Computer Vision

Computer vision development services that run on your hardware

We build computer vision prototypes and production systems for inspection, detection and segmentation, trained on your own images, tested against an accuracy target you set, and deployed on the cameras, CPUs, NPUs or drones you already use. Video stays on site unless you choose otherwise.

  • 99.7% per-space F1 for parked cars, on CPU-only servers
  • 99.2% dish recognition on an edge device, built in 6 weeks
  • 92% landmine detection from drone imagery, vs 70% by hand
  • 5.0 Clutch rating, 22 client reviews

Why teams call us

When do teams call a computer vision company? Usually when the current system fails on site

  • 01

    False rejects after a change

    Your inspection system floods the line with false alarms after a lighting, tooling or product change. We retrain and re-test it for the new conditions, as we did for Gentherm.

    See the Gentherm case
  • 02

    Too slow on site hardware

    The model works in the lab but misses its time budget on the servers you have. We pick and compress a model that fits, without new GPUs, as for Abel Sensors.

    See the Abel Sensors case
  • 03

    Objects that overlap or hide

    Bikes in two-tier racks, mines under vegetation, ingredients on a plate. We start the test set with the hardest images you have.

    See the landmine case
  • 04

    Video that can’t leave the site

    Privacy or bandwidth rules keep images on the premises. We run the whole pipeline on the edge device, with nothing sent to the cloud.

    See the food quality case

At a glance

Computer vision development services in brief

  • What we build

    Inspection, defect detection, counting, segmentation, tracking and pose models, trained on your images and wired into your systems.

  • Where it runs

    Intel CPUs with OpenVINO, NPUs, NVIDIA GPUs, drones and on-premise servers. The cloud only if you want it.

  • Proof

    99.7% per-space F1 on CPU servers, 99.2% dish recognition on an NPU, 92% landmine detection from drones.

  • Where to begin

    A feasibility check on your own images: from €5,000, a fixed fee for 2–3 weeks. Then a working prototype at a fixed price, before any production build.

Quick check

Do you need a custom vision system? And when to wait

You need this if

  • People inspect parts, products or footage by eye, and the volume keeps growing.
  • A vision system exists, but its false rejects or misses cost you money.
  • Images or video can’t leave your site, or the network is too slow to send them.
  • You know which error costs you most and can share sample images.

Skip it if

  • You need cameras, lenses or lighting only. A machine-vision integrator fits better; we work with them on the model.
  • An off-the-shelf tool already meets your accuracy target. Use it; the check would tell you so.
  • You want AI across many workflows, not vision. AI Transformation fits that.
  • The product is a regulated medical device. Start with Regulated AI & Compliance.

Delivery path

How does a computer vision project run? Five stages, one team

Check, Prototype, Build, Deploy, Monitor. The check is a fixed-scope project. Each later stage is quoted from what it finds, and you can stop after any of them, including after the prototype.

  1. 01

    Check

    2–3 weeks · fixed fee

    We test a baseline model on your own images, agree the target metric and try it on your hardware.

    Deliverable: a go/no-go, the metric in writing and a fixed quote.

  2. 02

    Prototype

    Fixed price · quoted from the check

    We label a first dataset and build a working model on your images and target hardware, so you see it run before you commit.

    Deliverable: a working prototype and its test report.

  3. 03

    Build

    Fixed price per phase

    Annotation specialists label at scale, and we compare model families and tune the best one until it meets the target on held-out images.

    Deliverable: a production model with a test report per class and per site.

  4. 04

    Deploy

    On your hardware

    We quantise and package the model for your CPU, NPU, GPU or drone, and connect it to the software that uses its output.

    Deliverable: a tested inference module or on-device pipeline.

  5. 05

    Monitor

    Monthly

    We track accuracy in production and retrain when the line, lighting or product changes.

    Deliverable: model versions you can trace and roll back.

What we build

What computer vision services do we build? Six kinds of system

Most projects are computer vision for manufacturing and inspection, where a missed defect or a false alarm has a price. The rest sit in mobility, field surveys, research and healthcare.

