RiO-DETRv2

Pretrained EMA detector weights, training logs, and configurations for RiO-DETRv2, a real-time oriented detector with distributional boundary and angle refinement.

RiO-DETRv2-X achieves 54.13% mAP50-95 on DOTA-v1.0, 53.94% on DIOR-R, and 59.05% AP50 on AI-TOD-R. Five model scales cover compact and high-capacity deployment.

Model Zoo

Accuracy values are percentages. Each metric triple belongs to the listed checkpoint.

DIOR-R

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

DOTA-1.0

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

AI-TOD-R

Model	Params	mAP50-95	AP50	AP75	Weights	Logs	Config
RiO-DETRv2-M	19.5M	18.98	55.53	7.39	pth	metrics / console	yml
RiO-DETRv2-X	64.6M	20.54	59.05	8.52	pth	metrics / console	yml

DOTA evaluation CSVs include epoch-level overall metrics and all 15 category AP values from the original training-time submissions. Metrics use 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.

Evaluation

Use the configuration listed for the selected weights. The checkpoint stores EMA detector parameters under the model key and loads through the implementation's -r argument.

CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 train.py \
  -c configs/rio-diorr/riov2_hgnetv2_x_diorr.yml \
  -r riov2_hgnetv2_x_diorr.pth --test-only --seed=0

The complete configuration tree is included under configs/. Preserve its directory structure when downloading YAML files with relative includes.

Metrics and Logs

• DOTA-v1.0: official test-server results from merged image predictions, using the two-GPU export setup.
• DIOR-R: geometric rotated-box IoU with VOC11 AP interpolation.
• AI-TOD-R: geometric rotated-box IoU with 101-point AP interpolation.
• manifest.json and dataset-level metrics.csv provide checkpoint-level evaluation results.
• Training JSONL files preserve epoch-level numeric records. DOTA submission events use export_epoch.
• Training-log evaluator conventions are recorded in each model manifest; model-zoo values follow the evaluation protocol declared above.
• Console logs are sanitized for machine-specific paths, user identifiers, network addresses, and credentials.
• AI-TOD-R-X includes both base-training and AP50-95 fine-tuning logs.

Integrity

Each model manifest includes the checkpoint SHA256. SHA256SUMS covers the public release files.

License

Apache 2.0. Third-party components and pretrained backbones retain their respective licenses.