Computer Vision
Computer Vision in Production: Turning Cameras Into Business Decisions
A vision model that scores 95% in a demo often drops to 70-80% in a real facility. Here is what separates deployed computer vision from pilot demos.
Quarza Team · · 7 min read
Most businesses already have cameras. Warehouses have them. Factories have them. Stores have them. Buses have them. Almost none of that footage produces a decision.
That gap is the entire opportunity in applied computer vision, and it is also where most vision projects quietly die. The model is rarely the hard part. Getting a model to keep working in a real facility, and connecting what it sees to something a person or system actually does, is the hard part.
Key Takeaways
- Almost every applied vision problem reduces to the same pattern: image to signal to decision to action.
- Accuracy measured in a demo is not accuracy in a facility. Reported drops from around 95% in controlled conditions to 70-80% in live environments are common.
- Domain adaptation on your own footage is a required step, not an optimisation.
- Data quality, not model architecture, is the most frequently cited cause of failure.
- The exception path deserves as much design attention as the happy path.
The Pattern Underneath Every Vision Project
Strip away the domain and the same structure appears every time:
- Capture
- Preprocess
- Detect or classify
- Validate
- Confidence
- Decision
- Action or escalation
A camera produces an image. Preprocessing corrects for the conditions the image was captured in. A model produces a detection, classification or reading. Validation checks that result against context and rules. Confidence determines whether the system can act alone. Then something happens: a record is written, an alert fires, a line stops, or a person is asked to look.
Everything domain-specific lives in the middle. Everything that determines whether the project succeeds lives at the ends.
Where This Applies
| Domain | What the camera sees | What the business gets |
|---|---|---|
| Logistics | Labels, barcodes, packages, damage | Verified shipment data, exception flags |
| Manufacturing | Surface defects, missing components, assembly errors | Inline quality control, reduced escape rate |
| Retail | Shelf state, product facings, price labels | Out-of-stock alerts, planogram compliance |
| Warehousing | Inventory positions, exceptions, worker safety | Fewer exceptions, better throughput |
| Transportation | Occupancy, driver state, incidents | Safety alerts, verifiable trip records |
| Documents | Printed and handwritten records | Structured data into existing systems |
The list looks diverse. Architecturally it is not. A team that has built a robust courier label pipeline has already solved most of the general problems in a shelf monitoring pipeline: variable lighting, occlusion, motion blur, class imbalance, and the need for a confidence-gated exception path.
That is what makes computer vision a horizontal capability rather than a single product.
Why Demo Accuracy Does Not Survive Deployment
This is the single most important thing to understand before commissioning a vision project.
A model evaluated on curated images under consistent conditions will report a number that has very little predictive power for how it behaves in a live facility. Practitioner reports from retail deployments describe models hitting around 95% in controlled evaluation and dropping to 70-80% in real stores with inconsistent lighting, crowded shelves and peak-hour motion blur. Manufacturing reports the same pattern under shift changes, lighting variance, tooling wear and new SKUs.
The causes are consistent across domains:
- Lighting variance. Daylight through a shutter, overhead sodium, mixed colour temperatures, glare on plastic and film.
- Occlusion and clutter. Objects partly hidden by other objects, by hands, by packaging.
- Motion blur. Anything moving, which in a warehouse is everything.
- Capture variation. Different camera positions, focal lengths, resolutions and mounting angles between sites.
- Distribution shift over time. New products, new packaging, new label formats, seasonal changes.
- Rare classes. The defect that matters most is often the one with the fewest training examples.
None of these are model problems. They are data problems, and they are why the most consistently cited reason for vision project failure is data quality rather than algorithm choice.
Domain Adaptation Is Not Optional
The practical consequence is that a model trained on generic or open-source data will underperform the moment it meets your environment. Fine-tuning on footage from your actual site, under your actual conditions, including your bad conditions, is a required project phase.
That has scheduling implications people frequently miss. Before a model can be adapted, someone has to collect representative footage across shifts, seasons and conditions, and someone has to label it. Labelling is slow, expensive and easy to get wrong, particularly for fine-grained classification where even human annotators disagree.
Two practices help materially:
Collect before you build. Data collection can start well before any modelling work, and its duration usually determines the project timeline more than the modelling does.
Deliberately collect failure conditions. The dataset should over-represent the difficult cases, because those are the ones that determine whether the system is usable. A dataset made only of clean images produces a model that works only on clean images, which in practice means it works in the demo and nowhere else.
The Exception Path Is Half the System
A vision system that only handles the cases it is confident about is not finished. What it does with the rest determines whether anyone trusts it.