Automated visual inspection
Defect detection and OK/NOK decisions on the line, with machine blocking where the process needs it.
Detection and counting
Objects, ingredients, vehicles or people, counted per frame, per plate or per space.
Segmentation
Pixel-level masks for overlapping objects, corrosion and coating damage.
Tracking and pose
Animals, people or parts followed across video, with keypoints.
Aerial and multispectral
Drone imagery turned into GPS-tagged detections for field surveys.

Need a camera rig as well? We work with your machine-vision integrator and own the model side.

Runs where your data is

Can computer vision run without a GPU or the cloud?

Yes. We choose the model for the hardware you have, then compress it until it meets your time budget.

  • Intel CPUs OpenVINO inference on the servers already on site, as in the parking project.
  • NPUs Quantised models on an NXP edge device, as in the food quality project.
  • NVIDIA GPUs We are an NVIDIA partner. Detection, pose and tracking run at 5–7 s per minute of video on one GPU.
  • Drones Multispectral survey imagery processed into flagged, geotagged hazards.
  • Your software Results delivered in the format your existing system already reads.

Before the check

What does a vision brief need? Six things we ask for first

Sample images
Images from real conditions, including the cases your current setup gets wrong.
Capture conditions
Cameras, lighting, distance, and how they change over a shift or a season.
The costly error
Which mistake costs more: a missed defect or a false alarm.
The decision
What happens with the output: reject a part, count stock, raise an alert.
Target hardware
The device or server the model must run on, and what you could add.
Time budget
How fast a result is needed: per frame, per batch or per day.

Send us these six and we scope the feasibility check in one call.

How we price

How much does computer vision development cost?

The fit call is free. The feasibility check is computer vision consulting with a deliverable: a fixed scope and a fixed fee from €5,000, agreed before you sign. The prototype and the production build are quoted from its results at fixed prices, and running the model is a monthly retainer.

  • Feasibility check

    From €5,000, fixed fee

    A baseline on your images, the target metric in writing, a test on your hardware and a fixed quote for the prototype.

    2–3 weeks · go/no-go report you keep

  • Prototype

    Fixed price, quoted from the check

    A working model on your own images and target hardware, tested against the agreed metric, so you see it run before you commit.

    Working prototype and test report

  • Build and run

    Quoted per phase, then monthly

    Production dataset and model, deployment on your hardware, then monitoring and retraining sized as workstream capacity, not hours.

    1 workstream = 160 story points a month

Every check includes: a baseline on your images, the metric agreed in writing, a test on your hardware and a call with your engineering lead.

What sets the fee: the number of classes, how many images need labels, the hardware target and how strict the accuracy target is.

Confidentiality: we sign an NDA before you share images, and a data processing agreement before we handle personal data.

Fit call: 30 minutes, free, to decide whether a check makes sense at all.

Need an investor demo rather than a production system? See Investor-Ready Prototypes. Need senior technical leadership for the whole product? See CTO as a Service. Not sure vision is the right first AI project? Start with AI Consulting.

Prices exclude VAT where it applies.

Results by hardware

How accurate is it, and on what hardware? Ten tasks from shipped projects

Accuracy reached on ten computer vision tasks, with the data and hardware each one runs on
TaskData and hardwareResultSource
Parking occupancy, carsIntel CPU servers, no GPU99.7% per-space F1 on 139 test images from 7 car parksCase: parking space segmentation
Parking occupancy, bikes in two-tier racksIntel CPU servers, no GPU82% average per-space F1Case: parking space segmentation
Dish recognition and ingredient countingNXP edge device, on the NPU99.2% dish recognition; 79–99% counting accuracy by ingredientCase: edge AI food quality control
Landmine and ordnance detectionDrones with multispectral cameras92% detection accuracy vs 70% by hand; 20× more ground a dayCase: landmine detection with drones
Rust and coating damageDrone imagery, segmentation models99% rust vs non-rust detection; 60% lower manual inspection costCase: drone-based corrosion monitoring
Heating-wire position on the lineProduction-line inspection camerasAbout 1.5 mm verification tolerance; 100% automated OK/NOK decisions with machine blockingCase: industrial computer vision quality control
Mouse detection, pose and trackingOne GPU0.91 AP50 for detection; 5–7 s per minute of videoCase: home-cage behaviour analysis
Breast cancer detectionMammography and ultrasound images+21% diagnostic accuracyCase: breast cancer detection platform
Skin condition and severityPatient skin photos18% more accurate than competing toolsCase: dermatology software platform
Blood pressure from face videoCamera video of the face90% accuracy in blood-pressure measurementCase: real-time vital signs detection

Figures as published in each case study. Results on your images are measured in the feasibility check.

