Fruhner,M.,Tapken,H.
Abstract:
Animal Re-Identification (ReID) is a computer vision task that aims at retrieving a query individual from a gallery of known identities across different camera perspectives. It is closely related to the well-researched topic of Person ReID, but offers a much broader spectrum of features due to the large number of animal species. This raises research questions regarding domain generalization from persons to animals and across multiple animal species. In this paper we present research on the adaptation of popular Deep Learning based person ReID algorithms to the animal domain as well as their ability to generalize across species. We introduce two novel datasets for Animal ReID. The first one contains images of more than 300 wild common toads. The second dataset consists of various species of zoo animals. We then optimize different ReID models on these datasets as well as on 20 datasets published by others to evaluate the performance of the models in a non-person domain. We show that the domain generalization capabilities of OSNet AIN go even beyond the person ReID task, although it is a comparatively small CNN, which enables us to further investigate real-time animal ReID on live video data.