L1 — Real-time detection

Full-frame scan · Weak-signal sensitivity

Scans the entire frame continuously and reacts to the faintest plume-like signals — a few pixels on the horizon line. Tuned for sensitivity; a few milliseconds per frame.

Tracking & signal descriptors

Every candidate, followed for its whole lifetime

Each candidate is tracked from first sighting to expiry. Its spatio-temporal dynamics and visual signature are characterised continuously — flickering artifacts are discarded at no cost, stable signals frame what the rest of the chain has to judge. No per-site training or configuration.

L3 — Semantic verification, on demand

On-device disambiguation · Off the real-time path

Only candidates that require disambiguation are examined, judged on their temporal mosaic — real fire smoke vs low cloud, valley haze, chimney, glare, dust. Runs fully on-device, outside the real-time critical path. Its verdict is crossed with the current descriptor state to reach the decision.

Decision, registry & zone memory

One gatekeeper for alerts · A memory of the scene

A single decision stage emits, holds or closes each event, with anti-burst policy and visual proof attached. After several concordant rejections at the same location, a temporary zone memory activates and stops spending the most expensive stage there. Nothing is suppressed for good: zones auto-expire, and any significant evolution of the signal — growth, intensity, plume-like signature — triggers a fresh verification. The operator can also qualify rejections (glare, road, low cloud, dust) from the web UI, enriching the site memory with no retraining and no integrator.

What frame-by-frame doesn’t tell you

An alert system is judged per fire detected, per delay, and per false-alarm load imposed on the crews. NumidAI natively operates in a temporal regime: it exploits the persistence of signals, so isolated, unstable detections never qualify as alerts. FIgLib, sampled at one frame per minute, sits far below that regime — so we adopt a measurement convention to bring the bench closer to it: an alert is declared after two consecutive detections, one minute of persistence, the benchmark’s own step. On a real stream this persistence is observed in seconds; the convention is an approximation, more constraining than real operation. Measured per event on the same split, strict OOD:

FIgLib is a true-positive-oriented bench, centred on known ignitions: it under-represents the diversity and frequency of field false positives (morning haze, low cloud, grazing light, dust, night lighting). That’s precisely where the full verification chain earns its keep — and what a field pilot has to measure.

Working prototype on target hardware, exercised on continuous or replayed public streams and benchmarked on FIgLib. Next critical step: a long-duration validation on operational streams, in the field — the production false-alarm rate and early recall (fires caught in their first minutes) are the metrics that matter to end users.

An alert system is chosen on its failure cases as much as on its averages. Known limits, stated upfront: night performance is not yet sufficiently characterised (thermal / NIR is the identified path); continuously panning cameras are not handled (step-rotating ones are, via scene recognition); the full L3 chain in continuous-time regime is still under validation.

Looking for a field pilot: one monitored massif, one season, real streams. Benchmark protocol and detailed result tables available on request.