| Summary: | Multi-class object detection in remote sensing imagery is an important and challenging topic in computer vision. Compared with the object detection of natural scenes, remote sensing object detection has some challenges such as scale diversity, arbitrary directions and densely packed objects. To resolve these problems, this paper presents a scale-aware rotated object detection. Firstly, we propose a novel feature fusion module, which takes full advantage of high-level semantic information and low-level high resolution feature. The new feature maps are more suitable for detecting objects with a large difference in scale. Meanwhile, we design a specific weighted loss, which contains an intersection-over-union (IoU) loss and a smooth L1 loss to further address the scale diversity. Besides, in order to detect oriented and densely packed objects more accurately, we propose a normalization strategy for the representation of rotating bounding box. Our method is evaluated on two public aerial datasets DOTA and HRSC2016, and achieves competitive performances. On DOTA, we boost the mean Average Precision (mAP) to 72.95% on oriented object detection.
|
|---|
