Summary: This paper proposes MetaAnchor, which is a anchor mechanism for object detection. In MetaAnchor, anchor functions are dynamically generated from anchor box bi, which describes the common properties of object boxes associated with i_th bin. It introduces a anchor function generator which maps any bounding box prior bi to the corresponding anchor function. In this paper, the anchor function generator is modeled as two-layer network for residual term R, added to the shared and learnable parameters for the anchor function theta *. The residual term R can also depends on input feature x, which introduces the data-dependent variant of anchor function generator. Using weight prediction mechanism, anchor function generator could be implemented and embedded into existing object detection frameworks for joint optimization.

We propose a novel and flexible anchor mechanism named MetaAnchor for object detection frameworks. Unlike many previous detectors model anchors via a predefined manner, in MetaAnchor anchor functions could be dynamically generated from the arbitrary customized prior boxes. Taking advantage of weight prediction, MetaAnchor is able to work with most of the anchor-based object detection systems such as RetinaNet. Compared with the predefined anchor scheme, we empirically find that MetaAnchor is more robust to anchor settings and bounding box distributions; in addition, it also shows the potential on the transfer task. Our experiment on COCO detection task shows MetaAnchor consistently outperforms the counterparts in various scenarios.