What cameras can and can't do
A 2 MP ceiling dome limits what any software can see. Read this before buying AI.
Every AI demo looks perfect because it runs on hand-picked footage. Real sites have dusty domes, backlit entrances, 15 fps recorders and cameras pointed at walls. This post is the camera map we draw with every customer in week 1.
Cameras are good at: who is where and for how long; counting people at gates, stairs and billing queues; vehicle presence at entries; night footage where lighting is decent. A well-placed 4 MP camera at a gate beats any algorithm on a bad angle.
Cameras struggle with: hands and small items seen from a ceiling dome; faces beyond 6–8 metres on 2 MP; reading plates on fast movers without a dedicated ANPR angle; anything through a dusty dome (clean it first — free accuracy).
Our rule: if a camera can't see it, no software can, and we say so before billing. Sometimes the right recommendation is moving a camera, not buying AI.
How we measure a pilot (and why we don't promise first)
Alerts per camera per day, share confirmed real, incidents found by search. Measured on your footage, or it doesn't count.
Accuracy percentages on a vendor slide are measured on somebody else's dataset, in somebody else's lighting, on footage you'll never see. We don't quote them. Here's what a 4-week NetraVision pilot measures instead.
Week 1 — baseline: your current review load. How many hours does footage review take per incident today? How many night alerts does the guard get, and how many are real?
Weeks 2–3 — assisted: alerts per camera per day, share confirmed real on human review, staged or past incidents caught by search. We tune weekly — keep, fix or drop each alert type.
Week 4 — readout: search time saved per incident, review hours avoided, coverage gaps found (cameras that see nothing useful). You decide with numbers from your site, not ours.
Cloud vs edge box: an honest guide
Start in the cloud in a day. Go offline where policy or connectivity demands it.
The cloud default: streams go to our cloud, alerts come back with clips to dashboard, SMS or call. Fastest start — a trial runs in days — and the right choice when connectivity is decent and policy allows it.
The edge box: an inference box on your premises. Existing cameras connect to it; analytics run locally; nothing leaves the site at runtime. Updates arrive as signed offline packages. Choose it when policy forbids video leaving the premises, or when links are too poor for stable streaming.
What doesn't change: the same analytics, the same clip-with-every-alert discipline, the same weekly tuning. The box is a deployment option, not a different product.
How to decide: if you hesitated reading 'video leaves the premises', you want the edge evaluation. Otherwise start cloud — you can move sites to the box later without retraining anyone.
Test it on your cameras →
