Few-Shot Object Detection of Remote Sensing Image via Calibration

Few-shot object detection (FSOD), which aims at detecting rare objects based on a few training samples, has attracted significant research interest. Previous approaches find that the performance degradation of FSOD is mainly caused by category confusion (high false positives). To solve this issue, we propose a two-stage fine-tuning approach via classification score calibration (TFACSC) for remote sensing images, which follows the flowchart of base training and few-shot fine-tuning to train the detector. First, the backbone with strong representation ability is employed to extract the multiscale features of the query image. Then, these features are aggregated by a novel multihead scaled cosine nonlocal module (MSCN). Next, the aggregated features are used to generate the objectiveness proposals by a region proposal network (RPN). Eventually, the generated proposals are refined by a bounding box prediction head for the final category and position prediction, and the category score is calibrated by a novel classification score calibration module (CSCM). Extensive experiments conducted on NWPU VHR 10 and DIOR benchmarks demonstrate the effectiveness of our model. Particularly, for any shot cases, our method greatly outperforms the baseline and achieves state-of-the-art performance.