How to Choose a Computer Vision Company: Criteria and 14 Questions

What should you look for when choosing a computer vision company? Accuracy measured on site, written pilot criteria, data security, integration, support and 14 questions to ask in the quote.

Translated from the Turkish original · Türkçe aslı

When choosing a computer vision company, first clarify what kind of need you have: an off-the-shelf product or project-specific development; workplace safety or on-line quality control. Then compare candidates on accuracy measured on site, a pilot with written success criteria, compatibility with existing cameras, where the data is processed, integration, maintenance, local support and pricing transparency. The final decision should come not from a demo video but from the result of a pilot with written criteria.

This guide sets out the questions on the desk of anyone looking for a supplier: how to define what you are looking for, how to test claims, what to ask for in writing in the quote, and which signs should make you stop. There are no company names or rankings; the criteria can be applied in the same way to every supplier, including us.

What should be clear before looking for a computer vision company?

Computer vision companies may go by the same name, but they do different things. Some set up cameras, lenses, lighting and PLC integration for on-line quality control; some run workplace safety, site security or production analytics on top of existing security cameras; some sell only software or only hardware. Knowing which category of supplier you are looking for makes the comparison meaningful.

Three questions are enough to define the need. The first is the type of problem: OHS events such as working without a hard hat or forklift proximity, or quality defects such as scratches, dimensional errors and missing parts? The quality side usually requires new industrial cameras and lighting, while the OHS side usually works with existing cameras; we explained the difference in our guide to quality control with computer vision. The second is whether the solution is off-the-shelf or custom: common scenarios are set up with ready-made models, while a defect or movement specific to your site requires model training and a longer timeline. The third is the definition of success: which event should be caught at what rate, and how many false alarms per shift are acceptable? Quotes requested without this sentence written down cannot be compared with each other.

How do you question an accuracy claim?

Every supplier quotes a high accuracy rate; what matters is where and how that number was measured. A rate measured in good light, on a selected camera and on a ready-made dataset does not represent your site. Four things should be asked: at which site and under what conditions was the measurement made, were the test images separate from those used in training, who labelled the real events, and how were missed events counted?

A single accuracy percentage is not enough either. Precision shows how many of the alerts generated are real; recall shows how many of the real events are caught. One can be high while the other is low, and their costs on site differ: a missed fall and an unnecessary alert are not the same thing. We explained how to read the two concepts in our article on confidence score, precision and recall.

A practical way to test this is to ask the supplier to work on a few hours of footage taken from your own camera. Footage that includes day and night, busy and quiet hours puts on the table in the first meeting the difficulties a demo video does not show. The privacy and KVKK (Türkiye's Personal Data Protection Law) framework for sharing the footage is also discussed at this stage.

How should a pilot be structured?

A serious supplier proposes a pilot before a large installation and itself asks for the success criteria to be written down before the pilot begins. The scope is kept narrow, with a single problem and a limited number of cameras; calibration and measurement periods are separated, and thresholds are not changed during measurement. Alongside the successful clips, the pilot report also includes missed events and examples of false alarms. A negative result is also a result; a supplier that accepts this possibility from the outset inspires confidence. You can find the step-by-step structure in our article on pilot design.

The pilot should also have owners on your side. The operator who will mark alerts as correct or incorrect, the business unit lead who will approve the success criteria, and the IT representative who will provide network and server access are identified from the start. A supplier asking about these roles unprompted shows that it sees the pilot as a measurement rather than a sales demonstration.

How are camera compatibility, data security and KVKK evaluated?

Whether your existing cameras can be used is one of the questions that most affects cost. A system that can read ONVIF and RTSP streams works independently of brand. Ask the supplier to write down, camera by camera across your camera inventory, which locations can be used as they are and which require an angle change or a new camera. In on-line quality control, by contrast, a quote saying the job will be done with security cameras should be questioned.

Where the data is processed is the second critical question. Is the footage processed on a server inside the facility (on-prem) or does it go to the cloud; if it does, is it raw video or only event records? Does the analysis continue if the internet connection is lost? From a KVKK perspective, you should ask whether face recognition or biometric processing is performed, whether recording is event-based or continuous, and what the supplier provides for access permissions and the privacy notice. We covered the legal framework in our article on workplace camera recording and KVKK.

Who is responsible for integration, maintenance and model updates?

An alert that stays only on the screen creates no value. How events will be transferred to MES, ERP, VMS, PLC or OHS software (REST API, webhook, MQTT, digital output) and whether this work is included in the quote should be in writing. In quality control, the decision must reach the PLC in a time that suits the line speed, and the reject mechanism and product tracking must be part of the design.

Once a model is deployed, the site changes: new products, new layouts, different lighting. You should ask how false detections are collected and fed back into retraining, who performs updates, how often and against which test set, and whether it is possible to roll back to a previous version. Response time for hardware failures and spare parts should also be included in the contract.

Information security is also part of this topic. On which network segment will the cameras and analysis server sit, by what method and with whose approval will the supplier's remote access be opened, and how will software updates be distributed? Getting answers to these questions at the quote stage also shortens your IT department's approval process.

Why do local support, references and ownership matter?

Computer vision is a system that lives on site; cameras shift, lenses get dirty, lines change. The distance between the supplier's team and your facility is felt both during the site survey and pilot weeks and at the moment of a failure. The distinction between problems that can be solved remotely and those that require on-site intervention, and the time to arrive on site, should be written into the contract.

