What a computer-vision shelf POC costs, and how long it takes
A computer-vision POC to measure shelf-out on one section of one store takes 6 to 10 weeks, if ground truth exists. Without it, the POC delivers a pretty video and no criterion, and dies at the moment of scaling.
Most vision POCs start with the right hypothesis: a camera can see the shelf better than a manual audit. What the team does not measure before starting is whether it can prove the camera saw correctly.
The realistic range
6 to 10 weeks, one section, one store, a small team. The range assumes cameras already installed or simple to install, and a section with visually distinct items.
The cost is not in the model. It is in labeling the images and the human verification that creates ground truth. That is the work nobody puts in the proposal and everybody pays for.
What “data in reasonable shape” means here
This is where the POC lives or dies. Reasonable means shelf images with a reliable label of what was there, at the moment, verified by a human. Without it, there is nothing to compare the camera’s reading against.
Lookalike items, packaging that changes every promotion, bad aisle lighting: each one raises the labeling cost. A general-merchandise section is harder than a beverage one.
The phases
Diagnóstico: 1 to 2 weeks, to choose the section and define the hit criterion. POC implementation: 4 to 6 weeks, model reading the shelf, compared against ground truth. Decision: 1 to 2 weeks, read the result and decide to scale or stop.
The POC does not deliver an operation. It delivers an honest number: does the camera hit accurately enough, in this section, to justify the next store?
What moves the timeline by 50%
Three decisions, and the third is the one that kills most.
- How many sections at once. One proves it; three become an ownerless project.
- Quality and quantity of available ground truth.
- Whether the POC has a written acceptance criterion, or “it looked good” will be the verdict. The POC with no criterion is the one that works in the room and dies in the operation.
The week-one criterion
The number that decides is not the model’s accuracy in the lab. It is its accuracy in your worst aisle condition, compared against the cost of the manual audit it would replace. And OSA has to be measured at the shelf, not in the system.
Take the vision POC on your desk. In ten seconds, name the three numbers that would tell you whether to scale it. If you stalled, the POC will deliver enthusiasm and not a decision.
Send me the three candidate indicators your POC will measure. In one hour I will send back which predict scale and which predict a pretty pilot that never becomes a store. If the shape confirms, the two-week Diagnóstico ends with a one-page document: three indicators, three ranges, three triggers, that the squad holds for the rest of the POC.