Discriminative Spectral Regression Metric Learning in Unconstrained Face Verification

This paper presents the robustness of the proposed metric learning formulation, dubbed Discriminative Spectral Regression Metric Learning in offering a simplistic solution for measuring the Mahalanobis metric to solve unconstrained face verification problems. It takes advantage of distance metric learning on pairs of doublets by adopting the merit of the quadratic kernel function in the verification task. To be specific, the spectral graph analysis and the linear discriminant analysis are unified into the distance metric learning process for better exploitation of the intrinsic discriminant structure of face data. The proposed formulation is evaluated with four benchmarked constrained and unconstrained face datasets, with different tuning parameters under the restricted protocol. The promising result of 89.07% verification rate evinces the effectiveness and feasibility of the proposed formulation in unconstrained face verification compared to the state-of-the-art methods.