The operating point is a business decision, and F1 does not know that

A detector emits a score. Turning that score into a decision needs a threshold, and the only rational way to pick one is the cost of being wrong in each direction. Scrapping a good part costs the part and the rework. Letting a defective part reach a customer costs the warranty claim, the field visit, and a share of the next contract. Those two numbers are never equal, and the ratio between them is the whole argument.

Set the costs — these are the facilitator's to change

Cost of a false rejection (good part scrapped or re-inspected)

Cost of a false acceptance (defect reaches the customer)

Parts per batch

Population
include the class held out of training

–––

Cost-optimal threshold

–

Expected cost there

–

per batch

If you had maximised F1

–

Penalty for choosing F1

–

per batch, every batch

Expected cost against operating point

The marked point is the cost minimum. Move the cost of a false acceptance up and watch it slide left; the model has not changed, only what you are willing to pay to avoid a miss.

What happens at the optimum

The uncomfortable reading

Why the population matters more than the curve Untick the population box. On the two texture families the detector was trained on, image-level separation is nearly perfect and almost every threshold costs nothing — a flat, useless curve that would let you claim any operating point you liked. A real cell also sees the third family. An operating point tuned on the training distribution is an operating point tuned on the wrong thing.

precomputed Built by scripts/train_baseline.py at 2026-08-09T19:23:35+00:00. Only the per-image detection scores are precomputed. Every threshold, count and cost on this page is computed during the request, which is why the inputs are editable. Re-run that script to regenerate it; nothing on this page is hardcoded.