One-Stage Arbitrary-Oriented Object Detection With Prediction Decoupling

Arbitrary-oriented objects widely appear in natural scenes, remote sensing images, text images, etc., and thus arbitrary-oriented object detection has received considerable attention. Arbitrary-oriented object detection is a comprehensive task including multi sub-tasks which are boundary regression, classification and orientation prediction. However, since there is feature misalignment among these sub-tasks, the performance of one-stage detectors without fine-tuning process is greatly limited. Thus, in this study, we explore feature misalignment among sub-tasks in one-stage arbitrary-oriented object detection and find that there are relevance and conflicts among the features of category prediction, foreground score prediction, orientation prediction and boundary prediction. In order to alleviate conflicts and extract especial features for each sub-task, we propose a decoupling detection head composed of three branches respectively for object classification, orientation prediction, boundary regression and foreground score prediction. Especially, boundary regression and foreground score prediction share one branch due to their relevance and feature similarity. What’s more, in order to construct an effective detection head for our decoupling model, we design two decoupling detection heads with different network depths. Experiments indicate that appropriate network depth can improve detection performance, and the proposed detection model achieves 68.91% mAP with single-scale testing on the challenging public dataset DOTA-v1.5. Compared with previous competitive method, our model gains 2.05% mAP.