The hardest vision problems are often not the model: optics, lighting, line speed, product variation and what happens after a detection all matter.
A good-looking score can hide a failed inspection
Imagine a production run of 10,000 parts: 9,900 are conforming and 100 are defective. A useless inspection rule that labels every part “good” reports 99% overall accuracy. It also misses all 100 defects. These are deliberately hypothetical numbers, not results from any named vision product. The example demonstrates a mathematical weakness of aggregate accuracy when the defect class is rare. A supplier deck can report a technically correct number that does not answer the plant’s real question: which defects did we miss, under which conditions, and what happened next?
Separate escape risk from reject burden
False accepts allow defective parts through; false rejects divert good parts. Those are not interchangeable errors. For each critical defect category, request a confusion matrix and the definitions of “good,” “defect,” “review” and “unclassified.” Specify the denominator and confidence interval, not only a percentage. If 1% of the 9,900 good parts in the hypothetical run were falsely rejected, 99 good parts would need unnecessary handling. That may be tolerable for a low-volume process and costly at another factory. The result depends on the agreed decision threshold and the real cost of both kinds of error.
A camera demonstration is not a line-acceptance test
Cognex documents integrated vision systems such as In-Sight 2800, while Basler describes cameras, optics and lighting as a combined image-formation problem. Their public documentation establishes product scope, not head-to-head inspection performance. A useful acceptance plan must test the actual optics, mounting, exposure, illumination, triggering, line speed, material batches, product changes and shift conditions. A model cannot evaluate a defect that the image-acquisition system fails to reveal. No general claim that one supplier has lower escape risk follows from those documents.
Test the consequences of uncertainty
Ask what happens to low-confidence or unrecognized images. Is the part rejected, quarantined, allowed through, or sent to review? How does the cell behave when a camera is misaligned, a lighting source degrades, the network connection drops or a recipe changes? Record the specific software or model revision associated with the inspected batch. Confirm how the inspection result reaches a PLC, robot, quality record or operator—and how a known-good configuration is recovered after an unsuccessful change.
Turn the score into an auditable buying decision
Before comparing proposals, give each vendor the same defect taxonomy, representative image set, throughput requirement, false-accept limit, false-reject budget and sample-handling rules. Require a documented acceptance protocol on production hardware, with agreed responsibilities for retesting and ongoing drift monitoring. A buyer should prefer defensible evidence over a bigger headline percentage. This is an engineering selection checklist; this site has not independently benchmarked Cognex, Basler or any other supplier.
Practical takeaway
If a machine-vision proposal leads with “99% accurate,” the next questions should be: 99% of what, on which population, at what operating conditions, with how many critical escapes, and what production decision did the result authorize? An inspection score is not a production disposition.
Make the next vendor conversation more productive.
Use the free project fit check to generate an initial list of questions for your team.
Open planning tool ↗Primary sources and limitations
These links support the public product or research context. They do not substantiate comparative performance conclusions. Source content may change; verify directly with the supplier.