本文构建了一个用于航空图像中旋转对象检测的大规模数据集DOTA。该数据集包含从不同的传感器和平台收集了2806幅航空图像，每幅图像的大小约为4000×4000像素，包含显示各种比例、方向和形状的物体，完全注释的数据集包含188282个物体实例，它比该领域现有的任何数据集都要大得多。与一般的自然图像相比，本文使用旋转边界框注释大量分布良好的旋转对象，这些遥感图像由航空图像判读专家使用15种常见物体类别进行标记。本文还建立了航空图像中目标检测的基准，并展示了通过修改主流检测算法生成旋转边界框的可行性，实验表明DOTA具有相当大的挑战性。
This paper constructs DOTA, a large-scale dataset for oriented-object detection in aerial images. The dataset contains 2806 aerial images collected from different sensors and platforms. Each image is approximately 4000 × 4000 pixels and contains objects showing various scales, orientations, and shapes. The fully annotated dataset contains 188,282 instances, which is much larger than any existing dataset in the field. Compared with general natural images, this paper annotates many well-distributed rotating objects with rotating bounding boxes. These remote sensing images are labeled by aerial image interpretation experts using 15 common object categories. This paper also establishes a benchmark for object detection in aerial images and demonstrates the feasibility of generating rotated bounding boxes by modifying mainstream detection algorithms. Experiments show that DOTA is quite challenging.
[文章链接]
[数据集地址]
