Multi-scale Dense Object Detection in Remote Sensing Imagery Based on Keypoints

Abstract

Object detection is a popular topic of computer vision and has attracted increasing attention in the field of remote sensing. However, object detection in remote sensing imagery is a very challenging task. Many existing detection methods depend on densely tiled anchor boxes, which require redundant computation resources and careful fine-tuning. In recent years, some anchor-free detectors have been proposed and show excellent performance. In this paper, we propose a novel anchor-free method for remote sensing object detection, named as multi-scale dense object detector based on keypoints (MDKD). Our method detects object in remote sensing imagery as a single keypoint, which needs neither predefined anchor boxes nor fallible pairing operation. For remote sensing scenario, we dedicate to improve the recall of detection results by allowing more keypoint samplings. In addition, we take advantage of balanced L1 loss to predict bounding boxes more accurately. Multi-scale test improves the performance further more. Our proposed MDKD method is verified on two widely used benchmarks, i.e. NWPU VHR-10 and DOTA datasets. The experimental results indicate that our keypoint based detector is effective on remote sensing imagery and outperforms several popular anchor-based methods.