Your cameras were sited for human reviewers, and humans are extraordinarily tolerant — we read backlit silhouettes, recognise people at odd angles, and mentally fill in what a pallet is hiding. A detection model has none of that. So a retrofit inherits mounting decisions made for somebody else's purpose, and the cameras are not wrong, they are optimised for the wrong consumer. The number that settles most arguments is pixels on target, which is not the same thing as resolution or coverage.
Vision projects get budgeted as software projects. Models, servers, integration, a licence. Then the system goes live and underperforms, and the conversation turns to better models — when the binding constraint was decided years earlier by an installer on a ladder who was solving a different problem.
That installer was not careless. A camera mounted for human review has to let a person work out what happened afterwards, and people are very good at that under bad conditions. Mount it high so nobody can reach it, angle it steeply to cover the most floor, accept the backlight from the doorway — all reasonable, and all actively hostile to a detector.
The same scene, two very different consumers
Pixels on target
This is the concept that converts a vague argument into a number, and it is worth insisting on early.
Pixels on target is how many pixels land on the thing you want detected — not what the sensor produces in total. A high-resolution camera with a wide field of view, covering a large area from a distance, can put very few pixels on a person at the far end of that zone. Below some threshold, detection stops working regardless of the model.
It reconciles two statements that sound contradictory and are both true: the camera covers that zone, and the system cannot see anything in it. Coverage and pixels on target are different properties, and only the second predicts whether detection will work. Once a team has the number for the far end of each zone, the conversation stops being about model quality and starts being about lenses.
Height and angle
Most training data is captured at or near eye level. A camera mounted high with a steep downward view sees something quite different: a person from directly above is largely shoulders and a hard hat, not a recognisable human form.
Security installers favour high mounting because it resists tampering and covers more floor. Both are good reasons for their purpose. For yours, lowering a camera or reducing its angle is frequently the single cheapest accuracy improvement on the site — and reliably the last thing anyone proposes, because it is not a software change and nobody owns it.
What happens after dark
Infrared is not dim daylight. It removes colour entirely, changes apparent texture, and produces hotspots where the illuminator is strongest with darkness beyond its range. A model trained and validated on daytime colour footage is being asked a materially different question after the switchover.
Performance often falls off a cliff at that moment rather than degrading gradually, which is why acceptance testing that runs during the working day tells you very little about the night shift. If the site operates after dark, the transition and the hours past it belong in the acceptance test, and night may warrant its own treatment rather than being assumed to follow from the daytime numbers.
Check the stream, not the spec sheet
The cheapest thing on this list, and the one most often skipped. The resolution, frame rate and compression arriving at the analytics server are frequently lower than the camera's specification implies, because somebody configured the recorder years ago for storage economy and nobody has looked since.
Heavy compression in particular destroys exactly the fine detail a detector depends on, while looking perfectly acceptable to a person reviewing footage. It costs nothing to check and it is invisible until you do. We went through the related constraint — how many of those streams a GPU can actually process — in how many cameras can one GPU handle.
The ordering that saves money
Physical first, then the rule layer, then the model.
Moving a camera, changing an angle, fitting a different lens, adding a light or cleaning a housing can each be done in an afternoon by someone already on site, with an immediate and measurable effect. The rule layer — zones, dwell thresholds, deduplication — is a configuration change, which we covered in PPE detection on a factory floor. Retraining is a full cycle of collection, labelling, training and validation, and it cannot recover information the camera never captured.
That last clause is the one worth repeating to anyone proposing a better model as the first move. If the subject was a silhouette, the pixels containing the answer do not exist. No model recovers them.
What a survey covers
Per camera: mounting height and angle; what the light does across a full day including the switch to artificial or infrared; what sits between camera and subject at the busiest time rather than the quietest; the lens and resulting field of view; and pixels on target at the far end of the zone. Then the stream as it actually arrives.
It is half a day on most sites and it reliably changes the project plan. On a retrofit it is the difference between a system that works and a procurement conversation about better models — the same point we made about working with cameras you already own in adding AI analytics to CCTV you already have.
Questions we get asked
Why do existing CCTV cameras underperform for analytics?
They were sited for a different consumer. Human reviewers tolerate backlight, odd angles and occlusion without noticing they compensated. A model has none of that tolerance. The cameras aren't wrong — they're optimised for somebody else's purpose, decided years before anyone mentioned analytics.
What is pixels on target?
How many pixels land on the thing you want detected, not what the sensor produces. A high-resolution wide-angle camera can put very few pixels on a person at the far end of a zone, and below a threshold detection simply stops. It's why "the camera covers that zone" and "the system can't see anything there" are both true.
How much does mounting angle matter?
A lot — training data is mostly near eye level, and a person seen from directly above is shoulders and a hard hat. Installers mount high to resist tampering and cover floor, which is right for their purpose. Lowering a camera is often the cheapest accuracy gain available and the last thing proposed.
What happens at night?
Infrared isn't dim daylight — no colour, altered texture, hotspots near the illuminator and darkness past it. Performance tends to fall off a cliff at the switchover rather than degrade. Acceptance testing during the working day tells you nothing about the night shift.
Move cameras or improve the model?
Move cameras, usually by a wide margin. An angle change or a different lens takes an afternoon and the effect is immediate. Retraining is a full cycle and it cannot recover information the camera never captured — if the subject was a silhouette, those pixels don't exist.
What should a site survey check?
Per camera: height, angle, light across a full day including the IR switch, obstructions at the busiest time, lens and field of view, and pixels on target at the far end. Then the stream actually arriving at the server — often lower resolution and more compressed than the spec sheet implies, and free to fix.
We start vision projects with the cameras rather than the models, because four of the five variables that decide accuracy are physical. If a system is underperforming and the proposal on the table is a better model, a survey is the cheaper thing to try first.