Semantic edge detection in poor data situations

Abstract

This paper deals with image processing techniques in poor data situations where the amount of learning data is insufficient. Specifically, given a small number of training data, this paper presents a method for quickly understanding the unseen category and detecting the boundary of the target object. To solve this problem, this work proposes a distance-based classifier equipped with an attention mechanism. The whole network is trained through the episodic training of meta-learning. In addition, a regularization technique is presented to train a more robust classifier. Since this paper is a pioneer work of few-shot edge detection technique, we construct two novel datasets and verify the performance of our model on those datasets.