Focused Small Aerial Object Detection with Improved Feature Pyramid

Aerial object detection is of great significance in engineering. Influenced by load and flight altitude, it could be challenging to detect some objects covering a small part of an image. To alleviate the problem, we propose a focused small object detection method in this work. Considering that small objects will be easier to be found when an image is resized in the real world, we utilize a larger feature map to build the feature pyramid. Moreover, deformable convolution is involved in impelling the network to focus on small objects. As a result, the proposed method is capable of multi-scale object detection while no complicated module is required. Experiments on public datasets demonstrate the effectiveness of the proposed components, and the method achieves state-of-the-art performance.