Learning Object-Wise Semantic Representation for Detection in Remote Sensing Imagery

Abstract

With the upgrade of remote sensing technology, object detection in remote sensing imagery becomes a critical but also challenging problem in the field of computer vision. To deal with highly complex background and extreme variation of object scales, we propose to learn a novel object-wise semantic representation for boosting the performance of detection task in remote sensing imagery. An enhanced feature pyramid network is first designed to better extract hierarchical discriminative visual features. To suppress background clutter as well as better estimate proposals, next we specifically introduce a semantic segmentation module to guide horizontal proposals detection. Finally, a ROI module which can fuses multiple-level features is proposed to further promote object detection performance for both horizontal and rotate bounding boxes. With the proposed approach, we achieve 79.5% mAP and 76.6% mAP in horizontal bounding boxes (HBB) and oriented bounding boxes (OBB) tasks of DOTA-v1.5 dataset, which takes the first and second place in the DOAI2019 challenge, respectively.