural networks and improves their robustness against noisy training samples. Spherical Gaussian Face (SCF) extends SCF to von-Mises Fisher distributions to model the uncertainty on a spherical feature space, which better aligns with the feature distribution of most current deep face recognition systems.