How well does it work?
Measured by the evaluator (make eval): synthetic swarms injected into real topic windows for detection, empty windows for false alarms, and hand-labelled live accounts for precision.
ROC: detection vs false alarms
Each point is one detector threshold (CFAR k). The marked point is the operating threshold.
Table
Detection by swarm size
Share of injected swarms found, by number of accounts, at the operating threshold.
Table
Known mistakes
Real false positives and misses, so the numbers above are not taken on faith.