Object Detection in Aerial Images Using a Multiscale Keypoint Detection Network

Automatic object detection in aerial imagery is being increasingly adopted in many applications, such as traffic monitoring, smart cities, and disaster assistance. In keypoint-based detectors, the prediction modules are usually generated from a fixed feature map scale. This configuration significantly limits the ability to detect multiscale objects in aerial scenes. The corner selection module in these detectors often ignores that a category in an aerial image is relatively unitary. In this article, a novel network, called the multiscale keypoint detection network (MKD-Net), is proposed to address these challenges. MKD-Net fuses multiscale layers to generate multiple feature maps for objects of different sizes. During the inference phase, both feature maps can be exploited for predicting corners. Moreover, a category attention module is designed to reduce the channel noise for a single-category scene. Experiments on benchmarks PASCAL VOC and DOTA show promising performance of MKD-Net compared with the baseline network. The code is available on https://github.com/jason-su/MKD-NET.