Title: MoEIoU: Re-evaluating Bounding-Box Regression Through a Mixture of Experts Lens

Abstract

Precise object localization in detection systems relies heavily on bounding-box regression, a core mechanism of the process. While conventional Intersection-over-Union (IoU)-centric loss functions attempt to enhance regression performance by integrating static geometric penalties—such as discrepancies in center distance and aspect ratio—these methods often fail to adapt to the evolving nature of training. Specifically, standard penalties do not reflect the optimization trajectory, where early epochs are dominated by significant errors in center alignment and shape, while later phases concentrate on maximizing overlap with ground-truth annotations.

To overcome this rigidity, we present MoEIoU, a novel regression loss grounded in a mixture-of-experts framework. This approach simultaneously models three critical localization factors: overlap, center alignment, and aspect-ratio mismatch. These elements are combined via a log-sum-exp function, a technique that highlights the most prominent localization error while ensuring that secondary terms contribute smoothly to the overall loss. Furthermore, we implement a curriculum-based weighting strategy that dynamically shifts focus: prioritizing the correction of box position and shape during the initial training stages, and subsequently emphasizing overlap refinement as training progresses.

Extensive evaluations of MoEIoU were conducted across PASCAL VOC, HRIPCB, and MS COCO datasets, utilizing various YOLO architectures alongside large-scale simulation studies. The results demonstrate that MoEIoU consistently surpasses both standard and recent state-of-the-art loss functions, delivering superior localization accuracy and accelerated convergence. Additionally, our findings indicate that this adaptive aggregation mechanism can enhance existing IoU-based losses, offering more robust optimization guidance for bounding-box regression within object detection pipelines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC