论文信息

论文标题：Automatically discovering and learning new visual categories with ranking statistics
论文作者：K. Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, A. Vedaldi, Andrew Zisserman
论文来源：ICLR 2020
论文地址：download 
论文代码：download
视屏讲解：click

1 介绍

　　提出问题：在已知类的基础上发现新类；

　　解决办法：特征向量的 top k 索引 ；

2 问题定义

　　Given an unlabelled dataset $D^{u}=\left\{x_{i}^{u}, i=1, \ldots, M\right\}$ of images $x_{i}^{u} \in \mathbb{R}^{3 \times H \times W}$ , our goal is to automatically cluster the images into a number of classes $C^{u}$ , which we assume to be known a priori. We also assume to have a second labelled image dataset $D^{l}=\left\{\left(x_{i}^{l}, y_{i}^{l}\right), i=1, \ldots, N\right\}$ where $y_{i}^{l} \in\left\{1, \ldots, C^{l}\right\}$ is the class label for image $x_{i}^{l}$ . We also assume that the set of $C^{l}$ labelled classes is disjoint from the set of $C^{u}$ unlabelled ones. While the statistics of $D^{l}$ and $D^{u}$ thus differ, we hypothesize that a general notion of what constitutes a "good class" can be extracted from $D^{l}$ and that the latter can be used to better cluster $D^{u}$ .

　　We approach the problem by learning an image representation $\Phi: x \mapsto \Phi(x) \in \mathbb{R}^{d}$ in the form of a $\mathrm{CNN}$ . The goal of the representation is to help to recognize the known classes and to discover the new ones. In order to learn this representation, we combine three ideas, detailed in the next three sections.

3 方法

4 总结

　　相似的样本其特征向量最大的索引是相似的。