The kernel classification-based metric learning in face verification

In this paper, a kernel classification distance metric learning framework is investigated for face verification. The framework is to model the metric learning as a Support Vector Machine face classification problem, where a Mahalanobis distance metric is learnt in the original face feature space. In the process, pairwise doublets that are constructed from the training samples can be packed and represented in a means of degree-2 polynomial kernel. By utilizing the standard SVM solver, the metric learning problem can be solved in a simpler and efficient way. We evaluate the kernel classification-based metric learning on three different face datasets. We demonstrate that the method manages to show its simplicity and robustness in face verification with satisfactory results in terms of training time and accuracy when compared with the state-of-the-art method