An Overview and Empirical Comparison of Distance Metric Learning Methods.

In this paper, we first offer an overview of advances in the field of distance metric learning. Then, we empirically compare selected methods using a common experimental protocol. The number of distance metric learning algorithms proposed keeps growing due to their effectiveness and wide application. However, existing surveys are either outdated or they focus only on a few methods. As a result, there is an increasing need to summarize the obtained knowledge in a concise, yet informative manner. Moreover, existing surveys do not conduct comprehensive experimental comparisons. On the other hand, individual distance metric learning papers compare the performance of the proposed approach with only a few related methods and under different settings. This highlights the need for an experimental evaluation using a common and challenging protocol. To this end, we conduct face verification experiments, as this task poses significant challenges due to varying conditions during data acquisition. In addition, face verification is a natural application for distance metric learning because the encountered challenge is to define a distance function that: 1) accurately expresses the notion of similarity for verification; 2) is robust to noisy data; 3) generalizes well to unseen subjects; and 4) scales well with the dimensionality and number of training samples. In particular, we utilize well-tested features to assess the performance of selected methods following the experimental protocol of the state-of-the-art database labeled faces in the wild. A summary of the results is presented along with a discussion of the insights obtained and lessons learned by employing the corresponding algorithms.

Can We Do Better in Unimodal Biometric Systems? A Rank-Based Score Normalization Framework.

Biometric systems use score normalization techniques and fusion rules to improve recognition performance. The large amount of research on score fusion for multimodal systems raises an important question: can we utilize the available information from unimodal systems more effectively? In this paper, we present a rank-based score normalization framework that addresses this problem. Specifically, our approach consists of three algorithms: 1) partition the matching scores into subsets and normalize each subset independently; 2) utilize the gallery versus gallery matching scores matrix (i.e., gallery-based information); and 3) dynamically augment the gallery in an online fashion. We invoke the theory of stochastic dominance along with results of prior research to demonstrate when and why our approach yields increased performance. Our framework: 1) can be used in conjunction with any score normalization technique and any fusion rule; 2) is amenable to parallel programming; and 3) is suitable for both verification and open-set identification. To assess the performance of our framework, we use the UHDB11 and FRGC v2 face datasets. Specifically, the statistical hypothesis tests performed illustrate that the performance of our framework improves as we increase the number of samples per subject. Furthermore, the corresponding statistical analysis demonstrates that increased separation between match and nonmatch scores is obtained for each probe. Besides the benefits and limitations highlighted by our experimental evaluation, results under optimal and pessimal conditions are also presented to offer better insights.