Shipped projects

Which computer vision systems have we shipped?

Each project ran on images from the client’s own sites, on the hardware they already had. Figures are taken from the published case studies.

  • 99.7% per-space F1 for cars across 7 car parks
    Smart parking · Abel Sensors

    Parking space segmentation for cars and bikes

    Instance segmentation on Intel CPU servers already on site, with no GPU; 82% per-space F1 for bikes in two-tier racks. See the case: parking space segmentation

  • 99.2% dish recognition on the edge device
    Food service · multi-site operator

    Edge AI food quality control for fast-food kitchens

    Detection and counting run on an NXP NPU with no kitchen video sent to the cloud, six weeks from brief to working pipeline. See the case: edge AI food quality control

  • 92% detection accuracy, vs 70% by hand
    Humanitarian demining

    Landmine detection from drone imagery

    Multispectral drone surveys cover 10,000 m² a day instead of 500 m² on foot. See the case: landmine detection with drones

  • 100% automated OK/NOK decisions with machine blocking
    Automotive Tier 1 · Gentherm

    Industrial computer vision quality control

    Inspection stabilised after a production-line and lighting redesign, verifying heating-wire position within about 1.5 mm. See the case: industrial computer vision quality control

  • 0.91 AP50 for mouse detection, 3× the off-the-shelf model
    Preclinical research

    Home-cage behaviour analysis from video

    Detection, pose and tracking process one minute of video in 5–7 seconds on a single GPU. See the case: home-cage behaviour analysis

More computer vision work: Drone-based corrosion monitoring · AI breast cancer detection · AI dermatology software · Vital signs from face video · Medical image-to-report conversion

Alternatives

Should you build, buy a platform or hire? Three ways to get a vision system

An off-the-shelf platform is quickest when your use case matches the product. A hire makes sense once vision is core to your business. A custom build fits when the hard cases, the hardware or the data rules are yours alone.

A custom computer vision build with Blackthorn AI, an off-the-shelf vision platform and an in-house hire compared
CriterionCustom build with Blackthorn AIOff-the-shelf vision platformIn-house hire
Fits your hardwareModel chosen and compressed for your CPU, NPU, GPU or droneRuns on the devices the vendor supportsDepends on who you hire
Accuracy on your imagesMeasured against your target before the buildMeasured after you buyMeasured once the team is in place
First resultA 2–3 week feasibility checkFast, if your use case matches the productAfter recruiting and setup
Hard casesCustom models for overlap, small objects and new defect typesLimited to what the product supportsPossible with the right specialist
Ongoing costMonthly retainer, or handover to your teamLicence per camera or deviceSalaries, tooling and labelling

Comparison reflects typical options, not specific products or people.

Why Blackthorn AI

Why choose a PhD-led team for computer vision?

  • Your images first. We test on your data before we quote the build.
  • Hardware you have. We fit the model to existing CPUs, NPUs or drones before suggesting new kit.
  • Hard cases are the job. Overlap, occlusion and small objects go into the test set first.
  • Plugs into what exists. Output goes into your current software, as the parking model fed the client’s occupancy logic unchanged.
  • You can stop after the check. The go/no-go report is yours either way.

Who builds it?

An AI architect owns the plan, computer vision engineers train and deploy the model, and annotation specialists label your data. The people on your project are named in the proposal.

  • Alex Gurbych, PhD, Co-founder and Director
  • Danylo, Senior Computer Vision Engineer
  • Maksym, Computer Vision Engineer

Page reviewed by Alex Gurbych, PhD, and Maksym, Computer Vision Engineer, 21 November 2026.

Meet the scientists

Client reviews

What do computer vision clients say?
22 reviews, 5.0 on Clutch

  • Verified on Clutch

    «Their team was flexible and agile»

    AI-driven drone system for landmine detection: computer vision model, detection software and field tests.

