Attention-Based Few-Shot Learning for Fine-Grained Vehicle Classification

While fine-grained vehicle classification has important applications in the security context, it is heavily limited by the availability of data. Particularly, the large number of vehicle models and the ongoing introduction of new models require regular and large dataset updates for a well-applicable vehicle classification system. In the field of few-shot learning, new visual classification approaches based on deep learning were proposed which claim to be more effective in terms of data usage. However, most few-shot approaches are evaluated in scenarios which only include a small number of classes. Thus, we evaluate attention-based few-shot approaches in a more difficult scenario which not only involves a significant number of classes with only a few images available but also including the classes of the base training for which abundant data is available. This new scenario better represents the challenges of few-shot learning for fine-grained classification in real-world scenarios where a classifier cannot afford to lose the capability of recognizing the base classes but also needs to be capable of being extended with new classes without large data collections. This scenario forces the approach to cope with the possibility of misclassifying a novel class as one of the many base classes rendering results more representative for real-world use cases. The results show that a modern transfer learning approach achieves good results even in this difficult scenario. Particularly, on a challenging dataset involving a high variety in terms of camera perspectives, unsupervised attention can further increase the accuracy.