Adaptive reppoint: an efficient approach for orientated object detection in remote sensing imagery

With the improvement of the remote sensing technology, more and more remote sensing imagery data can be obtained easily. Therefore, the object detection in remote sensing imagery becomes an important task in computer vision area. However, the object of remote sensing imagery are presented with multi-scales, arbitrary orientation and complex background, which make the task becomes challenging. Currently, the mainstream object detection methods of multi-scales and oriented objects, such as RepPoint, which focus on learning a set of adaptive points and generating the point set to the bounding box (with angle information) of the objects. However, the diversity of categories and complex background always make the learned points (in adaptive points set) deviating outside the ground-truth bounding box, which will yield inaccurate location of the objects and decrees the detection result. To deal with this problem, this paper proposes a spatial constraint method that can enforce penalties on outlier points within the point set. Firstly, a reasonable evaluation metric is proposed for quantifying the distance between the outlier points and the ground-truth; Secondly, an spatial constraint parameter is defined for the objects in different scales. Extensive testing was performed on the challenging DOTA and HRSC 2016 datasets. According to the results, the proposed method demonstrated a performance improvement of 1.84% and 0.4% for the respective datasets.