LOVD: Land Vehicle Detection in Complex Scenes of Optical Remote Sensing Image

Nowadays, there is a growing body of research about object detection in remote sensing data. However, the detection algorithms for small targets in remote sensing areas are inadequate, largely because of the unavailability of high-quality datasets. Most remote sensing datasets are comprehensive, which means that they include bridges, airplanes, and lots of other common categories. Compared with other categories, the number and diversity of weak objects, such as vehicles, are quite insufficient. These limitations greatly affect the detection of small targets in remote sensing images. In order to promote the development of algorithms for the detection of small targets in remote sensing images and also allow access to remote sensing data, we have established a large-scale dataset for the detection of vehicle targets in optical remote sensing images and called it LOVD. It contains 1196 pictures and 541751 instances, covering 13 categories. For the dataset, we have proposed in this article: 1) it is the largest one in terms of vehicle category and the total number of vehicle instances; 2) it contains images with various backgrounds in different weather and scenarios; and 3) all targets are marked by oriented bounding boxes (OBBs), and two label formats are provided. Finally, we test the state-of-the-art detection algorithms on our dataset and provide a benchmark for OBB detection.