VISIONROUTE

Computer-vision feasibility

Will image recognition actually work in your product?

You can answer that before committing an engineering quarter. Feasibility testing works in three stages: a plausibility check from your product context (free, minutes), an architecture and evaluation design for your exact use case, and a controlled test on a few hundred representative images with agreed acceptance criteria. The result is one of four explicit verdicts — including “not enough evidence” — rather than a demo-driven yes.

Definition

What feasibility testing is

Computer-vision feasibility testing determines whether any image-recognition approach can meet a product’s specific accuracy, latency, and cost requirements — before the product commits engineering effort to building it. It differs from a demo in one decisive way: it is scored against your definition of correct on your representative data, not on examples chosen to look good.

Why demos mislead

A demo shows the happy path: clean images, obvious cases, no negatives. Production sends blur, glare, edge cases, and inputs where the correct answer is “reject”. Feasibility lives or dies on exactly the cases a demo never shows — which is why the test set matters more than the model.

The unique asset

The four feasibility verdicts

A future proof run should end in one of these verdicts. The table is illustrative of the format — the verdict for your product can only come from testing your data.

Verdict format — illustrative; real verdicts cite the tests behind them.
VerdictWhat it meansWhat you do next
Feasible nowAt least one tested stack met your acceptance criteria on representative data, within latency and cost constraints, under recorded configurations.Ship the winning stack with its confidence thresholds and review path; keep the evaluation as a regression suite.
Feasible with constraintsA stack met the criteria only under conditions — e.g. above a confidence threshold, on a subset of input types, or with human review of low-confidence cases.Ship the constrained version; route the uncovered slice to people; expand the test set where the constraint binds.
Not yet feasibleNo tested stack met the criteria. The gap is specific: which slices failed, by how much, and at what cost point.Decide whether to relax the acceptance criteria, invest in data or fine-tuning, or defer the feature — with evidence for the decision either way.
Not enough evidenceThe test set was too small, too clean, or not representative enough to support any of the above with confidence.Collect the specific data the verdict names — usually hard cases and negatives — and rerun. No conclusion is forced.

How to run it

Feasibility in five steps

  1. State the decision, not the technology

    Write the visual decision in product terms: what images arrive, what must be determined, what the product does with the answer, and what a wrong answer costs. Feasibility is a property of this decision, not of “AI” in general.
  2. Check plausibility for free

    A free App Scan of your public URL frames the use case and identifies the plausible solution families and their known risks — in minutes, before anyone collects data.
  3. Define correct before testing

    Agree the expected answer for every test input and the acceptance criteria that decide pass or fail. If two people on your team disagree about what counts as correct, the model has no chance.
  4. Assemble a representative slice of reality

    A few hundred images that mirror production — common cases, rare-but-costly cases, negatives, and quality variation — beat ten thousand clean demo images. Include the inputs you are afraid of.
  5. Test candidates under identical conditions

    Run each viable stack on the same data with the same scoring, record every configuration, and read results by slice. The output is one of the four verdicts above — including, honestly, the fourth.

Limitations

What feasibility testing cannot tell you

  • It cannot certify inputs unlike the test set — a verdict is bounded by the data behind it, and the report states those bounds.
  • It cannot stay true forever: models, providers, and your own product drift. That is why the test set becomes a rerunnable regression suite.
  • It cannot rescue an undefined problem. If “correct” has no agreed definition, the honest first step is the definition, not the test.

In short

Computer-vision feasibility is testable before you build: define the decision and what correct means, test candidate stacks on a few hundred representative images under identical conditions, and accept one of four verdicts — feasible now, feasible with constraints, not yet feasible, or not enough evidence.

Related: how to test vision models properly · the twelve-item feasibility checklist · the production checklist · example free report