Computer Vision Development
Counting, checking and reading from images, measured on your own lighting.
What is computer vision?
Computer vision uses cameras and models to count objects, inspect items, read text or identify conditions from images, with accuracy that depends heavily on the lighting, angle and image quality of the specific place it runs.
Computer vision demonstrations are convincing because the images are good. Real deployments live on whatever the existing camera produces, in whatever light the building has at four in the afternoon.
That gap is most of the risk. A model reporting ninety five percent accuracy on a benchmark can perform very differently on your loading bay, and the only way to know is to test on your images.
The applications that work well here are narrow and countable. Counting units, reading plates, checking whether a required item is present, flagging visible damage. Broad visual judgement is a much weaker proposition.
Why we test before we quote
Duara builds computer vision from Nairobi for manufacturers, logistics operators, agricultural businesses and security operations.
We measure accuracy on your own images before quoting a build, and we have told clients the conditions were not good enough. Camera position and lighting frequently determine the outcome more than the model does, and that is a cheaper problem to fix first.
Who this is for
- Counting is done manually and inconsistently
- Visual checks depend on who is looking
- You have cameras already recording nothing useful
- Inspection is a bottleneck on a line
Industries we serve
- Manufacturing and packaging
- Agriculture and produce grading
- Logistics and yard operations
- Security operations
- Construction sites
- Retail loss prevention
Problems this solves
These are repetitive visual tasks, which is where vision works. Judgement heavy visual work is not.
- Manual counting is inconsistent. Two people counting the same load get different figures, and neither can be checked afterwards.
- Checks depend on the inspector. A visual standard is applied differently by different people and differently at the end of a shift.
- Cameras recording nothing useful. Footage exists and is only ever reviewed after an incident, which is the least valuable use of it.
- Inspection becomes the bottleneck. A line runs at the speed of the person checking, and adding people does not scale well.
What we deliver
Vision applied to narrow countable tasks, tested on your own images and conditions, with accuracy measured honestly and camera placement treated as part of the work.
Vision models trained for your specific task
Accuracy measured on your own images
Camera placement and lighting recommendations
Integration with your existing systems
A defined route for uncertain cases
Ongoing accuracy monitoring after deployment
Features
- Tested on your conditions first. Benchmark accuracy means little. We measure on your images before quoting the build.
- Camera and lighting treated as part of it. Position and light frequently matter more than the model, and fixing them is cheaper than compensating for them.
- Uncertain cases routed to a person. Low confidence results go to a human rather than being forced into a decision.
- Narrow tasks preferred. Counting, reading and presence checking work well. Broad visual judgement is a much weaker proposition and we say so.
- Accuracy monitored after launch. Performance is tracked over time, since lighting and camera condition change and accuracy drifts with them.
Benefits
- Counts become consistent. The same load counted twice gives the same answer, and the image is retained as evidence.
- Inspection stops limiting the line. Checking at machine speed removes a bottleneck that adding people does not solve.
- Existing cameras earn their keep. Footage moves from post incident review to producing usable information continuously.
- Standards get applied evenly. A model applies the same threshold at the start and end of a shift.
Technologies we use
- Object detection and counting. Identifying and counting items in an image or video stream.
- Optical character recognition. Reading plates, labels, batch codes and serial numbers from images.
- Condition and defect detection. Flagging visible damage or departure from a defined standard.
- Confidence thresholds. Routing uncertain results to a person rather than forcing a decision.
- Edge and server deployment. Running on site where bandwidth or latency requires it, or centrally where it does not.
Camera placement is the cheapest accuracy improvement available and it is almost always the first thing worth changing. Software cannot recover information the image never captured.
How the process works, step by step
A vision project usually takes six to twelve weeks, and it starts with images from your actual site rather than a specification.
-
Step 1
Free image assessment
We test on images from your site, since that is the only meaningful indicator of whether this will work.