A reference visit means seeing the case from the presentation on site: hearing from the user of a system running in a similar facility. Even if the customer's name is kept confidential, the supplier should be able to show a similar installation. Ownership is also discussed from the start: who owns the images collected from your site, the labels and the model trained with this data, and how will the data and records be handed over to you when the contract ends? The same transparency is expected in the pricing model; we explained how the cost items are built up in our article on AI camera system costs.

Which 14 questions should be asked at the quote stage?

The questions below help reduce the quotes to a single comparison table. Asking for the answer to each in writing prevents most disputes that might arise later.

1. At which site and under what conditions did you set up this scenario, and how did you measure accuracy? Ask for the measurement method along with the rate.
2. Can you report precision and recall separately, and what is the expected number of false alarms per shift?
3. What are the pilot's scope, duration, success criteria and measurement method; will these be put in writing before the pilot begins?
4. What happens if the pilot does not meet the criteria? How are the fee, the installed hardware and the next step handled?
5. Which of our existing cameras can be used as they are, and which require an angle change or a new camera? Is the assessment written up camera by camera?
6. Where is the footage processed? What data leaves the facility, in what form and to whom; does the system keep working if the internet connection is lost?
7. Is there face recognition or biometric processing? Are recordings event-based or continuous; how are access permissions and retention configured?
8. With which interface is MES, ERP, VMS or PLC integration done, and is it included in the quote?
9. On what basis was the server hardware sized? If we add another scenario, will additional hardware be needed?
10. How do model updates and retraining work, what are their frequency and cost, and is it possible to roll back to a previous version?
11. How quickly do you come on site in the event of a failure or a drop in accuracy, and under what rules is remote support provided?
12. Can we see a similar installation in person?
13. What items make up the price; is the licence per camera, per scenario or per server, what does annual maintenance cover, and what will the cost be in the second year?
14. Who owns the data collected from our site, the labels and the trained model; what is handed over to us when the contract ends?

Which signs are red flags?

During the quote process, the following signs are red flags:

• Selling with a demo video: a carefully selected video does not show what the system will do on your camera and in your lighting; a demo is no substitute for a pilot.
• A claim promising the same high rate in every scenario and at every site: accuracy depends on camera angle, lighting, distance and how the event is defined; a quote guaranteeing 99% everywhere is not explaining how the rate was measured.
• A large installation without a pilot: a supplier that wants to sell dozens of cameras and servers in one go and avoids writing down success criteria leaves the risk with you.
• No clear answer on where the data is processed.
• Integration left out of the quote and postponed until later.
• Only successful examples shown in the pilot report.

How does CX Teknoloji meet these criteria?

At CX Teknoloji, we work from GTU Technopark in Gebze and, alongside off-the-shelf solutions for OHS, site security, production and industrial quality, we develop project-specific models. We carry out site surveys, pilots and support on site at facilities in Kocaeli, Istanbul, Bursa, Sakarya, Tekirdağ and Yalova (Marmara service regions). Before installation, we assess your camera inventory free of charge, define the pilot's success criteria with you in writing, and measure and report precision and recall on your site. By default, processing takes place on a GPU server inside the facility (processing can also be done in the cloud if requested), and face recognition is not used; events are transferred to MES and ERP via REST API and webhook, and on-line decisions go to the PLC via digital output or a fieldbus network. The values on our solution pages, for example 99% for licence plate reading or 90% for hard hat detection, are field values and are measured again on your site during the pilot. For project-specific models, the typical deployment time is 12–14 weeks. You can ask us these 14 questions too; we will give our answers in writing.

Frequently asked questions

What should you look at first when choosing a computer vision company?

The first thing to look at is whether the company has results measured on site for your type of problem. On-line quality control and OHS analytics on security cameras require different expertise. Next, ask under what conditions and with what method accuracy was measured, whether the company provides written pilot success criteria, and where the data is processed. A demo video or a single percentage is no substitute for these questions.

How long should a pilot last in a computer vision project?

For an off-the-shelf solution, installation is completed within a week; the pilot usually fits into a 30-day period: a first week of calibration in which camera angle, zones and thresholds are adjusted, followed by a few weeks of measurement in which thresholds are not changed. For project-specific models, feasibility, data collection and training come first; typical deployment is 12–14 weeks. More important than the duration is that the success criteria and measurement method are written down before the pilot begins.

Can an accuracy rate be written into the contract?

It can, but it is meaningful only when written together with its method. The contract should define what counts as a correct detection, the out-of-scope conditions, and on which cameras, in which shifts and with how many days of data precision and recall will be measured. A percentage with an unclear method creates disputes. What will be done if the criteria are not met, for example an additional calibration round or a change of scope, should be set out in the same document.

What is the advantage of working with a local computer vision company?

Time to reach the site and ease of communication are the most tangible advantages. Computer vision systems may require on-site intervention for reasons such as camera drift, a dirty lens or a line change; a team in the same region can carry out this intervention quickly. Familiarity with KVKK and local legislation, and documentation and training in Turkish, also make a difference during operation. Having the support model in writing matters more than the product's origin.

Who should own the trained model and the data?

This should be written explicitly into the contract; otherwise problems arise when the contract ends. The images and event records collected from your site are generally regarded as the facility's data; the usage rights to the labels and the model trained with this data, however, depend on the parties' agreement. Which records will be handed over in which format when the contract ends, and how copies held by the supplier will be destroyed, should be determined in advance.

Is a demo meeting enough to make a decision?

No, a demo is useful only for initial screening. A short demo meeting shows whether the supplier understands your problem and whether your existing cameras are suitable. How the system will work in your lighting, at your camera angle and in your workflow can only be seen through a pilot with written criteria. A supplier that does not accept the difference between a demo and a pilot should be evaluated with caution.