We quote 97.8% counting accuracy. That sounds like a marketing number. It is actually a confession: 97.8% means 22 missed counts in every 1,000. This post is about those 22.
01 · WHERE THE MISSES COME FROMWhere the misses come from
Counting accuracy is not uniform across a room. The errors cluster, and knowing where they cluster is more useful to a customer than the headline percentage.
- Doorway tailgating. Two people crossing a threshold shoulder-to-shoulder in the same frame are the single largest source of miscounts. The model resolves most of them; a fraction read as one person.
- Extreme density. At very high occupancy, individual separation degrades. Counts stay directionally correct but absolute accuracy softens.
- Unusual geometry. Swing doors, mirrored surfaces and atrium voids each introduce their own characteristic error, which is why every deployment is calibrated per site.
02 · WHY WE PUBLISH THE FAILURE MODESWhy we publish the failure modes
A vendor who only quotes the headline number is implicitly promising uniform accuracy, and that promise breaks the first time a facilities manager checks a busy doorway by eye. We would rather the customer know, on day one, that the doorway is the hard case and that we calibrate for it.
Calibration is a 30-minute manual ground-truth pass per sensor: a human counts, the model counts, the difference tunes the deployment. We re-run it whenever a sensor's drift exceeds 2%.
03 · HUMILITY AS A FEATUREHumility as a feature
Occupancy data drives decisions HVAC schedules, safety thresholds, capacity planning. A decision made on a number that is quietly wrong is worse than a decision made on a number known to be approximate. So we ship confidence intervals alongside counts, and we support independent ground-truth audits on every pilot.
97.8% is not the end of the sentence. The end of the sentence is: and here is exactly which 2.2% to be careful with.