fireviewer/fire-smoke-yolo11m-strict-v1

FireViewer YOLO11-M strict v1

Research checkpoint trained for two-class visible smoke and flame detection on the FireViewer strict-clean detection corpus, revision 8e3cb52e86e15debdff7d37a3ce9eac39fbe1835.

This release is intentionally not presented as the primary FireViewer detector. Its independent flame performance and recall are too weak for that claim. It is published as a reproducible baseline and iteration artifact.

Classes

• 0: smoke_visible
• 1: flame_visible

Independent benchmark

The model revision a75aa44716658f02eb85b4e7bba6a358040e6f86 was evaluated on 512 images and 854 boxes from Hajorda/flameye-wildfire-detection@test, revision 361a3dea8b877482af9f4ed80eff77ffd39926ef. One exact overlap was removed; the final overlap with 80,190 FireViewer training/validation hashes was zero.

Metric	Result
COCO mAP50-95	0.1491
COCO mAP50	0.2784
Smoke AP50-95	0.2790
Flame AP50-95	0.0193
Calibrated threshold	0.10
Holdout precision / recall / F1	0.4274 / 0.2488 / 0.3145
Holdout negative-image false-alarm rate	0.3391

The threshold was selected only on a 129-image stratified calibration subset, then measured once on a 383-image holdout. See benchmark/independent-evaluation.json for the machine-readable protocol and metrics.

Artifacts

• model.pt: native Ultralytics checkpoint;
• model.onnx: portable FP32 ONNX export;
• inference.py: minimal invocation;
• training/, validation/ and provenance/: resolved configuration, validation metrics and source receipts.

Rights and safety

Ultralytics code and models are distributed under AGPL-3.0 unless covered by a separate enterprise licence. Training-data attribution and source-specific terms remain applicable; see RIGHTS_AND_ATTRIBUTION.md.

This model is not a certified warning system. It must not be used alone to confirm a fire, trigger an alert, order an evacuation or direct emergency response.