An Advanced IoU Loss Function for Accurate Bounding Box Regression

Abstract

Listen

Translate article

AbstractBounding box regression (BBR) is one of the important steps for object detection. To improve accuracy of recognition ability between true object and prediction object, many researches have developed loss functions for BBR. In existing researches, some main drawbacks can be shown: (i) both IOU-based loss functions and \(l{_n}-norm\) are inefficient enough to detect the object; (ii) the loss functions ignore the imbalance issues in BBR when the large number of anchor boxes have overlaps with the target boxes; (iii) the loss functions own redundant parameters which lead to extend training process. To address these problems, this paper is proposed a new approach by using an Advanced IoU loss function. Three geometric factors including overlap area, distances and side length are considered in the proposed function. The proposal focuses on the overlap area to improve accuracy for object detection. By this way, the proposal can relocate anchor box for covering the ground truth in the training process and optimize anchor boxes for object detection. The proposal is tested on VOC Pascal and MS COCO dataset. The experimental results are compared to existing IoU models and show that the proposal can improve accuracy level for Bounding box regression.KeywordsAdvanced IoUBounding box regressionObject detectionLoss functionGeometric factors