Enhanced Dynamic Anchor Learning for Oriented Object Detection on Satellite Images

In object detection, most of them use horizontal bounding box (HBB). This will be a problem when objects are adjacent to other objects in the oriented direction because it causes a misalignment of the bounding box. To determine the oriented bounding box (OBB), the horizontal anchor needs to be modified to become a rotation anchor. Through a two-stage process, namely by finding HBB and then converting it to OBB, it becomes inefficient. The method that can be used to overcome this problem is Dynamic Anchor Learning based on Convolutional Neural Network (CNN). This research proposes the use of a new building block that is more effective to overcome the weaknesses of ResNet when the network becomes very deep and complex. The proposed method can identify detailed features at multiple levels and expand the scope of what each network layer can perceive by creating relationships between parts of a single residue block. The experiments carried out proved to produce better accuracy than the baseline method with a mAP value of 92.38%.