Three lanes, as with document processing:
- Automatic. High confidence, validation checks passed, no exception rules triggered. The system acts.
- Assisted review. Lower confidence or a triggered rule. A person sees the image, the model's reading, and the specific reason it was flagged, and confirms or corrects in seconds.
- Full manual. Unreadable input, an unrecognised class, or a hard business rule that always requires human sign-off regardless of model confidence.
The third lane is important and often omitted. Some decisions should never be automated no matter how confident the model is, and encoding that as a deterministic rule rather than a threshold is the safer design.
The review interface is also where model improvement happens. Corrections captured in lane two are exactly the training data that lane one needs to expand. A system without a review path has no improvement path either.
Edge or Cloud
The deployment decision is usually forced by three constraints rather than chosen freely:
- Latency. If a decision has to happen before an object leaves the frame or a line moves, inference has to be local.
- Bandwidth. Continuous video from multiple cameras to the cloud is often impractical over available connectivity, particularly at industrial or vehicle sites.
- Data sensitivity. Footage containing people, children or patients frequently cannot leave the premises, which settles the question regardless of the other two.
A common workable pattern is edge inference for immediate decisions, with selected frames, metadata and low-confidence cases sent centrally for review, monitoring and retraining.
Monitoring After Go-Live
Vision systems degrade quietly. A camera drifts on its mount. A supplier changes packaging. A light fitting is replaced. Nothing errors, and accuracy falls without anyone noticing until someone downstream complains.
Production monitoring should track, at minimum:
- Confidence score distribution over time, where a shift indicates drift before accuracy reports do
- Escalation rate against its historical baseline
- Per-class performance, since aggregate accuracy hides collapse in a single important class
- Regular audited samples of automatically approved decisions, which is the only reliable way to catch a threshold set too loosely
- Camera health, including position, focus and exposure
The audited sample deserves emphasis. It is the check that finds errors before a customer, a regulator or an incident does.
How to Pilot Without Losing Six Months
The failure pattern in vision pilots is broad scope and vague success criteria. The corrective is uncomfortable specificity.
- Pick one decision. Not "improve quality." One decision: does this unit pass, is this shelf slot empty, is this the right courier.
- Define the metric before building. What accuracy, at what escalation rate, on what sample, judged by whom.
- Collect a calibration set from the real environment, weighted toward difficult conditions.
- Run in shadow mode. The system produces outputs, nothing acts on them, and outputs are compared against current practice.
- Automate the easy slice only. High-confidence, low-risk cases first, with everything else still reviewed.
- Widen from measured evidence, not from optimism.
This sequence is slower to start and much faster to reach production, which is the trade most pilots get backwards.
Frequently Asked Questions
How accurate is computer vision in real deployments?
It depends far more on conditions than on the model. Controlled evaluation figures around 95% commonly fall to 70-80% in live environments with variable lighting, clutter and motion, until the model is adapted on footage from that specific environment. Any accuracy figure quoted before someone has seen your site should be treated as a marketing number.
Do we need to label our own data?
In almost all cases, yes. Generic training data does not capture your lighting, camera angles, packaging or defect types. Labelled footage from your own environment is what closes the gap between demo performance and production performance, and collecting it usually drives the project timeline.
Can we use our existing cameras?
Often, partly. Existing CCTV is frequently positioned for human review rather than machine analysis, so angle, resolution and lighting may not suit the task. It is worth evaluating existing coverage first, since some use cases work with what is installed and others need purpose-placed cameras.
Should inference run on the edge or in the cloud?
Latency, bandwidth and data sensitivity usually decide it. Real-time decisions and sites with limited connectivity point to edge inference. Footage containing people, children or patients often cannot leave the premises at all, which settles the matter independently of performance considerations.
How do we stop the system from degrading over time?
By monitoring for it. Track confidence distributions, escalation rates and per-class performance, audit a random sample of automatically approved decisions on a schedule, and check camera health. Drift is usually visible in these signals well before it shows up as a complaint.
What should a first vision project be?
The narrowest valuable decision you can define, with a success metric agreed before any modelling starts, run in shadow mode against current practice. Broad scope is the most reliable way to spend six months and deploy nothing.
Have Cameras but No Decisions?
If you are already capturing images or video that nobody has time to look at, there is usually a specific decision buried in that footage worth automating first. Get in touch and we can help identify which one.
Keep reading
- Approach & Methodology
Our AI Consultancy Approach: Problem First, Architecture Second
8 min read - Logistics & Document AI
Shipping Label Intelligence: Turning Messy Logistics Documents Into Verified Data
9 min read - Healthcare & Document AI
AI Document Automation for Hospitals: Turning Paper Into Structured Records
8 min read
