Computer vision

Models that detect, classify and measure what's in images and video, in the cloud or on the device.

Confidence threshold 0.50

Table 1. At this threshold.

Detections	Precision	Recall
6	0.83	0.83

Figure 1. Detections on a shelf image at a confidence threshold of 0.50. A higher threshold keeps fewer, surer detections (precision); a lower one misses less (recall). Example image and scores.

Start a project

What we build

Computer vision solutions

We build models that inspect, count, read and monitor from your cameras and images, running in the cloud or on devices at the edge.

Visual quality inspection
Detect defects on production lines against agreed acceptance criteria.
Detection and counting
Count and locate items, vehicles or stock on shelves from camera images.
Document images
Read text, tables and fields from scanned documents and photos (OCR).
Safety monitoring
Detect people in restricted zones or missing personal protective equipment in video.
Damage assessment
Assess damage to vehicles, property or goods from photos to speed up claims and returns.
Grading and sorting
Grade produce, parts or materials by size, colour or quality from line cameras.

Security cameras on a pole against a blue sky

How to get started

Three steps to a plan for computer vision

1Tell us what you needUse the project form or book a call. A few sentences about the goal is enough to start.
2Free technical consultationWe go through your goals, users, existing systems and constraints with you.
3Your planA detailed plan covering the right tech stack, architecture, timeline and budget. Then you decide.

Start a project Ask a question first

How it works

Four kinds of model we build

The same camera frame, as each kind of model outputs it. Choose a carton to see its result and what the line does with it.

It outputs: A box around each object, with its class and confidence.
Measured by: mAP, using IoU to decide whether a box matches the true one.

Carton 2

Class: carton · 0.97
Defect: dent · 0.91
Box: [308, 138, 54, 46]

Diverted to inspection

Tuned to your line

We set the confidence threshold with you. Move it to see which detections count and what gets missed, and point at any box to see what it is.

Confidence threshold0.50

Correct

False alarms

Missed

Precision

88%

Recall

78%

Glare, shadows and price labels can look like products; half-hidden products score low. The threshold is set from validation on your own images, against the cost of a false alarm and of a miss.

Selected detection

Glare on the shelf edge · 0.55

Counted: false alarm

Correct False alarm Missed

Our approach

How the work runs

Define the task
Classification, detection or segmentation, with acceptance criteria per class.
Collect and label
Gather representative images, including rare cases, and label them to written guidelines.
Train
Adapt pretrained models to your images through transfer learning.
Validate in real conditions
Test across lighting, angles and cameras, with precision and recall per class.
Deploy and monitor
Run in the cloud or on edge devices, and monitor performance as conditions change.

What we'll need from you

Having these ready keeps the work moving.

Sample images or camera access
Images or video from the real cameras and conditions, including rare cases.
Acceptance criteria
Quality or operations staff who define each class and review labels.
Site access
Help with camera placement, lighting and network or edge hardware on site.
Privacy sign-off
Review of what may be captured and stored, where images include people.

Who's on the project

Our team, working with quality team from yours.

¹, ¹, ¹, ¹, ²

¹ Mayrian ² Your organization

Computer vision engineer: Selects, trains and evaluates the vision models.

Services

Computer vision services

Computer vision consulting
Use-case selection, a camera and image assessment, and acceptance criteria for each class.
Data collection and labelling
Representative images, including rare cases, labelled to written guidelines.
Model development
Pretrained models adapted to your images and tested in real conditions.
Deployment, integration and support
Inference in the cloud or on edge devices, connected to your line systems or apps, and monitored after launch.

Deliverables

What you receive

Appendix A. What you receive

All code, data and documentation are handed over in your accounts and repositories.

In your hands

What the documentation looks like

Every project ends with documents your team can run with. Here is an excerpt of one of them.

Labelling guideline and evaluation

Carton inspection: dent class

Line 2 cameras · 4,800 labelled images

Label as a dent when

The surface is pushed in by 3 mm or more
Draw the box tight to the deformed area, not the whole carton
Creases along fold lines are not dents

Per-class results on the validation set

Class	Precision	Recall	mAP@0.5
Dent	0.93	0.90	0.92
Tear	0.91	0.86	0.89
Missing label	0.98	0.97	0.98

Measuring success

How success is measured

What we report on in computer vision projects. Which measures apply, and their targets, are agreed with you at the start.

Table 2. What we report, and when. Targets are agreed with you at the start.

Measure	Reported
Precision and recall per class	Validation set, per class
mAP	Validation set
IoU	Validation set
Character or field accuracy	Validation set, for OCR
False reject rate	Pilot on the line
Throughput and latency	On the target hardware

Measure

Precision and recall per class

For each defect or object type, how many detections were right and how many real cases were caught.

