A comparative attention framework for better few-shot object detection on aerial images

Few-Shot Object Detection (FSOD) methods are mainly designed and evaluated on natural image datasets such as Pascal VOC and MS COCO. However, it is not clear whether the best methods for natural images are also the best for aerial images. Furthermore, a direct comparison of performance between FSOD methods difficult due to the wide variety of detection frameworks and training strategies. To this end, our contributions are twofold. First, we propose a benchmarking framework that provides a flexible environment to implement and compare attention-based FSOD methods. The proposed framework focuses on attention mechanisms and is divided into three modules: spatial alignment, global attention, and fusion layer. To remain competitive with existing methods, which often leverage complex training, we propose new augmentation techniques designed specifically for object detection. Using this framework, several FSOD methods are reimplemented and compared. This comparison highlights two distinct performance regimes on aerial and natural images: FSOD performs worse on aerial images. Our experiments confirm that small objects account for the poor performance. Small objects are difficult to detect, however in the few-shot regime, this challenge is largely reinforced. While the small object detection issue is well-known, to our knowledge this few-shot complication has never been reported in the literature. Second, always within the proposed framework, we develop a novel alignment method called Cross-Scales Query-Support Alignment (XQSA) for FSOD, to improve the detection of small objects. XQSA significantly outperforms the state-of-the-art on DOTA and DIOR, two aerial image datasets.