中科大联合华为发布旋转检测模型 RiO-DETR，端到端延迟 2.7 ms 达 78.4 AP50
RiO-DETR 让 DETR 旋转检测首次实现实时推理，2.7 ms 达到 78.4 AP50，获 ECCV 2026 Oral。
旋转目标检测在遥感影像等场景中比水平框更贴合目标，但 DETR 因其注意力计算和角度周期性问题难以实时。RiO-DETR 通过内容驱动角度估计、旋转校正正交注意力等设计，首次将端到端 DETR 旋转检测推入实时区间：最轻量版仅 2.7 ms 延迟，AP50 达 78.4。该工作已被 ECCV 2026 接收为 Oral。
正文摘录
.jpg)  In tasks such as remote sensing imagery, traffic perception, and scene text recognition, the objects to be detected are rarely arranged in neat horizontal or vertical patterns. Horizontal bounding boxes can enclose objects but often incorporate large amounts of irrelevant background；