AIoU: An Adaptive Bounding Box Loss with an Edge Alignment Penalty

Le résumé fourni par la source

Intersection-over-Union (IoU)-based losses, such as IoU and CIoU, have become the de facto standard for bounding box regression in modern object detectors. However, these losses mainly emphasize center distance and aspect ratio consistency, while providing limited supervision on explicit boundary alignment, which is particularly critical for small, dense, and overlapping objects. In this paper, we propose AIoU, an enhanced IoU-based loss that introduces an explicit edge alignment penalty together with a dynamic weighting mechanism. By directly penalizing boundary misalignment and adaptively balancing different geometric factors during training, AIoU provides more stable and accurate localization supervision. AIoU is simple to integrate into existing detection frameworks and introduces no additional inference overhead. Extensive experiments on COCO and VisDrone benchmarks across multiple YOLO variants demonstrate consistent performance improvements, validating the effectiveness and practicality of the proposed method.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.