https://bugs.kde.org/show_bug.cgi?id=525972
H <[email protected]> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |[email protected] --- Comment #5 from H <[email protected]> --- I can reproduce this problem as well with digiKam 9.1.0 on Windows, and I believe it has a particularly serious consequence for Face Recognition. My setup: * digiKam 9.1.0 * Microsoft Windows * Internal MySQL database * Large existing photo collection; photos are not moved or renamed by digiKam * Face Recognition settings: Face size = Small, Detection accuracy = 10, Identification accuracy = 8, all CPU cores enabled * Face Tags including face regions are enabled * Lazy metadata synchronization is enabled * I also tested with “Rescan files when modified” disabled; the problem still occurs. I already have confirmed face data in the database. For example, one person (“Venci”) has about 110 faces which are displayed with a face rectangle and “Venci [Accepted]”. The problem can be reproduced as follows: 1. Go to People → Unknown. 2. Select a face which is clearly a known person, for example Venci. 3. Enter the person's name and click the green Confirm button. 4. The face disappears from Unknown. 5. However, it does NOT become a confirmed face under that person. 6. Instead, it appears under that person's Unconfirmed faces. 7. If I return to the Unknown view, the same face appears there again. 8. Rejecting or ignoring faces shows the same general problem: the face disappears temporarily but then reappears. There is also a metadata synchronization symptom: * After confirming a face, digiKam puts the image into the pending metadata synchronization queue. * If I synchronize from database → image metadata, the face can return to the Unconfirmed/Unknown state. * However, the problem is not limited to synchronization: even without performing the synchronization, leaving the view and returning can show the face as Unconfirmed/Unknown again. This is especially serious for Face Recognition because newly recognized faces cannot reliably become confirmed training samples. For example, with my collection: * I had about 880 Unknown faces. * Running Face Recognition produced useful results. * At identification accuracy 7, about 37 new suggestions were produced. * At 6, about 76 suggestions were produced, but most were false positives. * At 8, about 34 suggestions were produced and the ones I checked were mostly correct. So the recognition engine itself is producing useful suggestions. The problem is that I cannot reliably confirm those correct suggestions and have them remain confirmed. This prevents the normal iterative training process: I can identify a face correctly, but I cannot permanently add that new correct example to the person's confirmed training set. Therefore the recognition model cannot progressively improve from additional verified examples. There is another related symptom: after rebuilding all training data, the “Recognize faces only” workflow sometimes reports: “There are no recognizable persons. Please add new persons.” This happens even though the database contains persons with many confirmed faces (for example, Venci with about 110 accepted faces). I believe the Unknown/Unconfirmed editing problem should therefore be investigated not only as a UI refresh issue, but also as a face-tag/database persistence problem, because it prevents confirmed recognition results from becoming stable training data. I would be happy to test a fixed version and provide additional reproduction information if needed. -- You are receiving this mail because: You are watching all bug changes.
