Few-shot learning in computer vision is a very challenging task. Many approaches have been proposed to tackle the few-shot learning problem. Meta learning, a method of learning to learn, is introduced into few-shot learning problem and has achieved pretty good results. But there is still a very big gap between the machine and our human in the few-shot learning tasks. We think it’s because the existing methods do not make full use of global knowledge (similar to the priori knowledge of human understanding of images) thus lacking a world view of the task. In other words, they focus too much on local information and neglect the whole task. In this paper, we rethink about the few-shot learning problem, and propose that we should make full use of global knowledge. Seen data set is used to obtain a embedding function between images and feature vectors, the images are embedded onto a hypersphere in the manner of cosine embedding. By taking this embedding function as global knowledge, we train a classifier to classify the corresponding embedded vector of images. The experiment proved that our approach significantly outperforms both baseline models and previous state-of-the-art methods. It surpasses most existing methods in terms of flexibility, simplicity and accuracy. Codes are available at https://github.com/SongyiGao/OICEFFSL.
CITATION STYLE
Gao, S., Shen, W., Liu, Z., Zhu, A., & Yu, Y. (2019). Only Image Cosine Embedding for Few-Shot Learning. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 11954 LNCS, pp. 83–94). Springer. https://doi.org/10.1007/978-3-030-36711-4_8
Mendeley helps you to discover research relevant for your work.