Abstract:Bounding box regression is a crucial step in most object detection algorithms, and directly affects the positioning accuracy and regression speed of convolutional neural networks (CNN). The existing loss functions commonly used in bounding box regression suffer two main disadvantages: firstly, the ln-norm loss does not match the evaluation metric Intersection over Union (IOU), leading to poor regression performance. Second, some recently proposed IOU-based loss functions are beneficial to IOU metric, but the negative effects of some terms in these loss functions on bounding box regression lead to slow convergence and inaccurate regression results. To solve these shortcomings, we proposed a Manhattan-Distance IOU (MIOU) loss function here. It takes into account that the Euclidean distance term in the Complete IOU (CIOU) loss and the Efficient IOU (EIOU) loss is unstable in training due to the huge gradient in the early stage of regression, and the Manhattan distance is added to effectively alleviate this defect. In addition, the denominator of the Euclidean distance term in the two loss functions discussed above has an antagonistic effect on loss reduction, and setting it as a normalized coefficient without participating in backpropagation can effectively improve the convergence speed. The effectiveness of the proposed MIOU loss was verified with designed simulation experiments. Moreover, object detection is usually applied to natural scenes and remote sensing scenes, but the application of detection methods are often limited due to varied image characteristics in different scene settings. We incorporated the MIOU loss into YOLO v4 and other mainstream object detection networks to examine its effectiveness in remote sensing and natural object detection scenarios. The experimental results on real remote sensing datasets DOTA and natural datasets MS COCO demonstrate that the MIOU loss has strong robustness in both remote sensing object detection tasks and natural object detection tasks. In summary, as a general regression loss function, the MIOU loss shows excellent performance in the above two types of scenes.

Optimized Loss Functions for Object detection: A Case Study on Nighttime Vehicle Detection

Low Illumination Enhancement for Object Detection in Self-Driving

Searching Parameterized AP Loss for Object Detection

Manhattan-distance IOU loss for fast and accurate bounding box regression and object detection

Improving 3D Object Detection Via Joint Attribute-oriented 3D Loss

Application of an Improved Focal Loss in Vehicle Detection.

Unified-IoU: For High-Quality Object Detection

Multiclass objects detection algorithm using DarkNet-53 and DenseNet for intelligent vehicles

Loss Function Discovery for Object Detection Via Convergence-Simulation Driven Search

Diag-IoU Loss for Object Detection

Optimization for Anchor-Free Object Detection via Scale-Independent GIoU Loss

Improved Design Based on IoU Loss Functions for Bounding Box Regression

Nighttime vehicle detection algorithm based on image translation technology1

Directly Optimizing IoU for Bounding Box Localization

IoU-aware Single-stage Object Detector for Accurate Localization

Research on night-time vehicle target detection based on improved KSC-YOLO V5

Improved Vision-Based Vehicle Detection and Classification by Optimized YOLOv4

Offset Bin Classification Network for Accurate Object Detection

Overcoming the Challenges of Long-tail Distribution in Night-time Vehicle Detection

Rethink the IoU-based Loss Functions for Bounding Box Regression