Get the right image
Camera and lens selection, lighting, positioning and calibration. Establish what can be observed reliably before choosing a model or promising a result.
Hardware meets software · Christchurch, New Zealand
Computer vision built around the real task.
Turn images into useful detections, measurements and records. We work across image capture, model evaluation and integration, so the result fits the process it needs to serve.
From camera to decision
The model is part of the system.
A useful result depends on what the camera sees, what the data represents and what happens after a prediction. We scope those parts together, whether you are exploring a new workflow or improving an existing implementation.
Build the vision task
Detection, segmentation, tracking, OCR and visual inspection. Prepare datasets and labels, evaluate existing models, and scope training or adaptation where the task needs it.
Measure the errors
Test representative conditions and difficult cases. Measure missed detections, false alarms and measurement error; select thresholds around the consequences of each mistake.
Connect it to the workflow
Integrate cameras, edge or cloud processing, operator interfaces and business systems. Define how results are recorded, uncertain cases are reviewed and equipment or software may respond.
Establish feasibility before expanding the build.
01
Define the task
Identify what must be recognised or measured, the operating conditions and the cost of mistakes. Review available images and the capture setup.
02
Evaluate a bounded approach
Agree a representative test set and compare a suitable implementation against the acceptance criteria. Record the conditions where it works and where it falls short.
03
Integrate and maintain
Scope the production implementation, review paths and monitoring. Keep model, data and configuration versions traceable so changes can be evaluated.
A result you can question.
An evaluation should show the task, test conditions, measured errors and unresolved cases. Changes in lighting, camera position or the items being observed may require a fresh evaluation. Automated actions need their own controls beyond the model score.
Owned specimen · synthetic images
Inspect the result. See the failure cases.
Four generated image cases show a conventional vision rule counting shapes, making a false positive, missing joined objects and withholding a result. Inspect the actual input, output and reproduction source.
Explore the vision specimenQuestions before a build
What computer vision work can AcesForge help with?
We scope and engineer image-based workflows covering detection, segmentation, tracking, OCR, visual inspection and measurement. The work can include camera and lighting choices, data preparation, model evaluation or training, operator software and integration.
Do we need a custom-trained model?
Not necessarily. A feasibility assessment can compare conventional image processing and existing models with the task requirements. Custom training is considered when the available approaches do not meet the requirements and suitable training data can be obtained.
Can you guarantee an accuracy percentage?
No percentage is promised before assessing the task and data. We agree metrics and test conditions, evaluate on representative held-out cases, and report limitations. A confidence score alone does not prove accuracy; thresholds need validation against labelled outcomes.
Can it run on our equipment rather than in the cloud?
Edge and cloud deployment can both be assessed. The decision depends on latency, connectivity, compute capacity, data handling and maintenance requirements. Hardware compatibility and performance are established during scoping and testing.
What do we receive from a feasibility assessment?
The agreed deliverables can include an image-capture recommendation, dataset and labelling requirements, a bounded prototype or evaluation, measured results and a staged implementation plan. The scope makes clear which of these are included before work starts.
What should we bring to an initial discussion?
Describe the task, current process, operating conditions and consequences of errors. Tell us whether sample images or equipment are available. We agree an appropriate way to share data before you send anything sensitive.
Start with the process you want to improve.
Tell us what needs to be seen, measured or extracted, and what happens with the result. We will discuss the evidence needed to assess feasibility and scope the next step.
Discuss your projectExplore hardware and test engineering, or get an AI system review for an existing implementation.