How it works
Four stations from CAD to a verdict
Viseena replaces weeks of defect collection and hand-labeling with a pipeline you can run per SKU. Here is what happens at each step, who does what, and how long it should take.
Process at a glanceTargets
- Steps
- 6
- New SKU
- 24 h
- Real frames
- 50–200 per SKU
- Shadow mode
- 1–2 weeks
- Go-live
- Engineer approval
Step by step
Timings are design targets for a part with CAD and an installed camera station.
Define
Upload CAD and material, pick defect classes from the taxonomy, and mark the faces that matter.
Generate
Synvex renders thousands of labeled variants per class across lighting, pose, camera and texture.
Train and align
Coreon fine-tunes an adapter and aligns it with 50 to 200 real frames of good and known-bad parts.
Validate
Run the model in shadow mode beside current inspection and compare verdicts on real parts.
Deploy
Switch the station to live, with thresholds for reject and human review.
Learn
Uncertain frames go to review. Confirmed ones retrain the adapter on a schedule you approve.
Who does what
| Step | Your team | Viseena |
|---|---|---|
| Define | CAD, material specs, defect priorities | Taxonomy mapping, scene setup |
| Generate | Review sample renders | Synthetic runs and realism tuning |
| Train and align | Provide real good and known-bad parts | Training, alignment, evaluation report |
| Validate | Run shadow mode, review disagreements | Accuracy report on real parts |
| Deploy | Approve thresholds and PLC actions | Edge setup, OPC UA wiring support |
| Learn | Confirm uncertain frames | Retraining and staged rollouts |
A station is one file
Each Veridex station is described in a short YAML file: which cameras, which model, which thresholds and which PLC action. The file is versioned in the control plane alongside the model.
- Thresholds for reject and review, per defect class
- PLC actions over OPC UA (planned)
- Opt-in learning per station
# station.yaml: one Veridex station (example) station: line-a-st2 sku: AX-204 cameras: - { id: cam1, model: gige-5mp, roi: top-face } - { id: cam2, model: gige-5mp, roi: edge-left } model: coreon://ax-204/adapter@v7 thresholds: reject: 0.90 # above: divert the part review: 0.55 # between: route the frame to a person actions: reject: opcua://plc-a/ns=2;s=Divert learning: send_uncertain_frames: true
What a pilot is designed to show
Defective parts the model passed, measured against your reference inspection.
Good parts the model would have scrapped.
Hours from CAD to a validated model for a new part.
Share of frames sent to a person.
About the process
How many real images do you need?
Our design target is 50 to 200 real frames per SKU, mostly of good parts plus any known-bad parts you have. Synthetic data covers the rare classes.
What if we have no CAD?
Synvex works best from CAD. Without it, a 3D scan of a good part can stand in, with more alignment work.
Who approves retraining?
You do. Retrained adapters are staged and need an engineer's approval before they reach a live station.
Walk through the process on your part
We will map your SKU, defects and station to these steps in one call.