Understanding Geometric Relationship Concepts in Few-Shot Learning

This paper is concerned with classifying geometric relationships in a few-shot learning (FSL) setting. Concept learning is proposed as a solution, combined with a new approach to use test samples during meta-testing. This new approach applies test images belonging to categories used during meta-training, but containing visually different content (e.g., circles, instead of rectangles, arranged in the same geometric relationship). The viability of the approach is demonstrated by training state-of-the-art FSL methods on datasets of synthetic images containing simple objects in various geometric relationships. The resulting models are able to generalise from images containing simple geometric shapes to images of more complex objects like artificial cracks and voids. The proposed method have application potential in many domains, where geometric relationships between, possibly complex, shapes are of importance, but only a handful of annotated samples are available per class. The FSL based approach may provide a useful alternative to presently used solutions using semi-supervised or self-supervised learning.