IoU 를 이용한 능동 학습의 샘플링 전략 평가방법

Abstract

Currently active learning is used to reduce the training data required for deep learning. In active learning, a few data are sampled from the training dataset to learn a model where the goal of sampling strategy is the sampled data to cover the distribution of the training dataset as much as possible. We propose a method for evaluating sampling strategies of the active learning in the instance segmentation task. We evaluate the sampling strategies by comparing the sampling data and its ground truth(G.T.) data with Mask IoU and Boundary IoU. It was confirmed that both evaluation methods were drawn as upward-sloping graphs as they passed through the sampling stage.