Multiple space based cascaded center point network for object detection

Abstract

For the purpose of object detection, numerous key points based methods have been suggested. To alleviate the imbalance problem that some objects may be missing when a single-center-point based network is used for object detection, we propose a brand-new multiple space based cascaded center point network (MSCCPNet) for object detection. Particularly, we first bulid a novel structure to alleviate the imbalance problem in detecting different scale objects by scanning more objects in different scale spaces. We then propose a cascaded center point structure to predict the category and confidence of the object by integrating the results of the two centers with the idea of choosing the high confidence and discarding the low confidence. Finally, we determine the object’s location by predicting the center point deviation as well as the width and height of the object. Our MSCCPNet shows competitive accuracy when compared with many sample classical object detection algorithms on GeForce RTX 2080Ti, according to the results of experiments on PASCAL VOC datasets and COCO datasets.