Setting up Inno Counter on a line — and what to check on day one

Four steps in order, three numbers to watch, and the one figure to reconcile at the end of the shift.

Before the camera counts anything

Four things, in this order, and none of them is optional.

1. Point the camera where product crosses cleanly. A view where items overlap for most of their travel is the one case the filter cannot rescue.
2. Choose a line or a zone. A line suits a belt. A zone suits a tray — on a real bread line a line counted nothing in 35 seconds where a zone counted fifteen.
3. Teach it the product. Click on it in a frozen frame, or upload photographs. Both feed the same recognition.
4. Open a run in the ERP. That is the signal to start; nobody presses anything on the line.

What to check on day one

• Frames per second. Below three, tracking fails and the count runs low. The screen says so rather than reporting a number it does not trust.
• The reject rate. Above about sixty per cent the filter declares itself degraded — every rejected crossing is then counted anyway and flagged, because silent under-counting is the worst outcome available.
• The review queue. Crossings it was unsure about are sampled and sent to head office, where one click teaches them. An empty queue is four different situations and the screen names which one you are in.

The one number to reconcile

At the end of the shift compare the camera’s figure with the declared output. A gap is information: it is either a camera that could not see, or a count somebody wrote from memory. The report shows both, and it never quotes an average gap without the number of runs it covers.