Add Anchor to the Anchor-Free Object Detection Algorithm

Anchor free framework is more in line with the principle of visual semantics, but it falls into the problem of visual regression of the target shape (width, height and rotation angle) directly, with low accuracy. In this work, we propose a method to improve the accuracy of anchor-free frames by using anchor techniques. On the basis of object location information estimation, only a few anchor-assisted target shape regressions are generated. The method integrates preset boxes into anchor-free framework skillfully and avoids the huge number of anchors in anchor-based scheme. In addition, we have implemented two detection algorithms for horizontal and rotating frames. Compared with the existing anchor-based and anchor-free schemes, our scheme achieves an optimal balance between detection accuracy and computation. The experiments on the two public data sets, MS COCO and HRS2016, fully showed that the precision of our AP reached state-of-the-art.