Scatter Plot from Scientific Research

In this visualization, we select 5 novel classes with 200 query samples per class from an unseen domain (Places [63]). Each point indicates the feature representation of RelationNet (RN) [47] with backbone network (ResNet-10 [16]), projected by LDA [15], [34]. We use the linear regression prediction accuracy (LR-Acc) to demonstrate the improved decision boundaries: baseline RN (left), BAN episodic training applied to RN (mid), and our further proposed FS-BAN applied to RN (right). See numerical results and comparisons in Sec. VI.

DOI