The FR10 trial folder calls a bright streak a scratch, although it moves between views of the same polished cover. The cover was returned to the production pile. I have asked people to stop using those labels until we have physical references we can identify and keep.
Leave those images as unverified examples. Build the comparison from identified accepted and rejected covers inspected under the agreed appearance method, with the imaging setup recorded.
Do you still have the original image files? Our supposed scratch changed character when we opened the source instead of the presentation copy. Still needed the part, but it explained why people were arguing.
Oscar, that can explain a display difference, but without the original cover it cannot resolve its physical condition. Callum should not retrospectively classify it from whichever image now looks more convincing.
Oscar, source files exist. They include the different presentations, not just crops. Henry, we are leaving that cover unclassified. Quality has retained new examples but disagrees about one borderline finish.
Keep the borderline finish in an undecided group until quality settles it. Don't force a binary label just because the software folder wants two directories.
Our training examples once omitted the difficult finish entirely. The test looked excellent. Include it once its judgement is agreed, not only immaculate parts and obvious damage.
Imran, agreed, with the intended range of surface and presentation variation represented. The sample selection needs to match what production will actually send to the station.
Callum, can the inspector and vision person look at the borderline cover together? Might be cheaper than another week of passing photographs around and each imagining a different mark.
They did. The disputed line is acceptable under our finish method; a separate shallow scratch is not. Both retained. Imaging comparison is still open because the scratch is weak in the proposed view.