AMB:Automatically Matches Boxes Module for One-Stage Object Detection
Zhandong Wu, Haiye Yu, Lei Zhang, Yuanyuan Sui · 2023
In computer vision, object detection is an important task. Whether it is a single-stage or two-stage method, it has been based on anchor-based detectors for several years. In recent years, due to multi-scale feature fusion and Focal Loss, anchor-free detectors have become popular. However, both anchor-based and anchor-free detectors have the problems of large computation, poor stability, many hyperparameters, and imbalance between positive and negative samples. This paper proposes a one-stage target detection network (AMBNet) that automatically matches ground truth boxes, avoid manual selection of anchors, which is difficult to match multi-scale targets. The size of the grid can be adjusted according to the size of the detection target. It is still regressing the target from rectangles that are closer to the ground-truth boxes, so the training is more stable. By calculation, when the number of grids is 64, no concentric rectangles are set, and the average IOU is 0.7753. When setting 4 concentric rectangles, the average value of IOU is 0.8211.