Enhancing Angle Perception in Rotation Object Detection: A Further Approach on Gaussian Distribution-Based Method

Substantial progress has been made in recent years in the detection of rotating objects in arbitrary directions, and Gaussian distributions have been used as a solution to the problem of boundary discontinuities. However, the existing Gaussian distribution design has certain limitations, and the angle perception of square-like objects is slightly weak, and the angle information of square-like objects cannot be expressed well. A detection scheme that can accurately express the spatial angle perception of square-like objects is needed when square-like objects exist in remote sensing images. In this paper, we analyse and summarise the limitations of the current Gaussian distribution modelling for the detection of square-like objects, and on top of that, we propose a Gaussian distribution design to enhance the angle perception, specifically, the rotating bounding box is modelled as a 2D Gaussian distribution, and the design of the 2D Gaussian distribution is improved by adding a new angular penalty term, and the different parts of the Gaussian distribution design form a complementary relationship within the formula, which makes it able to adaptively adjust to the detection of different morphological objects as a way to achieve enhanced angle perception. Our experimental results on two public datasets DOTA1.0 and HRSC2016 confirm the effectiveness of the method.