SIoU Loss: More Powerful Learning for Bounding Box Regression

TL;DR

This paper introduces SIoU, a novel loss function for bounding box regression in object detection that considers the direction of mismatch, leading to faster training and improved accuracy.

Contribution

The paper proposes SIoU, a new loss function that incorporates angle-based penalties, enhancing convergence speed and detection accuracy over existing methods.

Findings

SIoU improves training speed in neural networks.

SIoU enhances bounding box regression accuracy.

SIoU outperforms traditional loss functions in tests.

Abstract

The effectiveness of Object Detection, one of the central problems in computer vision tasks, highly depends on the definition of the loss function - a measure of how accurately your ML model can predict the expected outcome. Conventional object detection loss functions depend on aggregation of metrics of bounding box regression such as the distance, overlap area and aspect ratio of the predicted and ground truth boxes (i.e. GIoU, CIoU, ICIoU etc). However, none of the methods proposed and used to date considers the direction of the mismatch between the desired ground box and the predicted, "experimental" box. This shortage results in slower and less effective convergence as the predicted box can "wander around" during the training process and eventually end up producing a worse model. In this paper a new loss function SIoU was suggested, where penalty metrics were redefined considering the angle of the vector between the desired regression. Applied to conventional Neural Networks and datasets it is shown that SIoU improves both the speed of training and the accuracy of the inference. The effectiveness of the proposed loss function was revealed in a number of simulations and tests.