Arbitrary-oriented Object Detection Based on Sub-regional RoI Align

Abstract:
When detecting vehicles, ships and some other types of targets in satellite or UAV remote sensing images, due to the large camera pitch angle and long imaging distance, the targets to be detected is generally small and arbitrary-oriented, and the arrangement density is often large. Conventional object detection methods using axis-aligned bounding box are prone to miss detection in these scenarios. In order to solve this problem, a target detection method suited for remote sensing targets is proposed, in which a rotational region proposals network is introduced. in the training process of network, the definition of positive and negative anchors, the training strategy and the non-maximum suppression method are improved. Furthermore, a feature mapping method for arbitrary oriented region of interest is proposed. Experiment results on the DOTA dataset demonstrate that the proposed method can detect arbitrary-oriented targets effectively.