| Oriented object detection can locate objects with arbitrary orientation in remote sensing images more accurately than horizontal object detection and it is now the mainstream,in which the design of bounding box regression loss functions is concerned with the accuracy of object localization,and the improvement and optimisation of the oriented object detection is one of the current hot topics.The bounding box regression loss functions used by the existing oriented object detection methods have the following problems:(1)The calculation scheme of the existing GIoU is not suitable for oriented object detection.(2)The existing bounding box regression loss functions for oriented object detection can not dynamically adjust the learning intensity according to the Io U value.To address the above issues,this paper carries out the following works:In order to address problem(1),this paper proposes an improved GIoU calculation method for oriented object detection.For the critical aspect of the GIoU calculation,i.e.the area of the minimum external rectangle,a method is proposed for calculating the minimum external rectangular area of two oriented rectangular boxes,which provides the basis for subsequent design of the bounding box regression loss functions around the GIoU optimisation.In order to address problems(2),this paper proposes a smooth GIoU bounding box regression loss function.The loss function is designed in the form of the segmentation function,and different optimization strategies are used for different ranges of GIoU values.i.e.,the smooth GIoU loss function can adaptively adjust the gradient values according to the GIoU value and its scheme can be extended to other Io U-based bounding box regression losses.In this paper,the effectiveness of the smooth GIoU loss function is experimentally validated on the two large remote sensing datasets DOTA and DIOR-R.The ablation experiment demonstrates the rationality of the design of the smooth GIoU loss function in the form of the segmented function.The quantitative comparison with other bounding box regression loss functions indicates the superiority of the smooth GIoU regression loss function.The experiment of generalisation capability evaluation shows that the smooth GIoU loss function is suitable for all types of oriented object detection models.The quantitative comparison experiment with the popular algorithm shows that the smooth GIoU loss function has excellent performance on both large remote sensing datasets.The subjective evaluation experiment shows that smooth GIoU loss function can be used to locate oriented objects more accurately.The parallel computation capability evaluation experiment shows that increasing the number of display cards can significantly accelerate the training and inference speed with a small decrease in detection accuracy. |
