Critical direction projection networks for few-shot learning

With the development of deep learning, visual systems perform better than human beings in many classification tasks. However, the scarcity of labelled data is the most critical problem in such visual systems. Few-shot learning is adopted to tackle this problem, wherein a classifier should acquire the ability to identify some class is not in the training data when given only a few examples. In this paper, critical direction projection (CDP) networks are proposed for few-shot learning. Basically, two crucial steps are involved in CDP: The first step is to find the critical directions for each category in the embedding space, and the second step is to measure the similarity between samples and critical directions according to the projection length. It emerges that CDP networks can be effectively compatible with existing classification networks and achieve state-of-the-art performance on several benchmark datasets. Moreover, CDP achieves outstanding performance both on 2D image and 3D object classification. This study is a new attempt to achieve 3D object classification in a few-shot learning scenario. To summarize, our major research contributions are as follows: 1) a novel metric learning method, CDP, is proposed; 2) a new feature extraction module, EffNet, is introduced; and 3) a benchmark for few-shot 3D object classification is provided.