MetaAnchor: Learning to Detect Objects with Customized Anchors

TL;DR

MetaAnchor introduces a flexible, dynamically generated anchor mechanism for object detection that improves robustness to anchor settings and enhances transferability across different scenarios.

Contribution

It presents MetaAnchor, a novel approach for dynamically generating anchors, which outperforms traditional fixed anchors in various detection scenarios.

Findings

MetaAnchor outperforms traditional anchors on COCO detection tasks.

MetaAnchor is more robust to anchor settings and bounding box distributions.

MetaAnchor shows potential for transfer learning in object detection.

Abstract

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.