Collaborative learning in bounding box regression for object detection

Abstract

• We explore the defects of the existing localization loss function in object detection. • We propose a method to calculate the localization loss of the bounding box, i.e., SLIoU loss. • We propose a method for filtering many redundant detection boxes, i.e. SLIoU-NMS. • We integrate the proposed methods into the state-of-the-art one-stage object detectors. • Our experiments prove that the effectiveness and efficiency of the proposed methods. Object detection has attracted growing attention due to its extensive application prospect, in which bounding box regression is an essential component. Dedicated to collaborative learning in bounding box regression, we explore the unified framework of smooth ℓ 1 and intersection over union, named SLIoU. On the basis of that, we propose a SLIoU loss as localization loss, which focuses on the geometric relationships of pairs of rectangular bounding boxes in overlapping degree, central position and structural shape. Furthermore, we propose a SLIoU-NMS for suppressing redundant detection boxes, which adaptively maps the evaluation value of detection boxes to meet the evaluation metric using nonlinear representation. By incorporating SLIoU loss and SLIoU-NMS into the state-of-the-art one-stage detectors, the detection performance is considerably improved.