Three ways to read a paper target, measured against the same ground truth
Problem
A phone photo of a paper target is skewed, shadowed and often reused, with patch tape and old holes showing through. Classic blob detection, a vision-language model and a trained detector can all be made to report holes. Which one to trust is not a matter of opinion, but nothing settles it until each is scored against the same truth on the same photos.
Action
Every photo is warped into one canonical frame, so scoring is geometry. A detector only proposes: a person confirms every target, and both readings are kept side by side. That makes each detector's misses and false alarms countable on identical data, and the confirmed readings become the training set for the next model.
Outcome
On a held-out target marked again from scratch, the trained detector finds 34 of 37 holes, where classic CV finds 23. It pays in false alarms, 10 against none at its default threshold. With a person confirming either way, the gain is not fewer clicks: it is more of the holes found, placed closer on average. One target, one run: a direction, not a benchmark.
The detail worth keeping: the first held-out run read as a clear loss, and the model was fine. It was cut at the stock confidence threshold, tuned for detectors with eighty classes of everyday objects; a hole is one small class. A bad number from a model is a question about its operating point before it is a question about the model.
Discuss your project
Email me with what your system does, what is failing or missing, and your timeline.
Limited slots, and a short discovery call to confirm fit, scope and timelines before anything is agreed.