简体中文界面。帖子、指南及政策正文保留原文；未对原文作自动翻译。
LeoBell0623 · 2026年7月2日 06:25 UTC
Are the faceplates retained and linked to their images? Quality needs to examine the actual marked regions against the appearance requirement before those captions become reference decisions.
5分Keep the rejected-caption history when the labels change. I would want the dataset owner to know which physical decision replaced a guess, not simply receive a cleaner folder and assume all the old evidence agreed.
7分Liam, retained and numbered. Carla, the original captions are preserved as unverified, not approved defects.
9分For a teaching comparison I would include a confirmed defect as well as a cleared reflection. Removing false labels can improve the reference set without showing that the camera can detect the defects it must find.
9分How repeatable is the seating itself, Leo? Written settings are useful, but I wonder whether the part returns to them after removal, rather than only staying put through one image session.
3分Jane's question could explain some image variation, but it does not replace the physical classification. I'd pursue both: agreed reference regions and a measured repeatability check of how they're presented.
16分Quality has classified a shallow groove as reject and a bright brushed region as acceptable. Repeated seating still shifts their images.
18分Are those shifts associated with observed seating changes, Leo, or only image movement? Do not declare the holder faulty from the camera result alone, even if that is now the leading suspicion.
5分Tooling observed a side-stop pin missing from the holder's specified assembly; it isn't only an image inference.
17分That is a useful physical finding. I'd have the holder restored and its repeatability checked, then compare the same reference regions; otherwise an improved image could hide a simultaneous change to the lighting or labels.
4分Keep the earlier frames with the incomplete holder identified. They may still explain the false labels, but they should not be mixed into a supposedly controlled repeatability set.
9分Holder restored and reseating checked by tooling. Same labelled regions, same light settings; the repeat images now align consistently.
10分How do the accepted region and rejected groove appear across those repeats? Consistent positioning is progress, but an aligned image is not by itself an adequate inspection reference.
21分Both regions remain visible; their brightness ranges overlap. Presentation improved, but our proposed brightness threshold still cannot separate these references.
4分