On using periocular biometric for gender classification in the wild

Abstract

The periocular area is a reliable cue for automatic gender classification (GC).Each local descriptor and grid configuration report different GC accuracy.The score level fusion of local descriptors increases GC performance.Tests carried out in a challenging large and unrestricted dataset.The fusion of periocular and facial GC reduces the classification error in roughly 20%. Display Omitted Gender information may serve to automatically modulate interaction to the user needs, among other applications. Within the Computer Vision community, gender classification (GC) has mainly been accomplished with the facial pattern. Periocular biometrics has recently attracted researchers attention with successful results in the context of identity recognition. But, there is a lack of experimental evaluation of the periocular pattern for GC in the wild. The aim of this paper is to study the performance of this specific facial area in the currently most challenging large dataset for the problem. As expected, the achieved results are slightly worse, roughly 8 percentage points lower, than those obtained by state-of-the-art facial GC, but they suggest the validity of the periocular area particularly in difficult scenarios where the whole face is not visible, or has been altered. A final experiment combines in a multi-scale approach features extracted from the periocular, face and head and shoulders areas, fusing them in a two stage ensemble of classifiers. The accuracy reported beats any previous results on the difficult The Images of Groups dataset, reaching 92.46%, with a GC error reduction of almost 20% compared to the best face based GC results in the literature.