Precision measures how trustworthy a system's alerts are: of all the events it flagged, what fraction were real. Formally it is true positives divided by all positives the system raised — true positives over true positives plus false positives. A precision of 0.9 means nine in ten alerts were genuine and one was a false alarm. It is the metric that answers an operator's gut question: "when this thing cries wolf, how often is there actually a wolf?"
Precision is one half of the accuracy picture, and it cannot be read alone — its partner is recall, which measures how many real events were caught. The two trade off: you can make a detector extremely precise by only alerting when it is absolutely sure, but then it stays silent on borderline real events (low recall). Quoting a single "accuracy" number hides this trade; a serious specification gives precision and recall together for the actual operating conditions.
The pitfall is optimising precision in a vacuum. High precision with terrible recall is a detector that almost never false-alarms because it almost never alerts at all — useless for catching events. Which way to lean depends on the cost of the two errors: a watchlist or retail alert wants high precision (don't accuse the innocent), a perimeter at night wants high recall (don't miss the intruder), and you set the confidence threshold to pick the balance. Precision is a range tied to conditions and tuning — never 100% in the real world.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
Precision tells you how many alerts were real, and reading it without recall hides how many events were missed. Our seven-volume handbook covers the full video stack, and for precision in analytics start with Volume 6: detection, tracking and the ML minimum for product decisions. Volume 7 covers vertical playbooks and real-time pipelines. Written by Nikolay Sapunov, CEO at Fora Soft.
Need real precision and recall numbers for your video analytics?
Send us labelled clips or alert logs, and we will measure precision and recall per camera and propose threshold changes. 250+ video products shipped since 2005.
