RiO-DETRv2: Geometry-Aware Uncertainty Modeling Meets Real-Time Oriented Object Detection Transformer

Zhangchi Hu¹, Zilong He¹, Yifan Zhao¹, Xiangyu Yu¹, Jie Chen¹, Peixi Wu¹, Yansong Peng¹, Hebei Li^1,*, Xiaoyan Sun^1,2

1. University of Science and Technology of China

2. Institute of Artificial Intelligence, Hefei Comprehensive National Science Center

Corresponding author: lihebei@mail.ustc.edu.cn

Highlights · Demo · Model Zoo · Getting Started · Usage · Deployment

RiO-DETRv2 is a real-time, end-to-end oriented object detector that refines both object boundaries and orientation through learned distributions.

DOTA-v1.0 accuracy versus inference latency on NVIDIA T4.

Highlights

• Fine-grained object-aligned localization. Four-side boundary distributions support progressive refinement in an object-aligned coordinate system.
• Coarse-to-fine orientation refinement. Each decoder layer predicts an angular residual around its current reference angle, with progressively finer refinement ranges.
• Boundary-and-angle-aware confidence. Boundary distribution concentration and angular entropy inform detection scores, with object shape providing context for orientation ambiguity.
• A scalable real-time detector family. N, S, M, L, and X configurations share the same oriented-detection framework, with ONNX export and TensorRT FP16 deployment utilities.

Demo: Bogota Airport

We challenge RiO-DETRv2 with high-resolution Jilin-1 satellite footage of Bogota Airport, Colombia. Under moving cloud cover and dense aircraft layouts, the model delivers well-aligned oriented detections across successive frames, showcasing its capability in dynamic satellite scenes.

riov2_bogota_airport_1080p.mp4

RiO-DETRv2-X uses DOTA-1.0 pretrained weights and tiled inference. Video credit: Paul Sed, "Jilin-1 Live streaming from space - Bogota, Colombia".

Model Zoo

Download pretrained EMA checkpoints and training logs from Hugging Face.

Release metadata: manifest · SHA256 checksums.

DOTA-v1.0 Single-Scale

Results are evaluated on the DOTA-v1.0 test set by the official evaluation server, using 1024 x 1024 image patches and merged full-image predictions.

Model	Params	mAP50-95	AP50	AP75	Config	Weights	Logs
RiO-DETRv2-N	4.3M	49.20	79.10	50.94	yml	pth	train · console · eval
RiO-DETRv2-S	8.9M	51.82	79.56	56.36	yml	pth	train · console · eval
RiO-DETRv2-M	19.5M	53.20	81.20	57.30	yml	pth	train · console · eval
RiO-DETRv2-L	28.6M	53.63	80.77	58.69	yml	pth	train · console · eval
RiO-DETRv2-X	64.6M	54.13	80.99	59.44	yml	pth	train · console · eval

Evaluation CSVs contain the original epoch-level official test results, including VOC mAP, AP50, AP75, COCO mAP50-95, and all 15 category AP values. CSV metrics are fractions in [0, 1]; epochs are zero-based and match the checkpoint filenames. Released checkpoint epochs: N 142, S 132, M 102, L 84, X 64.

DIOR-R

Results use 800 x 800 inputs on the DIOR-R test split. AP is evaluated with geometric rotated-box IoU and VOC-style 11-point interpolation over the listed IoU thresholds.

