基于特征解耦的开放世界目标检测

Open world object detection based on feature disentanglement

开放世界目标检测是一项具有挑战性的视觉任务,填补了传统目标检测与真实世界目标检测的差距。与有限类别集合设定下的传统方法不同,开放世界目标检测不仅需识别和检测已知(可见)类别的目标,还要能够标记并逐渐学习未知(不可见)类别的目标。当传统的目标检测技术直接应用于开放世界场景时,常出现2个主要问题:其一,可能会将未知类视为背景而忽视;其二,可能将未知类错误地归类为已知类。为解决这些问题,提出采用退火算法分离已知与未知的特征,指导检测模型的学习过程。由于退火模块的引入,未知类精度有所提升,但已知类的精度略有下降,因此引入高效通道注意力模块提高已知类精度。与以往方法相比,该策略在检测已知类和未知类的目标上均表现出更优的性能。

Open-world object detection is a challenging visual task that bridges the gap between traditional object detection and that in real-world scenarios.Unlike traditional methods confined to a limited set of classes,open-world object detection requires not only the identification and detection of objects from known(seen)classes but also the ability to label and gradually learn objects from unknown(unseen)classes.When traditional object detection techniques are directly applied to open-world scenarios,two major problems often arise:first,they might treat unknown classes as background and ignore them;second,they might misclassify unknown classes as known ones.To tackle these problems,this study proposes the utilization of annealing algorithms to separate features of known and unknown classes,guiding the learning process of the detection model.The introduction of the annealing module leads to an improvement in the accuracy of unknown classes,but a slight decrease in the accuracy of known classes.To address this,an efficient channel attention module is incorporated to enhance the accuracy of known classes.Compared to previous methods,this approach demonstrates superior performance in detecting objects from both known and unknown classes.