RotCert research models
Author-trained detector parameters supporting Angle-aware conformal certification for oriented object detection in optical aerial imagery, by Zeyu Fu.
This public collection provides five final evaluated models, their portable configurations, and parameter-verification records, for the CODrone comparison and the four-family DIOR-R validation check. No account or access approval is required to download the files.
| Experimental scope | Models provided |
|---|---|
| DIOR-R train-only validation check | Oriented R-CNN, RoI Transformer, Rotated RTMDet-L and S2ANet, seed 0: 4 models |
| CODrone final-epoch comparison | EAV-DETR, seed 42: 1 model |
The manifest specifies each dataset, detector, modality, seed, experimental scope, configuration, source revision, evaluated parameter branch, original-checkpoint hash and public-file hash. Different aliases for the same saved checkpoint are not counted as additional models. The four DIOR-R families are Oriented R-CNN, RoI Transformer, Rotated RTMDet-L and S2ANet.
This is not the complete model inventory of the article. A separate preservation inventory identified 26 additional candidate checkpoints from DIOR-R main experiments, DOTA-v1.0, DroneVehicle and HRSID. Their full provenance checks, parameter exports and public downloads are not complete, so they are not provided or claimed as verified by this release. Older historical analyses require a further inventory. See the reproduction scope for the remaining work.
What the files preserve
inference.pth contains a CPU tensor mapping, not a complete training checkpoint.
For the MMRotate models, the mapping is the evaluated state_dict. For EAV-DETR,
the mapping is the evaluated model branch, also exposed under model for
loading; this original checkpoint does not contain an EMA branch. The RTMDet
mapping follows the saved evaluation branch;
the alternative optimizer/training EMA state is not substituted.
Every parameter key, dtype, shape and byte value was checked against its original evaluated branch. Each public file also passed a restricted CPU reload, finite parameter checks, and a static scan of serialized strings. No new training, checkpoint selection, or numerical modification was performed. Optimizers, schedulers, data loaders, training configuration objects and host paths are not included in the weights. Only filesystem deployment paths were changed in the separate configuration copies.
These checks establish parameter and file identity. They do not establish a new end-to-end forward-pass reproduction, portability across every runtime, or complete reproduction of every historical experiment. Training resumption and bitwise replay of optimizer trajectories require other records.
Related resources
- GitHub software and aggregate evidence.
- Versioned software archive.
- All software archive versions.
- Loading and evaluation guidance.
- Category order and inference settings.
- Reproduction scope and missing inputs.
- Rights, dataset sources and upstream attribution.
The software DOI identifies its own unchanged snapshot. These separately released model files are not retroactively part of that archive. Dataset images, complete annotations, proprietary fonts, unpublished manuscripts and review correspondence are not included here.
Scientific limitations
Historical DIOR-R method-development exposure is not removed by the validation check. CODrone validation and test share acquisition locations and recordings; tiles are not independent physical observations. Different seeds vary the full procedure, including source-role assignment and operating-point selection.
Native EAV margin-score coverage is not rendered-polygon containment, and its matched population differs from common-HCP comparisons. Localization coverage is conditional on the specified matched-target population; it is not a guarantee of recall, each prediction's conditional coverage, false-positive control, or safe deployment. Unsupported targets, infinite thresholds, failed finite-report gates and adverse results must remain explicit. The article assumes i.i.d. source scenes for its stated results; the release does not verify those assumptions.
Intended use: noncommercial research, inspection and experimental reproduction under the applicable data and upstream terms. These models are not certified for safety-critical operational use. Consult the rights document before redistribution, commercial use or reuse outside the reported experimental scope.