Score of the model that held this fire out of training. Dashed line: operating threshold chosen on validation fires. The clip opened by default is a clear example, not a typical one — use the filters to browse misses and false alarms. Orange line: FIgLib's annotated first visible plume (offset 0; not the ignition time). A strict detection is a persistent alarm whose run of above-threshold frames starts after that line; an alarm already running at the plume onset is not counted as a detection. Persistent alarm (alarm-v2) = 2 consecutive frames (≤ 180 s apart) above threshold, counted only after 20 min of clip history; a single frame above threshold is only a candidate frame. The detector compares each frame with the same camera's previous 5–20 minutes.
Loading evidence…
Check your own camera frames
Upload the frame to check and one or more frames from the same fixed camera taken roughly 5–20 minutes earlier. PlumeCheck compares them and shows where the model responded (a model attribution, not a verified smoke location). Uses the model trained only on fires before August 2021. Available only with the local research server; the upload path's preprocessing has not been checked against the training pipeline.