计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (10): 198-208.DOI: 10.3778/j.issn.1002-8331.2301-0103

• 图形图像处理 • 上一篇    下一篇

特征互斥化的目标检测域适应方法

李润泽,王子磊   

  1. 中国科学技术大学,合肥 230027
  • 出版日期:2024-05-15 发布日期:2024-05-15

Domain Adaptive Object Detection Method Based on Feature Mutual Exclusion

LI Runze, WANG Zilei   

  1. University of Science and Technology of China, Hefei 230027, China
  • Online:2024-05-15 Published:2024-05-15

摘要: 当前,蒸馏学习已成为目标检测无监督域适应领域中的一种常用技术手段。然而蒸馏带来的特征偏移会导致目标域上伪标签的准确性较低,不利于目标域的精确检测。因此提出特征互斥化方法,包括特征分布互斥化和特征属性互斥化。其中特征分布互斥化鼓励网络对不同类别的特征分布进行互斥,特征属性互斥化促使分类器对不同类别主要使用的属性进行互斥。还提出强弱增强一致性方法对网络的预测输出进行一致性约束,促使网络提取的特征中主要包含与目标域检测相关的属性,进一步提高特征互斥化方法的效果。所提方法在多个域适应场景上进行了广泛的实验,在相同实验设置下的结果表明,所提方法较其他先进方法具有更好的有效性。

关键词: 目标检测, 无监督域适应, 蒸馏学习, 计算机视觉

Abstract: Recently, distillation learning has become a common technical means in the field of unsupervised object detection domain adaptation. However, due to the feature shift of distillation, the accuracy of the pseudo-labels obtained on the target domain is not so accurate, which has a certain negative impact on the target domain precise detection. Therefore, a feature mutual exclusion method is proposed, including feature distribution mutual exclusion and feature attribute mutual exclusion. The feature distribution mutual exclusion is used to prompt the feature distribution of different categories to be mutually exclusive, while the feature attribute mutual exclusion realizes that the classifiers mainly rely on mutual exclusive attributes when classifying different categories of features. In addition, a strong-weak augment consistency method is proposed to constrain the consistency of the network prediction, so that the features extracted by the network will mainly contain attributes related to the target domain detection, thereby improving the effect of the feature mutual exclusion method. Extensive experiments are conducted on several domain adaptation scenarios. The results show the effectiveness of the proposed method compared with other state-of-the-art methods under the same experimental settings.

Key words: object detection, unsupervised domain adaptation, distillation learning, computer vision