Task interleaving and orientation estimation for high-precision oriented object detection in aerial images

Oriented object detection in aerial images has received extensive attention due to its wide range of application scenarios. Although great success has been achieved, current methods still suffer from inferior high-precision detection performance. Firstly, the classification scores cannot truly represent the localization accuracy of the predictions. Secondly, the orientation prediction in these detectors is not accurate enough for high-precision object detection. In this paper, we propose a Task Interleaving and Orientation Estimation Detector (TIOE-Det) for high-quality oriented object detection in aerial images. Specifically, a posterior hierarchical alignment (PHA) label is proposed to optimize the detection pipeline. TIOE-Det adopts PHA label to integrate fine-grained posterior localization guidance into classification task to address the misalignment between classification and localization subtasks. Then, a balanced alignment loss is developed to solve the imbalance localization loss contribution in PHA prediction. Moreover, we propose a progressive orientation estimation (POE) strategy to approximate the orientation of objects with n-ary codes. On this basis, an angular deviation weighting strategy is proposed to achieve accurate evaluation of angle deviation in POE strategy. TIOE-Det achieves significant gains on high-precision detection performance. Extensive experiments on multiple datasets prove the superiority of our approach. Codes are available at https://github.com/ming71/TIOE.