Computer vision outside the lab: light, angles and the last five percent
A detector that scores well on a test set can still fail at the gate at 6 pm. Glare, rain, night, dirty lenses and odd angles are where camera projects are won or lost.
Computer vision demos are easy to like. A box appears around every person, car or product, labelled and counted. The hard part starts when the same model has to work on a real camera, mounted where the building allows, through every hour of the day and every season of the year.
Most of the effort in a camera project goes into the conditions the demo never showed. Planning for them early is cheaper than discovering them after installation.
Where models fail in the field
- Light: low sun straight into the lens, headlights at night, a shop window reflecting the street.
- Weather: rain on the lens, dust in summer, fog in winter mornings.
- Angle: a camera mounted high looks down on objects the model only saw from the front.
- Motion: fast vehicles or people blur at low shutter speeds, especially at night.
- Change: new uniforms, new packaging, a repainted wall, seasonal decorations.
- Hardware: compressed video streams, dropped frames, a camera that reboots at midnight.
Collect data from the real camera, early
The single most useful step is recording a week or two of footage from the actual camera position before training anything serious. Public datasets and stock images help a model learn what a car or a box looks like. Only your own footage teaches it what your gate looks like at dusk.
Picking the hard frames
From that footage we pick frames on purpose: the brightest, the darkest, the busiest, the rainiest. A few hundred carefully chosen frames of hard cases often help more than thousands of easy ones.
Hard frames are picked on purpose. The model learns most from the cases it gets wrong.Fix it at the camera before the model
Some problems are cheaper to solve with hardware than with training data. Before adding more images, we look at the camera itself.
Field problems and the cheapest fix we usually try first.| Problem | Try first | Then |
|---|
| Glare from low sun | Move or tilt the camera, add a hood | Train on glare frames |
| Blur at night | Faster shutter, infrared lighting | Train on night frames |
| Objects too small | Zoom lens or a second camera closer | Higher input resolution |
| Rain and dust on lens | Hood and a cleaning schedule | Flag low-confidence frames for review |
Design for the frames it will get wrong
No vision system is right on every frame. A good one knows when it is unsure and says so. Every prediction carries a confidence score; below a threshold, the frame goes to a person instead of straight into the record.
- Combine several frames: a number plate read the same way on five frames is more trustworthy than one read once.
- Keep a review queue for low-confidence results, and use what people correct as new training data.
- Track accuracy per hour and per camera, so a dirty lens shows up as a drop on one chart, not as a mystery.
A vision system that says "not sure" on one frame in fifty is more useful than one that is confidently wrong on the same frame.
Frequently asked questions
Can you use our existing CCTV cameras?
Often, yes. We first check resolution, frame rate and placement on recorded footage. Some tasks work on ordinary CCTV; small details such as text at a distance may need a better camera.
Does it need a GPU on site?
Not always. Small models run on a modest computer or an edge device for a few cameras. Many cameras, or heavier models, need a GPU on site or a server the video can reach.
How accurate will it be?
We only quote a number after testing on footage from your cameras, across day and night. Accuracy on someone else's dataset says little about your gate.
What happens when we add a new camera?
We record from it for a few days and check accuracy on that footage before switching it on. A new angle or lighting can need a small round of extra training.