Outdoor fire detection fails when models are trained only on controlled indoor flames. Ember targets the scenes operators actually worry about: forecourts, loading bays, storage yards, and perimeter lines.
Performance snapshot
- 90% accuracy on outdoor fire and smoke validation splits under Innomium protocols
- 19MB ONNX for edge deployment
- Browser demo for rapid evaluation by safety teams
These metrics are protocol-specific. Sun glare, welding, dust, rain, and vehicle lights can all resemble a target event. A useful evaluation set includes those hard negatives alongside positive classes.
Designing an alerting workflow
A detection is not an incident. Safety teams need a path from signal to review: what triggers an alert, who receives it, which camera context is included, and when the event is closed. We recommend evaluating Ember with the actual operating protocol, including expected false alarms, escalation requirements, and the conditions in which human confirmation remains mandatory.
Outdoor scenes deserve particular care. A clearer threshold policy comes from calibrating against hard negatives from the real site, not only clean flame examples.
Challenge mechanism
Arena challenges can expand labeled outdoor data and pressure-test recall on difficult negatives. Challenge submissions are builder contributions under challenge rules — not client endorsements. Confirm current rounds and licenses on Arena and the model card before planning commercial use.
Related: [edge vision evaluation protocol](/updates/edge-vision-evaluation-protocol).