Dat-inf recipe for success: computer vision & software
We have summarized our experience from many successfully completed projects here.Define the goal clearly and concisely
What exactly should be detected or measured, under which conditions, within what maximum analysis time—and what should happen with the result?
Make success measurable with numbers
- Sensitivity and specificity: How often is a good part incorrectly classified as bad? How often is a bad part incorrectly classified as good?
- Statistical measures (e.g., correlation metrics for scores or continuous target values)
- Timing requirement: How quickly must the result be available at the latest (e.g., in the 95th percentile case)?
Stabilize image acquisition before optimizing software
Design lighting, camera, optics, positioning and triggering so that images look as consistent as possible. Stable acquisition often reduces computer vision effort dramatically.
Define reference data early and don’t keep changing it
A fixed, versioned dataset with typical good parts, real defective parts and borderline cases. This dataset ultimately decides “pass” vs. “fail”.
Start with the simplest solution and measure consistently
Try simple, robust methods first. Only switch to more complex approaches if the metrics are not sufficient. Important: always measure changes against the same reference dataset.
Don’t force borderline cases—handle them cleanly
For uncertain cases, define a clear outcome: “manual review” or “separate reject”, rather than guessing.
Plan pilot operation as a mandatory step
First run in monitoring mode and log results, then gradually output real decisions. Verify whether the metrics remain stable over time (material batches, wear, environmental conditions).
Make the handover operationally reliable
Provide a short guide for start/stop, calibration, typical issues, logging and responsibilities. Changes to variants or criteria should follow a simple change process.
For small projects
If you consistently implement just three things, that is often enough in practice to reliably deliver small custom projects successfully:
- stable image acquisition
- fixed acceptance data
- a few clear metrics