Few-Learning for Fine-Grained Vehicle Model Recognition

Recent vehicle recognition models have focused on using conventional Deep Neural Network approach to solve the problem. The most important limitation for such approach is the scalability issue. They require hundreds or even thousands of image samples per class to train effective models. However, new vehicle models become available yearly, which imposes constraints to source for hundreds of image samples for every new vehicle model and retrain the recognition algorithms on the entire dataset to introduce recognition capability for new classes. This process is data hungry, tedious, computationally intensive, and time consuming. To mitigate this issue, we propose using few-shots learning approach which allows for model scalability to new unseen classes with as few as 1 to 20 image samples per class rather than hundreds. In this paper, we introduce RelationNet++, an improved version of Relation Network, for recognition with few-shots-learning. Our experimental results show that the proposed model outperforms current state-of-the-art few-shot learning methods on the vehicle recognition dataset.