Fig. 3.
Overview of our active learning framework. (a) The base model S predicts semantic masks, which are post-processed to generate ROIs. We align them to the same orientation for better clustering. (b) Our method adds an additional stream of unsupervised feature extracted by Eu. We apply hierarchical clustering to partition the unlabeled data and suggest cluster centers as queries for annotation. (c) Annotators provide True or False annotations for query samples that are used to fine-tune both the based mode S (black dashed line) and the proposed Eu (red dashed line).