Almost every business already owns the raw material for computer vision: security cameras, phone photos, scanned documents, drone footage. Almost none of it is doing any work. Computer vision services exist to change that, by turning images and video into counts, measurements, alerts and structured records that software can act on.
This guide is for the person who has to buy that work. It covers what computer vision services include, how a project runs, what moves the price, and the questions that tell you whether a provider has shipped a system before.
The short definition
Computer vision services are the engineering work of getting software to read images and video for a specific job in a specific business. A provider defines the task, gathers and labels data, trains or adapts a model, deploys it where it needs to run, connects it to your systems, and keeps it accurate as conditions change.
The model is the smallest part. Most of the effort, and most of the difference between providers, sits in the data before it and the deployment after it.
The six kinds of work on offer
Different providers use different labels, but nearly every project is a combination of these.
- Object detection and tracking. Finding and following things in images and live video: people, vehicles, products, packages. The output is boxes, counts and events.
- Image segmentation. Deciding pixel by pixel what belongs to what. Used when you need shape and area, as in damage assessment, medical and scientific imaging, or mapping from aerial photos.
- Image classification. One label for one image: pass or fail, this product or that one.
- OCR and document extraction. Reading invoices, IDs, contracts, receipts and forms into structured fields.
- Visual inspection. Defect and quality checks on a production line, at line speed, with a record of every decision.
- Video analytics. Queue length, footfall, dwell time, shelf compliance, protective-equipment checks and restricted-zone monitoring, usually on cameras that are already installed.
If you want to know which model family sits behind each of these, our field guide to computer vision models walks through classifiers, detectors, segmenters and foundation models in plain language.
How a computer vision project runs
1. Define the task
The first deliverable is a sentence, not a model: what counts as a positive, what counts as a negative, which edge cases matter, and how many false alarms the operation can tolerate. A vague task produces a model nobody trusts.
2. Collect and label data
Footage and documents from your real environment, in your lighting, with your products. Part of it is set aside as a test set that the model never sees during training, so the accuracy figure you are shown means something.
3. Train and benchmark
Several candidate models are compared on your data. Sometimes a foundation model works with almost no training. Sometimes a compact custom network is the only thing fast enough for the hardware. The choice should come from the benchmark, not from the provider's favourite tool.
4. Deploy
On an edge device next to the camera when the site needs low latency, has poor connectivity, or cannot send video off the premises. In the cloud when scale and central reporting matter more. Either way, the output has to land somewhere useful: an alert, a dashboard, a row in the system your team already works in.
5. Maintain
Lighting changes, packaging changes, a camera gets moved. Accuracy decays quietly unless someone monitors it and retrains. A provider that has no answer for this step is selling a demo.
Ask to see the test set. A provider who cannot show you the examples their accuracy number was measured on does not have an accuracy number.
What drives the cost
Quotes for computer vision work vary widely, and the model architecture is rarely the reason. Five things move the number.
- Data collection and labelling. How much footage has to be gathered, and how much expert time each label takes. Outlining a region pixel by pixel costs far more than drawing a box.
- How rare the target is. A defect that appears once in ten thousand items needs far more footage than a product that appears in every frame.
- The accuracy and speed required. The last few points of accuracy, and real-time speed on small hardware, are where the engineering hours go.
- Deployment footprint. One cloud endpoint is cheap to run. Forty edge devices across several sites need provisioning, updates and monitoring.
- Integrations. Connecting to camera systems, line controllers, the ERP or the CRM is ordinary software work, and there is usually more of it than expected.
The safest way to control cost is to start narrow: one camera, one task, one measurable outcome. Our broader guide to what AI automation costs covers engagement shapes and running costs.
Custom model or off-the-shelf API?
Cloud vision APIs are a good answer when the task is generic and the volume is modest: reading printed text, detecting common objects, moderating images. They are a poor answer when the things you care about are specific to your operation, when conditions on your site differ from the photos the API was trained on, or when video cannot leave the building.
A sensible provider will try the cheap option first and tell you when it is enough. Be cautious of anyone who proposes training from scratch before testing what already exists.
Eight questions to ask a computer vision company
- What exactly will the model detect, and what will it ignore?
- How is accuracy measured, on which examples, and can we see them?
- What false-alarm rate should we expect, and what happens to a false alarm?
- Have you tried an existing model or API on our images yet?
- Where will it run, and what happens when the network drops?
- Who owns the data, the labels and the trained model?
- How will we know when accuracy starts to slip, and who retrains it?
- Who is on the keyboard, and are they the people on this call?
Common questions
What do computer vision services include?
Task definition, data collection and labelling, model training or adaptation, evaluation on a held-out test set, deployment to edge devices or cloud, integration with the systems that act on the result, and ongoing monitoring and retraining.
What drives the cost of a computer vision project?
Mostly the data and the deployment, not the model. The main drivers are how much footage has to be collected and labelled, how rare the thing you are detecting is, the accuracy and speed required, the number of cameras or devices, and the integrations around the model.
Should we build a custom model or use an off-the-shelf API?
Use an off-the-shelf API when the task is generic, such as reading printed text or detecting common objects, and volumes are modest. Build or fine-tune a custom model when the objects are specific to your operation, when accuracy on your own conditions matters, or when the system has to run on-site.
How long does a computer vision project take?
A prototype on a handful of your own images can be ready in about a week. A production deployment typically takes four to eight weeks, depending on data collection, model complexity and the number of devices.
The bottom line
Good computer vision services look less like research and more like plumbing: a clearly defined task, honest measurement, and a system that keeps working after the launch. That is how we approach it at BaigOps. See what we build on our computer vision services page, or book a strategy call and we will test the idea on your own images before any invoice exists.