FAST-Det: Feature Aligned SSD Towards Remote Sensing Detector

Object detection based on large-scale, high-resolution visible light Remote sensing images are widely used in military such as reconnaissance and civilian such as marine resource management. It is also an important task for the application of computer vision in remote sensing images. With the development of deep learning, more and more object detectors use deep network as the backbone, and accurate detection results and indicators can be obtained on conventional images. However, compared with conventional images, remote sensing images have more object numbers and object sizes, and the object distribution is also denser, which makes detection more difficult. At present, there are two types of object detectors: single-stage and two-stage. The single-stage detector directly obtains the detection result based on the feature map and pays more attention to the detection speed, while the two-stage detector generates the region of interest (RoI) by using feature map. More attention is paid to the accuracy of the test results when the test results are obtained through RoIs. This paper proposes a bilateral filtering refining method based on a single-stage detector, which refines the results obtained by a single-stage detector and approaches the performance of a two-stage detector without losing too much detection speed. Experiments conducted on the public large-scale visible light remote sensing dataset DOTA have proved the effectiveness of this method.