Anchor-Free Few-Shot Object Detection for Densely Arranged Small Targets in Drone Images

In many cases, drone is needed to detect objects from high altitude. The lack of training samples of this object category will have a bad impact on the task. In this paper, we design a few-shot detector for drone images. It adopts anchor-free one-stage framework, which lead to more reasonable definition of positive and negative samples and faster speed. We introduce attention mechanism to enable our model match the objects of same categories and distinguish the different class objects and propose a matching score map to utilize the similarity information of attention feature map. The similarity probability of each pixel region for support category is integrated into regression bounding boxes to obtain the similarity probability of each regression bounding box. Finally, through soft-NMS, the predicted detection bounding boxes for support category objects are generated. Compared with YOLOv3 on DOTA dataset, our model is proved to be effective for few-shot detection task of drone images.