Abstract

Ears have been discovered to have biometric importance for identifying people and/or verifying their identity. This is largely because of their complex inner shape structure, which is not only unique but also long-lasting regardless of ageing. In this paper, we make two important contributions relevant to analysis of ear in imagery captured in unconstrained conditions. That is, we present (a) the first, to the best of our knowledge, annotated database with ear landmarks and use it in order to build statistical deformable ear models in-the-wild and (b) a database of 2058 labelled unconstrained ear images with 231 subjects and use it for ear recognition/verification. We perform extensive comparisons for ear alignment using many state-of-the-art techniques and extensive experiments. Finally, we conducted extensive experiments for ear recognition using both handcrafted, as well as learned features (i.e., using deep learning). All annotated data and code will be publicly available.