Dynamic Anchor: A Feature-Guided Anchor Strategy for Object Detection

The majority of modern object detectors rely on a set pre-defined anchor boxes, which enhances detection performance dramatically. Nevertheless, the strategy suffers some drawbacks, especially complex hyper-parameters anchors, seriously affecting performance. In this paper, we propose feature-guided generation method named dynamic anchor. Dynamic mainly includes two structures: generator and feature enhancement module. leverages semantic features to predict optimized shapes at locations where objects are likely exist in maps; by converting predicted shape maps into location offsets, module uses high-quality anchors improve Compared with hand-designed scheme, discards all boxes avoids hyper-parameters. addition, only one box is for each location, dramatically reduces calculation. With ResNet-50 ResNet-101 as backbone one-stage detector RetinaNet, achieved 2.1 AP 1.0 gains, respectively. proposed can be easily integrated anchor-based replace traditional scheme.