More Focused Remote Sensing Object Detection Based on Improved Bounding Box Regression Loss

Bounding box regression loss in object detection is a metric that measures the difference between the predicted and actual bounding boxes of objects in a picture.The goal is to reduce the loss and improve the precision of object localization.Small objects tend to be difficult to localize, and traditional loss functions are failing to solve this critical issue.For small objects, minimizing the distance of the predicted box is crucial, and focusing on aspect ratios, as seen in CIoU loss, can hinder localization performance.Additionally, these previous losses experience degradation to IoU, especially when the inclusion is present and the center coordinates and box ratios are identical.In such situations, all state-of-the-art losses revert to IoU loss.To address this degradation and place more emphasis on normalizing the distance, we propose a More Focused IoU (MFIoU) loss.By normalizing the bottom-left and top-right corners of both boxes, MFIoU loss becomes more reliable in these scenarios.This paper also incorporates the proposed loss into Non-Maximum Suppression (NMS) to select the optimal predicted box from redundant ones.Our findings demonstrate that the proposed loss delivers superior overall performance on YOLOv4 tested on the PASCAL VOC 2007 dataset, achieving a 34.88% improvement in AP compared to IoU loss as the baseline, a 1.80% improvement on the UAV dataset, and a 35.50% improvement on YOLOv7-Tiny compared to IoU loss.By this, our proposed MFIoU loss and MFIoU-NMS is more suitable for tasks like detecting small objects.