Confine Keypoint Triplets for Object Detection

Abstract

Corner-based object detection algorithm is a hot topic in academia and industry. Scholars have applied it to different scenarios such as 3D object detection, object tracking, pose estimation and so on. The core of this method is to detect and match corner points. Although a considerable number of papers have been studied on this issue, there are still a lot of missed detection and false detection. In this paper, we pro-posed a one-stage object detection method with confine key- point triplets(ConfineNet), which applies center pooling and corner matching limit in local feature region. This restrictions can provide stronger correlation features with the object for subsequent prediction of center points, which can help to improve the prediction results of center points and corner matching. In addition, ConfineNet learns the association distance between the upper-left corner point and the lower-right corner point, which can confine the corner matching within a local area, thereby reducing matching errors. The AP value of Confine- Net reaches 48.2% on MS-COCO test-dev. our ConfineNet not only outperforms all existing anchor-free detectors but also achieves comparable performance to the state-of-the-art of two-stage detection approaches.