Abstract
Few-shot object detection has made substantial progressby representing novel class objects using the feature repre-sentation learned upon a set of base class objects. However,an implicit contradiction between novel class classificationand representation is unfortunately ignored. On the onehand, to achieve accurate novel class classification, the dis-tributions of either two base classes must be far away fromeach other (max-margin). On the other hand, to preciselyrepresent novel classes, the distributions of base classesshould be close to each other to reduce the intra-class dis-tance of novel classes (min-margin). In this paper, we pro-pose a class margin equilibrium (CME) approach, with theaim to optimize both feature space partition and novel classreconstruction in a systematic way.
Abstract (translated)
URL
https://arxiv.org/abs/2103.04612
