Oriented objects as pairs of middle lines

Oriented object detection in remote sensing images is an active yet challenging task. State-of-the-art detectors adopt oriented bounding box as a basic representation of an object. The design of them always multiplies the number of anchor boxes with arbitrary angles. Coupled with the rotated Non-Maximum-Suppression (NMS) algorithm, their computational complexities and design difficulties are dramatically increased. In this paper, we propose a novel representation form of oriented objects as pairs of middle lines. Along with the new representation, Oriented Objects Detection Network (O-2-DNet) is proposed to encode and detect these paired middle lines. By eliminating the pre-designed set of anchor boxes and complicated rotated-NMS post-processing, our model runs in an anchor-free and NMS-free manner. Thereby, compared with state-of-the-art bounding-box and anchor-NMS based detectors, O-2-DNet is simpler and more efficient in the oriented object detection task. Experiments show that O-2-DNet achieves excellent performances on the DOTA dataset. In fact, the horizontal objects can also be regarded as a unique form of oriented ones with an angle of 90 degrees. O-2-DNet can also achieve competitive results on the horizontal box dataset, such as DIOR.