Fast and reliable probabilistic face embeddings based on constrained data uncertainty estimation

Probabilistic Face Embeddings (PFE) can improve face recognition performance in unconstrained scenarios by integrating data uncertainty into the feature representation. However, the matching speed of PFE is too slow to be applied to large-scale face recognition or retrieval applications. Moreover, since deep learning-based data uncertainty estimation tends to be over-confident, the recognition performance of PFE is unstable. This paper proposes a probabilistic face embedding method to improve the robustness and speed of PFE. Specifically, the mutual likelihood score (MLS) metric used in PFE is simplified by using a one-dimensional variance to approximate the data uncertainty of face feature, to speed up the matching of probabilistic embedding pairs. Then, a unilateral constraint loss is proposed to limit the variation range of the lower part of the estimated data uncertainties, which can solve the problem of accuracy degradation in high-quality images. In addition, a feature fusion method based on temperature scaling is proposed, which can adjust the fusion weights of different quality images to improve the performance of video face recognition. Comprehensive experiments show that the proposed method can achieve comparable or better results in 6 benchmarks than the state-of-the-art methods with a less computational cost at the matching process. The code of our work is publicly available in GitHub (https://github.com/KaenChan/ProbFace).