Introduce GIoU into RFB net to optimize object detection bounding box
Xiaofan Li, Haibo Pu, Yi Wei, Liu JiangChuan, HongXiang Xu · 2019
RFB algorithm is a very advanced algorithm in the direction of Object Detection based on Deep Learning, RFB Net proposed RFB module to simulate Receptive Fields (RFs) in human visual systems and gain higher accuracy. RFB, however, also have other Object Detection algorithm of faults, that is the detection of choosing predicted bounding box is a box with the highest scores, rather than the most close to the truth boundary box, and the positioning accuracy of bounding box is a core problem, therefore, we put forward a kind of method is replaced IoU(Intersection over Union) with GIoU as predicting boundary box plus or minus a sample selection criteria which can improve the prediction location of bounding box, GIoU (Generalized Intersection over Union) is an improvement over IoU, based on the function of IoU, GIoU considers the offset angle, the distance of truth boundary box and predict boundary box to some extent. We try to replace IoU with GIoU as the discriminant criterion. In the original paper of GIoU algorithm, it is applied to two-stage algorithm, in one-stage algorithm, it only validation in YOLO, while this paper is applied to one-stage algorithm of RFB. Extensive experiments on recognition benchmarks like Pascal VOC2007, Pascal VOC2012, we showed the advantages of using GIoU instead of IoU.