Utility inspectors need one defect type named on pole photos. The vision service still cannot label that type after the team supplied examples, and a stable set of labeled photos already exists. Which approach fits?
Select an answer to reveal the explanation.
Short Explanation
When the stock catalog has no card for the part, and you already have a labeled drawer of them, you make the card. You train a custom model on the labeled photos because the vision service still cannot name this defect. Waiting on the same service repeats the miss.
Full Explanation
A custom model fits when the tools you already have cannot produce the behavior and the labeled data is stable. The behavior is naming one defect type. The vision service still cannot label it after examples, and the utility already has a stable set of labeled photos, so training a custom model on those photos is the justified path. Keeping the same photos on the vision service repeats a call that has already failed to emit the label. Asking inspectors to stop photographing the defect hides the condition the business needs to see, and a chat model that invents a label is not a vision judgment grounded on the photo. A general article about poles is text, so it cannot mark this defect on an image, and it discards the labeled photos that make a custom model possible. Exam caveat: examples that fail are the evidence the prebuilt skill is the wrong tool. They are not a reason to invent a label with no image. Operational check: hold out a few labeled photos, train aside from that holdout, and confirm the custom model names this defect type on the holdout before it is wired to the inspector's flow.