An Advanced IoU Loss Function for Accurate Bounding Box Regression

EasyChair Preprint 8186

Abstract

Bounding box regression (BBR) is one of the important steps for object detection. To evaluate accuracy level between true object and prediction object, loss functions for BBR have considered in many researches. In existing researches, the loss functions have some main drawbacks. Firstly, both $l{_n}-norm$ and IOU-based loss functions are inefficient enough to perform the object detection in BBR. Secondly, the loss functions ignore the imbalance issues in BBR that the large number of anchor boxes which have small overlaps with the target boxes contribute most to the optimization of BBR. Thirdly, loss functions own redundant parameters which make for extending process. To address these problems, we propose a new approach by using an Advanced IoU loss function. Three geometric factors are considered in the proposed function including: (i) the overlap area; (ii) the distances; (iii) the side length. 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 performed on MS COCO and VOC Pascal dataset. The results are compared to existing IoU models and show that the proposal can improve detection ability of object for bounding box regression.

Keyphrases: Advanced IoU, Bounding Box Regression, Geometric factors, loss function, object detection

@booklet{EasyChair:8186,
  author    = {Ho-Si-Hung Nguyen and Thi-Hoang-Giang Tran and Dinh-Khoa Tran and Duc-Duong Pham},
  title     = {An Advanced IoU Loss Function for Accurate Bounding Box Regression},
  howpublished = {EasyChair Preprint 8186},
  year      = {EasyChair, 2022}}