领域自适应目标识别综述

[摘要] 针对目前域自适应目标识别问题的学习方法，进行系统总结。首先，提出目标识别的两个基本主题：基于域自适应的目标分类和目标检测；然后，围绕这两个主题，从特征和样本两个角度，展开具体综述。认为对于域自适应目标分类，几种算法的主要问题为：忽略了样本所构成的流形几何结构，如果能利用几何结构来约束特征表达，将有利于样本特征鲁棒性的提高。对于域自适应目标检测，其问题为：现有方法对源样本和带标签的目标域样本存在依赖，这一问题使得现有的方法很难适用于某些真实的应用场景。

[关键词] 域自适应学习；目标分类；目标检测

[Abstract] In this paper, a systematic review on domain adaptation of object detection is presented. At first, the two basic subtopics of object recognition—object classification and object detection based on domain adaption are proposed; and then, from the views of feature and samples, two problems are reviewed in detail. For the domain adaptation object classification, the existing methods ignore the manifold structure of samples. In fact, the geometric information is helpful to obtain robust representation. For the domain adaptation object detection, the existing methods depend on the source samples and labeled target samples, which makes these methods hard to be employed into some real applications.

[Keywords] domain adaptation study; object classification; object detection