Universal domain adaptation (UDA) aims to transfer the knowledge of common classes from source domain to target domain without any prior knowledge on the label set, which requires to distinguish the unknown samples from the known ones in the target domain. Recent methods preferred to increase the inter-sample affinity within a known class, while they ignored the inter-sample affinity between the unknown samples and the known ones. This paper reveals that exploiting such inter-sample affinity can significantly improve the performance of UDA and proposes a knowability-aware UDA framework based on it. First, we estimate the knowability of each target sample by searching its neighboring samples in the source domain. Then, we propose an auto-thresholding scheme applied to the estimated knowability to determine whether a target sample is unknown or known. Next, in addition to increasing the inter-sample affinity within each known class like previous methods, we design new losses based on the estimated knowability to reduce the inter-sample affinity between the unknown target samples and the known ones. Finally, experiments on four public datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.
翻译:通用域适应(UDA)旨在将共同类别的知识从源域转移到目标域,而没有事先对标签集有任何了解,这就要求将未知样本与目标域已知样本区分开来。最近的方法倾向于增加已知类别中不同样本之间的相亲关系,而忽视未知样本与已知样本之间的相亲关系。本文显示,利用这种相亲关系可以大大改进UDA的性能,并在此基础上提出一个了解UDA的框架。首先,我们通过在源域中搜索其相邻样本来估计每个目标样本的可知性。然后,我们提出一个自动保存方案,用于估计可知性,以确定目标样本是否为未知或已知。接下来,除了增加与以往方法一样的每个已知类别之间的相亲和性外,我们还根据估计的可知性设计新的损失,以减少未知目标样本与已知样本之间的相亲近性。最后,我们对四个公共数据集的实验表明,我们的方法大大超出现有状态。