-
Step 2
Conditions addressed
Camera position and lighting are corrected first, because software cannot recover what the image never captured.
-
Step 3
Accuracy measured
Performance is established on your conditions and reported plainly, including when it is not good enough.
-
Step 4
Integrated and thresholded
The system connects to your existing records with uncertain cases routed to a person.
-
Step 5
Monitored after launch
Accuracy is tracked over time, since cameras get dirty and lighting changes.
Quality assurance and how we work
The commitments attached to a technology whose performance depends on conditions.
- We measure on your images before quoting. Benchmark accuracy tells you nothing about your loading bay, and we have told clients the conditions were not good enough.
- Camera and lighting come before modelling. Position and light usually matter more than the model, and fixing them is cheaper than compensating for them.
- Uncertain results go to a person. A low confidence result routed to a human is correct behaviour, not a failure of the system.
- We report the real accuracy figure. Including where it is lower than you hoped, because a vision system trusted beyond its accuracy causes the damage.
- We decline broad visual judgement. Narrow countable tasks work. Asking a model to assess something subjective is a weaker proposition and we say so.
Frequently asked questions about computer vision
Will it work in our conditions?
That is exactly what we test before quoting. Benchmark accuracy tells you nothing about your site at four in the afternoon.
What if our images are not good enough?
We say so. We have told clients the conditions were not workable, because taking the project anyway would waste their money.
Do we need new cameras?
Sometimes, and it is often the cheapest improvement available. Software cannot recover information the image never captured.
Can it use our existing CCTV?
Frequently yes, though position and resolution decide it, and we assess that on your actual footage.
What tasks work best?
Narrow countable ones. Counting units, reading plates and codes, checking presence, flagging visible damage.
What does not work well?
Broad visual judgement. Asking a model to assess something subjective is a much weaker proposition and we would say so.
What happens with a blurred or awkward image?
The result goes to a person. Routing an uncertain image to a human is correct behaviour rather than a failure.
What accuracy will we actually get?
A figure measured on your own images, including when it is lower than you hoped, because trusting a vision system beyond its accuracy is what causes the damage.
Can it read number plates?
Yes, and yard and gate operations are among the more reliable applications, given reasonable camera angles.
Does it work at night?
Depends entirely on lighting, and that is one of the first things we test rather than assume.
How long does a vision project take?
Six to twelve weeks, starting with images from your actual site rather than a specification.
Does it need internet?
Not necessarily. Edge deployment runs on site where bandwidth or latency requires it.
Can it count produce or livestock?
Yes, and agricultural counting is a common request, though accuracy varies considerably with how the items are presented.
Will accuracy stay the same?
No, and that is why we monitor it. Cameras get dirty and lighting changes, and accuracy drifts with them.
Can it identify people?
We scope facial recognition very carefully and often decline it, since it raises data protection questions most clients have not considered.
Does it integrate with our systems?
Yes, writing counts and results into your existing records rather than sitting apart as another screen.
What about privacy of footage?
Images stay in your infrastructure. Where processing happens on site, they need not leave the building at all.
Can it detect safety violations?
Some, such as missing protective equipment, though we would be clear that it supports supervision rather than replacing it.
Who owns the trained model?
You do, along with the labelled images, which are the harder asset to recreate.
What does a computer vision project cost?
Camera count and task complexity determine it, quoted after the image assessment.
Glossary of terms
Computer vision terms in plain language.
- Object detection
- Locating and identifying specific items within an image.
- Confidence threshold
- The certainty level below which a result is sent to a person instead of accepted.
- False positive
- Reporting something that is not there, which is usually less costly than missing something that is.
- Edge deployment
- Running the model on site rather than sending images to a server.
- Training images
- Examples from your own conditions used to teach the model what to look for.
- Drift
- Accuracy declining over time as cameras, lighting or conditions change.
Related services in AI Development
Talk to us about computer vision
Send us fifty images from the camera you would use. That tells us more about feasibility than any specification.
Start your project