| 基于高斯层级损失的开放场景物体检测 |
| |
| 作者姓名： | 王琳 陈熙霖 |
| |
| 作者单位： | 1. 中国科学院上海微系统与信息技术研究所, 上海;2. 中国科学院计算技术研究所, 北京 100190;3. 上海科技大学信息科学与技术学院, 上海 201210 |
| |
| 基金项目： | 国家自然科学基金（61390510）资助 |
| |
| 摘 要： | 计算机视觉中的物体检测包括2个目标——对物体的定位和识别。由于现有大多数的物体检测方法都是类别依赖的，因此无法应对开放场景中新类别的检测。注意到检测中定位和识别2个目标从已知类向未知类迁移的难度不同，定位具有更好的普适性，同时受人类在认知未知物体过程中层次关系的启发，提出一种高斯层级损失模型，在物体检测中采用物体类别层次化建模，在学习层级结构中每个类别多维高斯分布的同时，使用KL散度描述类别之间的层级关系，增强已知类到未知类的迁移性，从而提升物体检测方法在开放场景下对未知类的识别能力。实验表明，所提出的方法可以在不损失已知类性能的前提下，提升对未知类的检测能力。
|
| 关 键 词： | 物体检测 开放场景 类别层级关系 高斯层级损失 |
| 收稿时间： | 2020-08-11 |
| 修稿时间： | 2021-01-26 |
Hierarchical Gaussian loss based object detection in open world |
| |
| Authors: | WANG Lin CHEN Xilin |
| |
| Institution: | 1. Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai;2. Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China;3. School of Information Science and Technology, ShanghaiTech University, Shanghai 201210, China |
| |
| Abstract: | Object detection targets both object localization and categorization. Most of existing object detection methods are category-dependent, and can not deal with the detection task in open world with unknown categories. In aware that the difficulties of localization and categorization are different during transfer from known categories to unknown ones, and the localization is more universal than categorization, and inspired by the process of human categorization on unknown objects, we propose a Gaussian hierarchical loss model which applies an hierarchical modeling of object categories in object detection. KL divergence is used to describe the hierarchical relationship among categories while learning all the class multidimensional Gaussian distributions and enhance the transfer ability from the known classes to the unknown classes. Therefore the proposed method can extend the existing object detection methods to unknown categories in open world. Experimental results show that the proposed method can improve the detection ability of unknown categories without losing the performance on known categories. |
| |
| Keywords: | object detection open world hierarchical categories relationship Gaussian hierarchical loss |
|
| 点击此处可从《》浏览原始摘要信息 |
| 点击此处可从《》下载免费的PDF全文 |
|
