MEAD: a Mask-guidEd Anchor-free Detector for oriented aerial object detection

Object detection in aerial images is a challenging task due to various orientations of objects and the lack of discriminative features. Existing methods are usually in a dilemma between accuracy and speed. While one-stage anchor-free detectors inference more quickly than two-stage frameworks, their predictions are not as accurate as that of the opposite. This paper proposes a quick and accurate detector, Mask-guidEd Anchor-free Detector (MEAD). It can rapidly locate oriented objects in aerial images by means of per-pixel prediction. Furthermore, it embeds a cascade architecture to locate targets more precisely. To enhance feature discrimination, the mask-guided branch is employed to force features to attend the foreground regions. Comparative experiments are conducted on DOTA and HRSC2016 datasets. The results show that MEAD is better than current state-of-the-art anchor-free detectors, that is, mAP 74.33 on DOTA and 89.83 on HRSC2016.