Model	Params	mAP50-95	AP50	AP75	Config	Weights	Logs
RiO-DETRv2-N	4.3M	46.46	72.17	47.67	yml	pth	train · console
RiO-DETRv2-S	8.9M	49.95	75.11	51.27	yml	pth	train · console
RiO-DETRv2-M	19.5M	52.60	77.27	54.30	yml	pth	train · console
RiO-DETRv2-L	28.6M	52.88	77.53	54.41	yml	pth	train · console
RiO-DETRv2-X	64.6M	53.94	78.30	56.11	yml	pth	train · console

Getting Started

Installation

Create a Python 3.10 environment and install PyTorch with a matching CUDA runtime. For CUDA 12.1:

conda create -n rio-detr-v2 python=3.10 -y
conda activate rio-detr-v2

pip install torch==2.2.2 torchvision==0.17.2 \
  --index-url https://download.pytorch.org/whl/cu121
pip install "numpy<2" -r requirements.txt

Run the following commands from the repository root. HGNetv2 ImageNet-pretrained backbone weights are downloaded automatically and cached under pretrain/hgnetv2/.

Download Pretrained Models

Download an individual checkpoint with the Hugging Face CLI:

pip install "huggingface_hub>=0.34,<1"
hf download RicePasteM/RiO-DETRv2 \
  diorr/riov2_hgnetv2_x_diorr.pth \
  --local-dir pretrained/riov2

Omit the filename to download the complete model collection, configurations, and training logs.

Training with DINOv3

The training configurations use DINOv3 as a feature teacher. Arrange its source and ViT-B/16 pretrained checkpoint as follows:

third_party/
└── dinov3/
pretrain/
└── dinov3/
    └── dinov3_vitb16_pretrain_lvd1689m.pth

The teacher is used only during training. DINOv3 source and pretrained weights follow their respective licenses.

Data Preparation

The default configurations use repository-relative dataset paths. Create symlinks under datasets/ or edit the matching dataset YAML file for your installation.

DOTA-v1.0

Prepare the original images and polygon annotations as 1024 x 1024 patches using the standard DOTA patch naming convention:

datasets/dota-1.0/
├── single-scale/
│   ├── trainval/
│   │   ├── images/
│   │   └── annfiles/
│   └── test/
│       ├── images/
│       └── annfiles/
└── multi-scale/
    ├── trainval/
    │   ├── images/
    │   └── annfiles/
    └── test/
        ├── images/
        └── annfiles/

Each annotation file contains one polygon per line:

x1 y1 x2 y2 x3 y3 x4 y4 class_name difficulty

For unlabeled test patches, provide an empty annotation file with the same image stem. Paths are configured in single-scale and multi-scale dataset definitions.

The offline evaluator merges patch predictions and generates a submission ZIP for official DOTA evaluation. DOTA-Auto-Eval provides a companion workflow for submission management.

DIOR-R

datasets/dior-r/
├── JPEGImages-trainval/
├── JPEGImages-test/
├── Annotations/
│   └── Oriented Bounding Boxes/
└── ImageSets/
    └── Main/
        ├── train.txt
        ├── val.txt
        └── test.txt

The detector trains on the combined train/validation split and evaluates on the test split. Paths are defined in dior_detection.yml.

AI-TOD-R

datasets/ai-tod-r/
├── images/
│   ├── trainval/
│   └── test/
└── labels_dota/
    ├── trainval/
    └── test/

Use DOTA-format polygon annotations with the eight AI-TOD-R class names. The dataset definition specifies the trainval/test paths, evaluated image IDs, and 101-point evaluator.

FAIR1M-2.0

Multi-scale training configurations are also provided for FAIR1M-2.0. Convert annotations to DOTA-format polygon text files and arrange the processed patches under:

datasets/fair1m-2.0/
├── train/
│   ├── images/
│   └── annfiles/
└── validation/
    ├── images/
    └── annfiles/

Usage

Available Configurations

Dataset	Scales	Configuration
DIOR-R	N, S, M, L, X	configs/rio-diorr/riov2_hgnetv2_{n,s,m,l,x}_diorr.yml
DOTA-v1.0 SS	N, S, M, L, X	configs/rio-dota_1_ss/riov2_hgnetv2_{n,s,m,l,x}_dota_1.yml
DOTA-v1.0 MS	M, X	configs/rio-dota_1_ms/riov2_hgnetv2_{m,x}_dota_1.yml
FAIR1M-2.0 MS	M, X	configs/rio-fair1m_ms/riov2_hgnetv2_{m,x}_fair_2.yml
AI-TOD-R	M, X	configs/rio-aitodr/riov2_hgnetv2_{m,x}_aitodr.yml

AI-TOD-R-X also provides AP50-95 fine-tuning, with AP50-95 checkpoint selection and a dedicated output directory. Use -r with a base X training checkpoint to continue training with this configuration.

