A Novel Approach of Hybrid Bounding Box Regression Mechanism to Improve Convergency Rate and Accuracy

Abstract

Bounding box regression is a commonly used technique aimed at enhancing the precision of object localization, which is crucial in the field of object recognition. The intersection over union (IoU) metric, which calculates the overlap between predicted and ground truth bounding boxes, is frequently used to evaluate the performance of object detection models. However, the MSE loss function used previously is not compatible with the IoU-based evaluation and has shown sensitivity to differences in object scales. The use of IoU as the basis for loss functions has become more common in recent years, and as a result, new techniques such as the Generalized IoU (GIoU) and complete IoU (CIoU) losses have grown to be developed. This paper introduces a hybrid mechanism called GCIoU loss, which combines GIoU and CIoU losses with the aim of further enhancing localization accuracy and convergence speed. According to our findings, the average precision (AP) is greatly improved by the GCIoU loss in comparison to the GIoU and CIoU losses by 7.72% (highest) to the basis of IoU loss, and 0.87% improvement to the CIoU loss. The GCIoU loss performs consistently well across different thresholds, especially at higher levels, and has improved AP75 by 3.07% compared to the CIoU loss as the default configuration. Additionally, GCIoU loss converges faster and even more robustly in the simulation experiments by taking 14% fewer epochs than the CIoU loss, leading to localization more precisely. By this, GCIoU loss is showing its usefulness in object identification and model optimization.