Inner-IoU: More Effective Intersection over Union Loss Function with Auxiliary Bounding Boxes for Improved Detection Performance
This paper proposes the Inner-IoU loss function, which introduces auxiliary bounding boxes and a scaling factor ratio to adaptively adjust the auxiliary box scale for different IoU samples. High IoU uses small boxes to accelerate convergence, while low IoU uses large boxes to improve performance. Integrating it into existing IoU losses further enhances detection performance.
Existing IoU loss functions cannot adaptively adjust according to detection tasks and have weak generalization. Inner-IoU uses auxiliary bounding boxes and a scaling factor ratio to compute losses with different scale auxiliary boxes for different regression samples: small boxes for high IoU to accelerate, large boxes for low IoU to improve. After integration, detection performance is further improved.
The core mechanism of Inner-IoU is introducing auxiliary bounding boxes, controlling their scale via a scaling factor ratio. For high IoU samples, small auxiliary boxes are used to accelerate convergence; for low IoU samples, large auxiliary boxes improve regression. This method can be integrated into existing IoU loss functions, and its effectiveness and generalization are validated through simulation and comparative experiments. The code is open-sourced.
This work has direct value for the object detection field, improving the accuracy and convergence speed of detectors in scenarios such as autonomous driving and security surveillance, reducing training costs, and promoting practical improvements to IoU loss functions.
It is recommended to integrate Inner-IoU loss into object detection model training, replacing the original IoU loss, which can improve detection accuracy by about 1-2% mAP, especially suitable for industrial detection products with high precision requirements.
Future work needs to validate generalization on more detectors (e.g., YOLOv8, DETR) and large-scale datasets (e.g., COCO, LVIS), and explore automatic learning mechanisms for ratio.