Reported

Validation set, per class

How we score accuracy

In the evaluation report, a detection counts only if it overlaps the labelled box enough. Drag the predicted box to see the overlap change.

Drag the blue box, or focus it and use the arrow keys.

Intersection over union

0.35

overlap÷both boxes

Too little overlap: scored as a miss and a false alarm

Detection accuracy (mAP) is reported at an IoU threshold, commonly 0.5, or averaged from 0.5 to 0.95 for stricter tasks such as measuring defects.

How to start

From first call to production

Start where you are. Each step ends with a decision, so you commit to the next one only when it makes sense.

Protocol

How an engagement runs

Each step ends with a decision on whether to continue

Free technical consultation

One or two sessions

Talk through the goal, the data you have and the systems involved.

Outputs

(a) A shortlist of use cases, ranked by value and feasibility
(b) A recommended next step
(c) A plan covering stack, architecture, timeline and budget

Book the consultation

Estimate the value

What it could be worth to you

Enter your own figures. The formula is shown, and the estimate can go with your enquiry.

Estimate

Manual inspection time

Computer vision · from your own figures

Inspection time freed = Items × seconds each × share automated ÷ 3,600 × hourly cost × 250 working days(1)

Items inspected a day12,000Seconds per manual check4 sShare checked automatically80%Loaded cost per hour$32

Result

$85,333

a year

Add to my enquiry

Build or buy

When you don't need
a custom build

Part of the free technical consultation: when an existing product covers the need, we recommend it instead of a custom build. These are the options we weigh, alongside the tools you already have.

Related work

Existing products that may be enough

[1]Google Cloud Vision, Amazon Rekognition or Azure AI VisionWhen general tasks such as common object labels or reading printed text.
[2]Document AI services such as Azure AI Document Intelligence, Amazon Textract or Google Document AIWhen standard documents such as invoices, receipts and forms, using their prebuilt models.

Our approach

When a custom build is worth it

Your products, defects or parts aren't in general models
Accuracy must be proven per class under your conditions
Inference must run on edge devices or on site
Results must drive line systems or applications

Technologies and standards

Chosen for your project

Built on the cloud you already use. Choose yours to see the services involved; we recommend the full stack in the free technical consultation.

Table 2. Managed services for each layer, by cloud. The highlighted column is the one you chose.

Layer	AWS	Azure	Google Cloud
Ready-made vision and OCR	Amazon Rekognition, Amazon Textract	Azure AI Vision, Document Intelligence	Cloud Vision API, Document AI
Custom model training	Amazon SageMaker AI	Azure Machine Learning	Gemini Enterprise Agent Platform
Labelling	SageMaker Ground Truth	Azure ML data labeling	CVAT or Label Studio on GKE
Edge devices	AWS IoT Greengrass	Azure IoT Edge	LiteRT on the device

Also runs on any of the three: NVIDIA Jetson, ONNX Runtime, Ultralytics YOLO, OpenCV.

Models and libraries

PyTorch
TensorFlow
OpenCV
Ultralytics YOLO
Detectron2

Labelling

CVAT
Label Studio

Edge deployment

NVIDIA Jetson
ONNX Runtime
TensorRT
OpenVINO

Cloud vision services

Amazon Rekognition
Azure AI Vision
Google Cloud Vision API

Next section

Generative AI

Assistants grounded in your content, evaluated before and after launch.

Previous: Machine learning

Data Analytics & AI

6 Generative AI

Can it read, write and act for us?

Questions

Common questions
about computer vision

How many images do we need?

We establish it with a pilot on your own images. Starting from pretrained models keeps the number needed down, and the pilot shows the accuracy to expect before a full rollout.

Can it run on our existing cameras?

In many cases. We assess your camera feeds, resolution and placement early, and recommend changes where they limit what a model can see.

What about people's privacy?

We design systems that avoid identifying people where it isn't needed: faces are blurred, images can be processed on the device, and only the results are stored.

How is a project priced?

Well-defined scopes are delivered as fixed-price engagements; when requirements are still evolving, we provide a dedicated team instead. Either way, the free technical consultation ends with a plan covering tech stack, architecture, timeline and budget, so you know the cost before work starts.

Computer vision

Computer vision solutions

Visual quality inspection

Detection and counting

Document images

Safety monitoring

Damage assessment

Grading and sorting

Three steps to a plan for computer vision

Four kinds of model we build

Tuned to your line

How the work runs

Define the task

Collect and label

Train

Validate in real conditions

Deploy and monitor

What we'll need from you

Who's on the project

Computer vision services

Computer vision consulting

Data collection and labelling

Model development

Deployment, integration and support