Evaluation of CNNs for Wildcats Classification in Real World Scenario

Biodiversity is threatened by habitat destruction and fragmentation; in this context, the knowledge and monitoring of patterns of occupancy of species is crucial for management of areas and planning proper actions. This study focuses on the monitoring of the presence of the European Wildcat (Felis silvestris Schreber) which conservation is threatened by several factors, including hybridization with domestic cats. This paper presents the evaluation state-of-the-art Convolutional Neural Networks (CNNs) for classification purpose. The employed dataset includes images of three main classes of cat species (i.e., Domestic, Wild, Hybrid) with three sub-classes each (i.e., ALIVE, DEAD, NECRO) captured in real world scenario using trap cameras or smartphones. The challenges are mostly related to the quality of the dataset, which is strongly biased and imbalanced, due to the nature of this task. Moreover, the images span within a wide range of qualities and acquisition settings including low lighting, specific poses, and backgrounds. The experimental results have highlighted that it is a highly complex task, and the data exhibit high specificity.