Skip to content
dhii.ai

Model Training & ML

Computer Vision Development

Cameras and images turned into structured, actionable data — defects caught, items counted, documents read and events flagged.

Feasibility test
1–2 weeks
Production model
4–10 weeks
Edge deployment
+2–4 weeks

Seeing at scale

People are excellent at visual judgement and terrible at doing it for eight hours straight. Inspecting products on a line, checking photos submitted with insurance claims, counting stock on shelves, reviewing CCTV for safety incidents, or reading meter photos — all of these are repetitive visual tasks where attention drifts and errors creep in. Computer vision handles the volume and lets people focus on the cases that need them.

Your conditions, not benchmark conditions

Public benchmarks don’t have your lighting, your camera angles or your particular defects. The quality of a vision system depends almost entirely on how well its training data reflects reality. We begin by collecting representative images — including the awkward ones: reflections, motion blur, partial occlusion, the rare defect that matters most — and we label them with written guidelines so annotators agree.

Choosing the approach

Today there are two broad options. Vision-language models can describe or classify images from a prompt with no training — ideal for varied, lower-volume tasks. Trained detectors and segmenters such as YOLO-family or DETR-family models are faster and more precise for fixed tasks at scale. We prototype with both and pick based on accuracy, latency and running cost.

From model to operation

Predictions need to reach people and systems. We build pipelines that ingest from cameras, uploads or storage buckets, run inference on cloud GPUs or edge hardware, and push results to your MES, ERP, dashboard or alerting tools. A review interface lets staff correct predictions, and those corrections become training data for the next version. Operating thresholds are set with you, trading off missed defects against false alarms according to what each costs your business.

Deliverables

What you get

  • 01

    Feasibility test

    A quick experiment on your images to estimate achievable accuracy and the data needed.

  • 02

    Labelled dataset

    Annotated images or video frames with a labelling guide and quality checks.

  • 03

    Trained vision model

    Detection, segmentation, classification or OCR model tuned for your conditions.

  • 04

    Inference pipeline

    Batch or real-time processing, on cloud GPUs or edge devices.

  • 05

    Review interface

    A simple UI to inspect predictions, correct errors and feed corrections back.

  • 06

    Performance report

    Precision and recall per class, failure examples and recommended operating thresholds.

Process

How a computer vision development project runs

  1. Step 1

    Capture

    Collect representative images, including hard cases — glare, angles, rare defects.

  2. Step 2

    Label

    Annotate with clear guidelines, using model-assisted labelling to save time.

  3. Step 3

    Train

    Fine-tune modern pretrained architectures and compare against vision-language baselines.

  4. Step 4

    Deploy

    Optimise for the target hardware and integrate with your systems.

  5. Step 5

    Improve

    Collect misclassified examples and retrain on a schedule.

Engagement models

Ways to work together

Every engagement starts with a fixed, written quote. See engagement models & pricing.

FAQ

Computer Vision Development: common questions

How many images do we need?

With modern pretrained models, a few hundred well-labelled images per class is often enough to start, and more for rare or subtle defects. The feasibility test tells you where you stand.

Can vision-language models replace custom training?

For low-volume or varied tasks, a vision-language model prompted well can be good enough and needs no training. For high-volume, real-time or precision-critical work, a trained detector is faster, cheaper and more consistent. We test both.

Can it run on-site without the cloud?

Yes. Models can be optimised to run on edge devices such as NVIDIA Jetson or a local GPU server, which helps with latency, bandwidth and privacy.

What about privacy when cameras capture people?

We minimise personal data: blur or crop faces where they aren't needed, avoid facial recognition unless there is a clear lawful basis, and keep retention short. Your local regulations always apply.

Related services

Used in

Start a project

Have a workflow that should run itself?

Tell us what you want automated or built. You get a written assessment of feasibility, risks and a fixed quote — usually within two business days.