TianchenGuan/FireSmockDetect-models

FireSmockDetect — wildfire smoke detectors

Trained checkpoints from the FireSmockDetect project (code, protocols and full reports: https://github.com/TianchenGuan/FireSmockDetect). Task: single-class ("smoke") detection on RGB images from fixed lookout cameras and drones, tuned for FAINT early-stage plumes.

Checkpoints

file	arch	params	train data	role
expB_y11s_deployed.pt	YOLO11s @1280	9.4 M	pyro-sdis train + D-Fire train	production default of the project service
v5bA_y11m_best.pt	YOLO11m @1280	20.1 M	same recipe, retrained (v5b arm A, seed 3)	best fixed-camera model on the frozen test
v6A_y11n/v6A_y11n_best.pt	YOLO11n @1280	2.59 M	pyro-sdis train + D-Fire train (real only)	v6 edge model (+ ONNX / TFLite fp16 / TFLite INT8 3.6 MB)

Headline numbers (frozen pyro-sdis TEST, matched negative-image FP rate, IoU≥0.4)

model	R@5%FP	R@10%FP	R@20%FP
v5bA_y11m_best	0.287	0.388	0.553
expB_y11s_deployed	0.258	0.354	0.501

Full per-source tables, aerial-vs-fixed splits, and negative results (data, distillation and ensemble variants that did NOT beat these) are in the GitHub reports (reports/v5b/FINAL_REPORT_V5B.md).

Usage

from ultralytics import YOLO
model = YOLO("v5bA_y11m_best.pt")
res = model.predict("image.jpg", imgsz=1280, conf=0.10)

Pick the operating confidence from your own tolerated false-positive rate; the project calibrates thresholds per deployment scenario (fixed camera vs aerial) — see the repo's eval tooling.

Limitations

• Faint-smoke recall at low FP is the binding constraint (see table); this is an assistance tool, not a certified alarm system.
• Thresholds do NOT transfer between fixed-camera and aerial imagery.
• No early-warning (time-to-first-alert) claim is made: qualified smoke-onset footage was unavailable (documented in the v5 protocol).
• Training data: pyro-sdis and D-Fire (original authors' labels). FIgLib / HPWREN (CC BY-NC-ND) was never used for training.

License

AGPL-3.0 (derivatives of Ultralytics YOLO11). Dataset licences remain with their owners.

v6 edge release (YOLO11n, real-data-only)

v6A_y11n/ contains the frozen v6 deliverable: PyTorch checkpoint, ONNX, TFLite fp16 and TFLite INT8 (3.6 MB), plus infer.py — a single-command offline image/video inference script (PyTorch / ONNX / TFLite backends) — and the export report.

Frozen pyro-sdis TEST (matched negative-image FP, IoU≥0.4): R@5% 0.243, R@10% 0.344 — the best nano-scale model of the v6 matrix. Deployment validation of the INT8 file itself on the project val set: no loss at the 2%/5% FP operating points, −2.1 pt at 10% (traced to graph conversion, not quantization; fp16 TFLite is bit-identical in accuracy). Suggested confidence thresholds from the INT8 file's own val curve: 0.73 / 0.68 / 0.62 for ~2 / 5 / 10 % negative-image FP. CPU x86 median forward latency 45.8 ms (proxy; on-device latency unmeasured).

v6 synthetic-data result (pre-registered, reported as-is): neither procedural compositing nor repaired generative image-to-image synthesis improved real-data evaluation at matched optimizer steps; seed spread (±3 pt) exceeded all between-arm differences. Full protocol and reports in the GitHub repo (reports/v6/, reports/v6/repair/, reports/v6/sprint/).