RiO-DETRv2

RiO-DETRv2 is a real-time oriented-box DETR. It predicts four box sides as distributions in the box's own frame, refines the angle coarse to fine and scores each box with an angle-aware quality. LibreYOLO ports it inference-only, in five sizes, with DOTA-v1.0 and DIOR-R checkpoints.

Задачи
oriented boxes
Установка
pip install libreyolo
Уровень поддержки
Только инференс, начиная с v1.8.0. Только предсказание, валидация и экспорт. Функции обучения неприменимы.
Исходный проект
RiO-DETRv2 от RiO-DETRv2 authors (RicePasteM), Apache-2.0. Статья, исходный код
Лицензии
Код: Apache-2.0, веса: Apache-2.0. Коммерческое использование

Install

RiO-DETRv2 needs no optional extra.

bash
pip install libreyolo

Predict

Weights download from the LibreYOLO Hugging Face mirrors on first use. The models are trained on aerial imagery, so use an overhead photo.

Python
from libreyolo import LibreYOLO model = LibreYOLO("LibreRiODETRv2s-obb.pt")        # DOTA-v1.0, 15 classesresult = model("aerial.jpg", save=True) print(result.obb.xywhr)   # (N, 5) centre, size and angleprint(result.obb.conf, result.obb.cls)
DIOR-R
from libreyolo import LibreYOLO model = LibreYOLO("LibreRiODETRv2s-obb-diorr.pt")  # DIOR-R, 20 classesresult = model("aerial.jpg")
CLI
libreyolo predict model=LibreRiODETRv2s-obb.pt source=aerial.jpg save=True

result.obb holds the rotated boxes: xywhr (centre, width, height and angle), conf and cls, with class names from the checkpoint. See the oriented detection task page for the payload and plotting.

Variants

Five sizes, n, s, m, l and x, each with two checkpoints:

CheckpointDatasetClassesCanvas
LibreRiODETRv2<size>-obb.ptDOTA-v1.0151024 px
LibreRiODETRv2<size>-obb-diorr.ptDIOR-R20800 px

The native canvas is read from the checkpoint. Raw outputs and decoded rotated boxes match the upstream Apache-2.0 release exactly (max absolute difference 0) for all ten checkpoints on CPU.

RiO-DETRv2 is inference-only. train() raises: the upstream recipe (DINOv3 feature distillation, OBB Hungarian matching, boundary and angle distribution losses) is not ported. For trainable oriented-box detection use GTR or YOLO-NAS.

Export

ЗадачаONNXTorchScriptExecuTorchTensorRTOpenVINOPaddleMNNRKNNAxelerancnnTFLiteCoreMLCore AI
Oriented boxesOriented boxes to ONNX: поддерживаетсяOriented boxes to TorchScript: поддерживаетсяOriented boxes to ExecuTorch: поддерживаетсяOriented boxes to TensorRT: поддерживаетсяOriented boxes to OpenVINO: поддерживаетсяOriented boxes to Paddle: не поддерживаетсяOriented boxes to MNN: не поддерживаетсяOriented boxes to RKNN: не поддерживаетсяOriented boxes to Axelera: не поддерживаетсяOriented boxes to ncnn: не поддерживаетсяOriented boxes to TFLite: не поддерживаетсяOriented boxes to CoreML: не поддерживаетсяOriented boxes to Core AI: не поддерживается

ONNX and TorchScript exports round-trip: the exported file loads back through LibreYOLO() and returns the same Results.

Python
from libreyolo import LibreYOLO model = LibreYOLO("LibreRiODETRv2s-obb.pt")model.export(format="onnx")model.export(format="torchscript")
CLI
libreyolo export model=LibreRiODETRv2s-obb.pt format=onnx

Checkpoints

Все перечисленные выше файлы уже доступны в организации LibreYOLO и скачиваются при первом использовании.

Licensing

Проверяйте лицензию в репозитории конкретных весов на Hugging Face. Она указана для каждого чекпойнта в организации LibreYOLO и может различаться даже внутри одного семейства. Этот репозиторий считается авторитетным источником, а сводка ниже описывает условия на момент последней проверки страницы.

Это описание соответствующих лицензий, а не юридическая консультация. Если ответ важен для коммерческого использования, прочитайте лицензии самостоятельно и проконсультируйтесь с юристом.

Оригинальная работа
RiO-DETRv2, RiO-DETRv2 authors (RicePasteM)
Лицензия исходного проекта
Apache-2.0
Исходный код проекта
github.com/RicePasteM/RiO-DETRv2
Код LibreYOLO
MIT
Веса
Apache-2.0, повторно опубликованы на huggingface.co/LibreYOLO
Толкование
Apache-2.0 is a permissive license, so the RiO-DETRv2 code and checkpoints can be used in commercial and closed-source products. Redistributing them means keeping the license text and notices, and the license grants a patent license. About 33 minutes after its initial commit (ab8b22df5d800a927774c8c67e77fd29c0fe6967), upstream changed its LICENSE file to AGPL-3.0; later commits change only the README and LICENSE, not the code. LibreYOLO's port relies on the Apache-2.0 grant of that initial release, uses nothing from later commits, and preserves the Apache-2.0 snapshot at LibreYOLO/RiO-DETRv2 (tag apache-2.0-ab8b22d). The adapted decoder keeps Apache-2.0 with its RT-DETR, D-FINE, DEIM and RT-DETRv4 copyright lines (libreyolo/models/riodetrv2/LICENSE and NOTICE); the HGNetv2 backbone, encoder and decoder blocks reuse LibreYOLO's dfine and rtdetrv2 families, and the wiring around them is LibreYOLO's MIT code. The official checkpoints (Hugging Face RicePasteM/RiO-DETRv2 at revision 5f578f57929aded1962ebf8c5584a4771461124f, tagged apache-2.0) are mirrored as LibreYOLO/LibreRiODETRv2{n,s,m,l,x}-obb (DOTA-v1.0) and -obb-diorr (DIOR-R) with the upstream LICENSE and NOTICE; conversion renames keys and drops a training-only projector, learned tensors are unchanged. The DOTA-v1.0 and DIOR-R training sets carry their own terms.

RiO-DETRv2 is by RicePasteM (RicePasteM/RiO-DETRv2). Code and weights are Apache-2.0, so commercial use is fine. Upstream changed its LICENSE file to AGPL-3.0 about 33 minutes after its initial commit, with no code changes. LibreYOLO's port uses only that initial Apache-2.0 release (commit ab8b22d) and preserves the snapshot at LibreYOLO/RiO-DETRv2. The official checkpoints are mirrored as LibreYOLO/LibreRiODETRv2{n,s,m,l,x}-obb and -obb-diorr, with keys renamed and learned tensors unchanged. The DOTA-v1.0 and DIOR-R training sets carry their own terms.

Citation

@misc{hu2026riodetrv2geometryawareuncertaintymodeling,
      title={RiO-DETRv2: Geometry-Aware Uncertainty Modeling Meets Real-Time Oriented Object Detection Transformer}, 
      author={Zhangchi Hu and Zilong He and Yifan Zhao and Xiangyu Yu and Jie Chen and Peixi Wu and Yansong Peng and Hebei Li and Xiaoyan Sun},
      year={2026},
      eprint={2610.29159},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.29159}, 
}

Скопировано из блока цитирования авторов на странице arxiv.org/bibtex/2610.29159.

Проверено с LibreYOLO v1.8.0. Таблицы поддержки, чекпойнты и результаты бенчмарков на этой странице сгенерированы из выпущенной библиотеки и опубликованных весов, а не написаны вручную.