    Oleksii Stupak, CEO, Aleannlab LLC

  • Verified on Clutch

    «All our needs were addressed in a very professional manner»

    Computer vision and image annotation for skin diagnosis, from strategy to training and validation.

    Yoanna Gouchtchina, CEO, ONE.SKIN Technologies

  • Verified on Clutch

    «blackthorn.ai quickly delivered a working prototype»

    AI tools for a biotech company’s scientific teams, delivered as a working prototype on time.

    Scientist, under NDA, BioTech company, San Diego

FAQ

Computer vision development questions, answered

01 What do computer vision development services include?

A feasibility check on your images, a working prototype, data labelling, model training against an agreed accuracy target, deployment on your hardware, and retraining when conditions change. At Blackthorn AI each stage is quoted separately, so you can stop after the check or the prototype.

02 How much does a computer vision project cost?

The 30-minute fit call is free. The feasibility check is a fixed fee agreed before you sign: from €5,000 in Europe, £4,700 in the UK and $6,000 in other regions, excluding VAT where it applies. The prototype and the production build are quoted from the check’s results at fixed prices, and running the model is a monthly retainer.

03 Can you build a computer vision prototype first?

Yes. After the feasibility check we build a working prototype on your own images and target hardware, tested against the agreed metric, at a fixed price quoted from the check. You decide on a production build only after you have seen it run.

04 How long does a computer vision project take?

The feasibility check takes 2–3 weeks. After that it depends on the task: the on-device food quality pipeline went from brief to a working system in 6 weeks, while the drone landmine project ran for 30 weeks, including field tests on several terrains.

05 Can computer vision run without a GPU?

Yes. The parking occupancy model runs on Intel CPU servers already on site, with OpenVINO, and meets a budget of 10 seconds for a batch of 90 camera frames. The food quality models run on the NPU of an NXP edge device. Where a GPU is the right choice, we work as an NVIDIA partner.

06 Can images and video stay on our premises?

Yes. The food quality system runs entirely on the edge device and sends no kitchen video to the cloud. We can also train and run models inside your own environment.

07 How do you measure accuracy?

On held-out images from your own sites, with the metric agreed before the build. For parking we reported per-space F1 on 139 test images from seven car parks; for detection tasks we report average precision per class.

08 How many images do we need?

It depends on how many classes you have and how much they vary. The feasibility check measures a baseline on the images you already have and tells you how many more need labels before the build. A small set of hard cases is worth more than a large set of easy ones.

09 Do you label the data?

Yes. Annotation specialists label boxes, masks and keypoints in tools such as CVAT. The food quality project had four annotation specialists on the team.

10 Which models do you use?

Whichever performs best on your data and hardware. The parking project moved to RF-DETR instance segmentation; corrosion detection compared PSPNet, ESPNet and U-Net segmentation models; the food quality pipeline uses quantised detection and segmentation models on an NPU.

11 What happens when our line, lighting or product changes?

We retrain and re-test the model against the same target. The automotive quality control project started exactly there, after a full production-line redesign with new lighting and tooling.

12 Can you work with our existing camera or vision system?

Yes. The parking model’s output fed the client’s existing occupancy logic and reporting unchanged. We can work alongside your machine-vision integrator, who handles cameras and lighting while we build the model.

13 Is computer vision the same as machine vision?

Not quite. Machine vision usually means cameras, lighting and rule-based software on a production line. Computer vision is the wider field of models that interpret images and video. We build the models and connect them to the cameras and machine-vision systems you already have.

14 Do you work on medical images?

Yes. Our breast cancer detection platform raised diagnostic accuracy by 21% on mammography and ultrasound, our dermatology models were 18% more accurate than competing tools, and our face-video system measures blood pressure with 90% accuracy. Software that is a regulated medical device follows the IEC 62304 and MDR route on our Regulated AI & Compliance page.

15 How do you handle personal data in images?

We sign an NDA before you share images and a data processing agreement before we handle personal data. Models can be trained and run in your environment or on the device, so images stay with you, and faces or number plates can be blurred where the task allows.

Bring sample images. Leave with a go/no-go.

Share images from real conditions and the error that costs you most. We tell you what accuracy is realistic on your hardware, and what it takes to get there.