Oriented Object Detector Based on Bivariate Gaussian Distribution in Remote Sensing Images

One fundamental task in graphic information processing for remote sensing is to detect interesting objects. However, the boundary discontinuity and the independent regression of parameters for the oriented bounding box (OBB) detection have become a bottleneck restricting the improvement of convergency speed and identification accuracy. In this article, we propose a one-stage framework based on the bivariate Gaussian distribution (BGD) to achieve high-precision arbitrary-oriented object detection in optical aerial images. Specifically, a bivariate Gaussian kernel is proposed to map the center zone of targets into the heatmap, and the mapping zone is related to the target’s location, shape and rotated angle. Then the OBB of the object is regarded as a rotated bivariate Gaussian distribution to solve the problem of boundary discontinuity, and the Wasserstein distance is selected to measure the deviation between two distributions. Besides, we propose a normalized Gaussian Wasserstein distance to further improve the positioning accuracy of the regression of the bounding box. Contrast experiments are conducted on the datasets HRSC2016 and UCAS-AOD, and state-of-the-art performances demonstrate this approach has obvious advantages in rotated object detection.