Recall measures how complete a system's catching is: of all the real events that actually happened, what fraction did it detect. Formally it is true positives divided by true positives plus false negatives — the events caught over all the events that occurred. A recall of 0.8 means the system caught eight of every ten real events and missed two. It answers the question that matters for security: "of the things I needed to see, how many did this system actually flag?"
Recall is the other half of accuracy, inseparable from precision. Pushing recall up — by alerting on anything remotely suspicious — catches more real events but also raises false positives, lowering precision; tightening to avoid false alarms raises precision but misses more real events, lowering recall. This is why a single accuracy figure is misleading: the honest statement is a precision/recall pair measured under the real scene, lighting, and tuning.
The pitfall is chasing recall without counting the cost. Recall of nearly 1.0 is easy if you accept a flood of false alarms — alert on everything and you miss nothing, but you bury operators and trigger alert fatigue, so the true events drown anyway. The right operating point depends on the cost of a miss versus a false alarm: lean to high recall where missing an event is unacceptable (perimeter intrusion, weapon detection), and accept more false positives there, managed by human verification. Recall is a conditions-dependent range, never 100% in practice.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
Recall decides how many real intrusions or incidents an analytics system misses, and teams often tune it blind by moving a confidence slider. Our seven-volume handbook covers the full video stack, and for recall start with Volume 6, which covers detection, tracking and segmentation, the ML minimum needed to judge models, and how an AI feature moves from prototype to production. Written by Nikolay Sapunov, CEO at Fora Soft.
Missing too many real events after tuning your analytics for fewer false alarms?
We will review your detection pipeline and propose a threshold and evaluation plan that balances recall against operator alert load. 250+ video products shipped since 2005.
