Textual Knowledge Fusion Reasoning Detector for Remote Sensing Few-Shot Object Detection

Due to the difficulty and time consumption of object-level annotations, remote sensing few-shot object detection has recently become a popular research topic. However, remote sensing objects exhibit significant variations in orientation and size, resulting in the misclassification of novel classes. Except for annotation samples, the text information between novel and base classes remains constant. Thus, in this paper, we explore how to leverage text information to resolve class confusion. Specifically, we propose a textual knowledge extractor and a cross-modal attention fusion module. The former is designed to extract textual knowledge for each class, while the latter aims to fuse it with visual knowledge. Experiments on the DIOR dataset demonstrate the effectiveness of our proposed method.