Arbitrary Oriented Few-Shot Object Detection in Remote Sensing Images

Abstract

Few-shot object detection (FSOD) aims to identify novel objects using only a limited number of samples. While impressive results have been achieved in natural scene images, object detection in remote sensing images (RSIs) presents unique challenges due to significant variations in orientation and size. Existing approaches often rely on horizontal bounding boxes, which may include a substantial quantity of irrelevant background because of object orientation, thus hindering accurate detection in RSIs. To address this limitation, we propose a metalearning-based method for arbitrary-oriented FSOD in RSIs, called AOFS. In our method, three key components are added to the YOLOv5-based one-stage detection architecture: a hierarchical metafeature encoder, an adaptive feature modulator, and an oriented bounding box decoder. We use features extracted from a small set of annotated samples to reweight the metafeatures of the objects. The oriented bounding box decoder provides outputs for object orientations. Experimental results on the benchmark datasets (NWPU and DIOR) indicate the effectiveness of our model. AOFS achieves 65.1% and 48.9% mean average precision in 3-shot setting on NWPU and DIOR, respectively, exceeding the second-best method by 4.4% and 5.4%.

Keywords

• Arbitrary-oriented object detection
• few-shot learning
• metalearning
• remote sensing images (RSIs)
• you-only-look-once (YOLO)