Two alerts per camera per 81-minute window, hard budget. Which camera do you look at first?
PlumeRank ranks a wildfire-camera network so the camera most likely showing a new plume reaches the top of a dispatcher's watch list — under a hard budget of two alerts per camera‑night, with the threshold solved on fire-free windows rather than picked by hand. (A camera‑night here is one 81-minute camera window, the length of a FIgLib sequence — not a 12-hour shift.) OpenCV 5 does the per-frame work: background modelling, optical flow and plume geometry, fused into one hazard and scored against each camera's own recent history.
32.0 min median time-to-detection on the development set: 17 of 29 fires detected. Threshold solved to 1.99 alerts per camera-night on the confusion set; 2.37 realised on the development set. In-sample — the constants were fit on these fires (results/arm_dev_2026-10-10.json). The held-out result has not been run yet and is the only number this project will call its headline.
make verifyabout 3 min — the tests plus a small committed sample, re-run. The development numbers and the held-out run each have their own command.results and the build string · the 30-second path for judges
connecting to the live stream…
Replaying the development fires (the constants were fit on them) — one frame per minute of footage, replayed faster than real time on screen. The headline will be measured on the held-out sequences the system has never seen.
waiting for the first replayed frame.