Object Detection with Auto-Learning Anchor Algorithm

As an effective auxiliary means for object detection task, region anchors are widely adopted in most of state-of-the art detectors. However, anchor's location and shape in those works are normally determined by experience or some preprocessing methods, i.e., clustering, which leads to time consumption and limits the flexibility of anchors. In this paper, we explore the possibility that networks can predict bounding boxes and simultaneously learn anchor's location and shape end to end. Specifically, we propose a new anchoring scheme, named Automatic Anchor Learning, which can be integrated into any object detectors and enable detectors to learn the location and size of anchors while training, without sampling anchors over any predefined set of scales and aspect ratios. The proposed method first predicts where the centers of objects of interest might exist and then predict the shape of anchor that should be placed in this location. By applying the proposed Automatic Anchoring Learning method to Yolov3 model, we achieve around 3.3% and 1.6% higher recall and mAP on MS COCO with 80% less anchors, and 10% more FPS than the original Yolov3. Additionally, we also integrate our method into other object algorithms, i.e., Fast R-CNN and RetinaNet, we respectively improve their detection mAP by 2.5% and 1.1%.