MetaAnchor: Learning to Detect Objects with Customized Anchors
This work addresses the need for more adaptable anchor settings in object detection systems, offering a novel approach that enhances robustness and transfer potential, though it is incremental as it builds on existing anchor-based frameworks.
The authors tackled the problem of inflexible anchor mechanisms in object detection by proposing MetaAnchor, which dynamically generates anchor functions from custom prior boxes, resulting in improved robustness and performance on the COCO detection task.
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 transfer tasks. Our experiment on COCO detection task shows that MetaAnchor consistently outperforms the counterparts in various scenarios.