RUBIKSCORPAI in factories
Why camera-based counting fails on the factory floor
Lighting changes, overlapping items, vibration and camera angle all confuse vision models trained on clean samples. A count that looks accurate in a demo can drift badly across shifts.
Test the counting method on the actual line
Define what a count means before collecting footage. Does an item count when it enters a region, crosses a line or leaves a station? Decide how rework, reversals, stoppages and overlapping items should be handled. Review the camera position, field of view, exposure and lighting. Counting often depends on tracking the same item through time, so a good detector alone may not solve the task.
What to check
- Collect permitted samples across shifts, speeds, product types and lighting conditions.
- Include crowded scenes, empty frames, reverse motion and partially hidden items.
- Compare the reference counts from independent reviewers and resolve disagreements in the counting rules.
What to change
Improve camera placement and lighting before compensating with a more complex model. Set a clear counting boundary and evaluate tracking across missed detections. Run a pilot with manual reference counts. Log uncertain intervals for review and show operators when the camera or data feed is unavailable instead of displaying a misleading zero.
How to evaluate the result
Compare overcounts and undercounts separately, along with cumulative error over a shift. Check results by operating condition. A good average can hide a consistent failure on one product or at peak line speed.
What the result does not establish
Performance can change after a camera moves or the line setup changes. Document the tested conditions and repeat acceptance checks after material changes.
Further reading
Want us to fix this for you?
Getting it right starts with the real conditions on your line. Register and we’ll help you scope a pilot that is tested where it will actually run.