SAIOU: A SYNERGISTIC ADAPTIVE IOU LOSS FOR BOUNDARY BOX REGRESSION

Bounding Box Regression (BBR) is a critical component of modern object detectors, with its performance heavily dependent on the loss function design. However, existing IoU-based losses exhibit significant shortcomings in synergistically optimizing for target distance, shape, and prediction quality. This often leads to slow convergence, insufficient localization accuracy, and imbalanced penalties for small objects. To address these challenges, we propose a novel Synergistic Adaptive IoU (SAIoU) loss. Within a unified framework, SAIoU integrates three synergistic mechanisms: a scaleadaptive distance penalty, aspect ratio-guided shape-aware modeling, and a prediction quality-driven gradient modulation. Extensive experiments on the MS COCO and African Wildlife datasets demonstrate that integrating SAIoU into advanced detectors like YOLOv8 consistently outperforms nine mainstream loss functions, including PIoU. SAIoU shows significant advantages in localization accuracy and generalization, validating the effectiveness of its synergistic optimization strategy, particularly for small objects and complex backgrounds.