Image models that count, inspect and recognise in real-world conditions, monitored and maintained after launch.
Trusted by leading organisations
Computer vision is the branch of AI that enables software to interpret images and video: finding, counting, classifying or precisely outlining objects. Data Science Lab develops computer vision applications for organisations that currently count, inspect or check things by eye, covering everything from cameras and labelling to integration with your systems and ongoing management.
At GVB, passengers on vehicles were counted by hand: labour-intensive, prone to errors and only ever a snapshot. In a factory, a quality inspector checks samples at the end of a shift; anything they miss is noticed by the customer. A camera paired with a model can do this work continuously, provided the model can handle real-world conditions.
A standard model trained on a few good photos can make for a convincing demo. In production, it fails because of what those photos left out: backlighting, reflections, a dirty lens, a camera repositioned after maintenance or a new product. Training images therefore need to reflect the real environment, across different times of day, seasons, weather conditions and viewing angles.
A computer vision project starts with where the camera is mounted and the decision that needs improving, not with the model. It is one of the AI applications we build.
A detection model draws a box around every object it recognises: a person, a vehicle or a pallet. By tracking each object across successive frames, you can count what passes by or is present without anyone having to review the footage.
In classification, the entire image receives a label, such as approved or rejected. For Floating Farm, we use image classification to gain insight into the feeding process. Labelling is simpler than with detection, but the model does not tell you where something is.
Segmentation identifies, pixel by pixel, which parts of an image belong to an object. You need this when shape or area matters, such as when measuring the extent of damage. Labelling takes more work because someone has to trace the outline precisely.
Resolution, viewing angle, lighting and frame rate determine what a model can see. Training data therefore needs to include backlighting, reflections, blur and night-time images. Sometimes an extra light or a different camera position achieves more than a larger model.
A model learns from labelled examples. We write instructions for edge cases, such as partially obscured or cropped objects, and have more than one person assess ambiguous cases. For detection, every object must be labelled; a partially labelled image teaches the model the wrong thing.
On a device near the camera, known as the edge, you process images locally and send only counts or alerts onwards. This saves bandwidth, works offline and keeps images on site. The cloud offers more computing power, but the images then have to be transmitted.
Describe the setting and the decision. Together, we’ll assess whether computer vision is the right tool.
We establish which decision needs to improve, what the model needs to see to support it, and which error is worse: a missed detection or a false alarm.
We collect images from the actual cameras at different times and in varying light. Where needed, we advise on camera position, lenses or lighting.
Using clear labelling instructions, we build a training set and a separate test set containing challenging situations that the model does not see during training.
We measure precision (how many alerts are correct) and recall (how many real cases the model finds) for each condition, such as day and night.
We deploy the model where it best fits the bandwidth, privacy and management requirements, and send its output to the system people work in.
After launch, we track the results and a team member reviews samples. This helps us spot drift early and retrain the model with newly labelled images.
It’s great to have been involved in this Data for Good project from the start. An end-to-end data solution, from mounting and installing the cameras ourselves to implementing our computer vision model.
Once people can be identified in footage, you are processing personal data and the GDPR applies. If cameras are installed in the workplace, you need a legal basis, often legitimate interest. You must inform employees and obtain consent from the works council or employee representatives. You may not use footage to monitor how employees do their jobs. Facial recognition is generally prohibited unless a legal exception applies, and the EU AI Act prohibits emotion recognition in the workplace and in education, except for medical or safety reasons. A DPIA is mandatory for systematic, large-scale monitoring of publicly accessible areas.
It therefore pays to design the system to generate as little personal data as possible: process data at the camera where you can, store only counts and do not retain footage if it is not needed. At GVB, footage from the security cameras is not viewed or shared; it is only analysed by the model. Your privacy officer or legal adviser should assess whether your application is permitted and on what legal basis.
A model may perform differently in practice than it does on the test set, and its performance can change: seasons, new products and repositioned cameras alter the images it receives. This is called drift. The built-in monitoring in Azure Machine Learning is designed for tabular data, so for images we set up our own indicators, such as the distribution of confidence scores and the number of detections per hour. For Rijkswaterstaat, we built a generic monitoring package to detect drift and outliers in production models.
Any organisation can start an AI pilot. The challenge comes with the next step: making AI part of the organisation. That’s where Data Science Lab makes a difference.
Read more about us
10+
8.7
200+
Computer vision is the branch of AI that enables software to interpret images and video. Its main tasks are classification (what is in the image), detection (where each object is) and segmentation (which pixels belong to an object). Counting requires a different approach from measuring the extent of damage, so the question determines the task.
That depends on the task, the number of classes and how much conditions vary. With a pre-trained model and transfer learning, you can start with fewer images of your own. Variety matters at least as much as quantity: images from one sunny morning teach a model little about conditions at night. After an initial assessment, we can give you a more specific answer.
Yes, if their resolution, viewing angle and frame rate suit what the model needs to see. At GVB, footage from security cameras in the vehicles is used to count passengers. For the Floating Farm, we installed cameras ourselves. We assess the existing footage first.
Either is possible. Processing near the camera makes sense when connectivity is limited, a fast response is needed or images cannot leave the site; only the result is then sent on. The cloud offers more computing power. A combination is also possible: running the model locally while deploying and monitoring it from the cloud.
That depends on the purpose, location and legal basis. Counting is different from identifying someone: under the GDPR, data is only considered biometric if technical processing enables or confirms unique identification. Remote biometric identification and, where permitted, emotion recognition are classified as high-risk under the AI Act, with obligations applying from 2 December 2027. We help you assess what is involved and classify the risks; your legal counsel or compliance team makes the decision.
Through monitoring, spot checks and retraining when the environment changes, and by keeping the labels, test set and model under your own control. Platforms change too: Microsoft discontinued data labelling in Azure Machine Learning on 30 September 2026 and will continue to support Custom Vision until 25 September 2028. We can manage applications we built ourselves through management and continuous improvement.
If a simple sensor, such as a light barrier, load cell or barcode scanner, can measure the same thing directly, it is easier to manage than an image-based model. Another approach is often more sensible for rare events with hardly any example images, or when errors could have serious consequences and there is no human oversight. Sometimes better recording in the existing system is enough.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Describe the location, the cameras and the decision you want to improve. Together, we’ll assess whether computer vision is the right approach.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
