Where AI actually fits in a digital twin, without the hype
Object detection finds things in a twin faster than a person scrolling through it. It does not replace the judgement of the person deciding what those things mean.
Every platform in this space now claims some form of AI. Usually the claim is vague on purpose, because vague is easier to sell than specific. We would rather say plainly what our detection does, how it works, and where the line sits between the software and the person relying on it.
Two models, one job
Detection inside a twin runs on a pairing of two established computer vision approaches. An object detection model finds and draws a box around things of interest, a fire extinguisher, a piece of plant, a visible defect. A segmentation model then traces the precise outline of that object within the twin, rather than a rough rectangle around it. Used together, the result is detection accurate enough to tag an object's exact location and extent in three dimensional space, not just flag that something is probably present somewhere in a room.
What detection is genuinely good at, and what it is not
Consistent, tireless scanning of a large capture for a defined set of known object types, far faster than a person reviewing footage.
Judging severity, context or intent. A crack the model flags still needs a competent person to decide whether it matters.
Work well on a class of object it has not been trained on. Detection accuracy depends directly on relevant training data for that specific use case.
That last point is worth dwelling on. A model trained to spot damaged fire doors on construction sites will not reliably spot fraud indicators on an insurance claim, and we do not pretend otherwise. Each detection use case on the platform is trained and validated for its specific job, not sold as one generic capability that does everything.
Why we say this out loud: a client who understands detection as a filter, not a verdict, uses it correctly, checks the flags that matter, and trusts the platform for longer. Overselling the AI gets a great first demo and a disappointed client six months in.
Detection is one of three layers behind every twin. The live sensor data layer, and how it keeps a twin current between captures, is next in this series. Read the next post.
See detection running on a real twin
We can show you exactly what our detection models are trained to find on your vertical, and where we would recommend a human review step regardless.
See the product