PL-FSOD: Pseudo-Labeling for Few-Shot Remote Sensing Object Detection—Mitigating Annotation Gaps and Forgetting

Few-shot object detection (FSOD) in remote sensing images (RSIs), unlike those in natural images, is highly challenging yet critical. Key hurdles include difficulties and incomplete annotations, drastic scale variations, and high inter-class similarity. Existing methods often underutilize K-shot support samples, perform poorly on novel categories, suffer from overfitting that degrades base-class performance, and lack efficacy for small objects. To address these, we propose Pseudo-Labeling FSOD (PL-FSOD) based on a teacher-student network. We design an Attention Path Aggregation Feature Pyramid Network (APAFPN) to boost small-object detection via multi-scale attention fusion. The Multi-Branch Region Proposal Network (MBRPN) and Pseudo-Labeling Bounding Box Head (PLBBH) enhance proposal quality and novel-object recall. Additionally, a metric-based loss function strengthens classifier discriminative power.Experiments on DIOR, NWPU VHR10.v2 show our method outperforms best FSOD approaches, validating the effectiveness of all designed modules.