Recently, the rotating target detector has been widely used in remote sensing images. However, the existing methods often use a large number of preset rotating anchors to cover the target, ignoring the interference caused by the shape change of target with high aspect ratio on network training, which is mainly reflected in the following two points: 1) missing high-quality positive samples containing slender targets with critical location information; 2) the gradient with the sharply changing causes training instability. In order to meet these challenges, we have designed a sample assignment strategy that can adapt to targets with different aspect ratios, and a training strategy that can more stably and accurately regression the bounding box with high aspect ratio. Specifically, first of all, the designed Shape Adaptive Label Assignment strategy introduces a weight function based on the IoU. Secondly, Gradient Equalization Regression Loss function is proposed to effectively alleviate the gradient instability of large aspect ratio targets during regression and make the model have better convergence. A series of experiments on DOTA and HRSC2016 datasets have confirmed the effectiveness of the proposed strategy.

not-yet-known not-yet-known not-yet-known unknown Abstract text. Recently, the rotating target detector has been widely used in remote sensing images. However, the existing methods often use a large number of preset rotating anchor to cover the target, without taking into account the interference caused by the shape change of the aerial target with high aspect ratio during the training, which is mainly reflected in the following two points: 1) missing high-quality positive samples that cover the critical information of slender targets 2) the gradient with the sharply changing causes training instability. In order to meet these challenges, we have designed a sample assignment strategy that can adapt to targets with different aspect ratios, and a training strategy that can more stably and accurately regression the bounding box with high aspect ratio. Specifically, first of all, the designed Shape Adaptive Label Assignment strategy introduces a weight function based on the IoU. Secondly, Gradient Equalization Regression Loss function is proposed to effectively alleviate the gradient instability of large aspect ratio targets during regression and make the model have better convergence. A series of experiments on DOTA and HRSC2016 datasets have confirmed the effectiveness of the proposed strategy.