What happens to an edge AI demo when it leaves the booth?
Why a demo of edge AI needs more than a camera and a screen, and which conditions in operation raise the real questions.
I like good demos. But putting “AI at the edge of the network” on a screen is not enough to build a good demo. The hard part lies elsewhere.
What makes a demo good?
A good demo finds a scenario where someone stops and understands right away what it is about. Three questions must be answerable at a glance:
- What does the camera see?
- What happens on the device?
- Why should I care?
Finding the scenario is the real work. The technology alone does not carry a demo if it is not clear what is being shown. Someone who stops should see within a few seconds what the camera captures, what the device makes of it and what use that is.
What comes after the demo?
Then comes the question that interests me even more: what happens when the device leaves the booth?
In operation, things look different from the booth:
- The light is different.
- The camera is a different one.
- The latency limit is tighter.
- There is no convenient connection to the cloud.
This is where the polished demo ends and the interesting engineering conversation begins.
Why are these four points so important?
Each point changes an assumption the demo was built on.
With the light, what the camera actually sees changes. A system that worked with a certain image at the booth gets a different one in operation.
With the camera, the source of the data changes. A different camera delivers a different image, and the processing on the device has to cope with it.
With latency, the room to move changes. If the permitted time gets shorter, less time is left for processing on the device.
With the cloud, the place of the computation changes. If there is no convenient connection, the processing has to happen on the device itself. The device has to do everything that is needed on its own.
Why do I call this the uncomfortable part?
Because it is hard to fit into a demo. It consists of deviations: light that does not match the test, a camera that behaves differently, a time limit that cannot be moved. They are the part that is hard to show in a demo. A demo shows that an idea works in principle. The conversation about light, camera, latency and missing cloud shows whether it can work under the conditions of operation.
What is meant by “on the device”?
The device is at the center. The camera delivers images, the processing happens there, and the result has to be ready within the time limit. A connection to the cloud cannot be assumed. That is exactly why the trade fair booth, with its stable light and comfortable surroundings, is a different place from the hall where the device will later work.
How can you work with this?
Anyone planning a machine vision project can take the demo as a starting point and then collect the parts that do not fit neatly into a demo. I am interested in exactly these parts. The brochure can wait.
One question can be asked at any time: what would look different in my project once the device is no longer on a booth table but working where it is needed? The answers describe the work that comes after the demo.