OpenDroneKitRegistered models
- Crack segmentationSegFormer-B5 (run crack_segformer_b5, 40 epochs). Weakness: At 0.85 thin and faint cracks are missed, so an empty mask is not evidence of a sound surface.
- Corrosion severitySegFormer-B2 (run corrosion_segformer_b2_cs, 80 epochs, log-inverse class weighting). Weakness: Errors are almost entirely between adjacent grades: of the 653,071 severe pixels it recovers 514,827 (recall 0.788) and loses 124,241 of the rest to 'poor', one step down. A reported grade can be one step optimistic.
- Solar cell defects (electroluminescence)YOLO11l (run pvel_ad_yolo11l, 60 epochs). Weakness: EL cell-level imagery only (module manufacturing and lab inspection). It is not an aerial or thermal model and must not be applied to drone photographs of installed panels.
- Solar thermal anomaliesResNet18 (run solar_thermal_cls, 26 epochs, early-stopped). Weakness: Soiling is the weak class at 0.367 recall and 0.344 precision on 30 validation samples: roughly two in three soiled modules are missed, and a third of soiling calls are wrong.
The registry also lists structural, rail and land-cover models; their figures are in docs/features/registry.py. These four are the ones whose numbers appear on this site.
Rejected models
Trained, measured, and not shipped. The reasons are kept in the repository rather than the runs being deleted.
- RGB solar panel-condition detector (Clear/Dusty/Damage/Snow): mAP50 0.318; Dusty class 0.073. Dusty, the class Indian sites most need, was its worst class. Deliberately not registered: soiling looks like a radiometric comparison problem rather than a detection one. source
- Single-class corrosion detector (YOLO11l, 498 images): mAP50 0.254, recall 0.257. Three corrosion sites in four went unfound. Rejected rather than registered and replaced by severity segmentation. source
- Mine change detection: IoU 0.295 on about 15 held-out mines. 60 per cent of its flags were wrong; out of v1 scope by decision. source
How the numbers are kept honest
- A number appears only with the split it was measured on.
- Each registry entry names the model’s known weakness; the cards repeat it next to the number.
- The weights’ SHA-256 is recorded, and a different installed file is reported as a mismatch.
- A missing model produces a refusal with a reason, never a heuristic reported as AI.