Training

Train RiO-DETRv2-X on DIOR-R with two GPUs:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
  --nproc_per_node=2 --master_port=7001 \
  train.py \
  -c configs/rio-diorr/riov2_hgnetv2_x_diorr.yml \
  --use-amp --seed=0

Set --nproc_per_node to the number of visible GPUs. The configuration's total_batch_size is shared across all processes. Training checkpoints and logs are saved to output_dir; use --output-dir to select a different destination.

Resume a training run with its full training checkpoint:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
  --nproc_per_node=2 --master_port=7001 \
  train.py \
  -c configs/rio-diorr/riov2_hgnetv2_x_diorr.yml \
  -r outputs/riov2_hgnetv2_x_diorr/last.pth \
  --use-amp --seed=0

Evaluation

Evaluate a DIOR-R checkpoint and report AP50, AP75, and mAP50-95:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
  --nproc_per_node=2 --master_port=7001 \
  train.py \
  -c configs/rio-diorr/riov2_hgnetv2_x_diorr.yml \
  -r /path/to/riov2_x_diorr.pth \
  --test-only --seed=0 \
  -u 'evaluator.iou_thrs=[0.50,0.55,0.60,0.65,0.70,0.75,0.80,0.85,0.90,0.95]'

Generate a DOTA-v1.0 single-scale submission with the two-GPU evaluation setup:

CUDA_VISIBLE_DEVICES=0,1 torchrun \
  --nproc_per_node=2 --master_port=7001 \
  train.py \
  -c configs/rio-dota_1_ss/riov2_hgnetv2_x_dota_1.yml \
  -r /path/to/riov2_x_dota_1_ss.pth \
  --test-only --seed=0

Merged submission archives are written under <output_dir>/dota_results/. Submit the ZIP to the official DOTA server for test-set metrics. DIOR-R and AI-TOD-R report local metrics from their labeled test splits.

Deployment

The deployment tools support ONNX export, TensorRT engine construction, and GPU inference benchmarking.

Export a 1024 x 1024 DOTA model to ONNX:

pip install onnx onnxsim

python tools/deployment/export_onnx.py \
  -c configs/rio-dota_1_ss/riov2_hgnetv2_x_dota_1.yml \
  -r /path/to/riov2_x_dota_1_ss.pth

The ONNX file is saved next to the checkpoint. With TensorRT installed, build an FP16 engine and benchmark it:

python tools/deployment/export_engine.py \
  --onnx /path/to/riov2_x_dota_1_ss.onnx \
  --saveEngine /path/to/riov2_x_dota_1_ss_fp16.engine \
  --fp16

python tools/deployment/benchmark_tensorrt.py \
  --engine /path/to/riov2_x_dota_1_ss_fp16.engine \
  --warmup 100 --num-runs 1000

The benchmark uses GPU-resident buffers and CUDA events to measure TensorRT execution latency. Preprocessing, host/device transfers, and host-side postprocessing are outside this measurement. Add --use-cuda-graph for CUDA Graph benchmarking.

License

RiO-DETRv2 is released under the GNU Affero General Public License v3.0. Third-party code, foundation models, and pretrained weights retain their respective licenses.

Acknowledgements

RiO-DETRv2 builds on the real-time oriented-detection foundation of RiO-DETR. We thank the authors of RT-DETR, RT-DETRv4, D-FINE, DEIMv2, DINOv3, and RHINO for their open-source contributions.