| Summary: | digikam face recognition ignores imported face tag data | ||
|---|---|---|---|
| Product: | [Applications] digikam | Reporter: | Justin <jr.freester> |
| Component: | Faces-Recognition | Assignee: | Digikam Developers <digikam-bugs-null> |
| Status: | CONFIRMED --- | ||
| Severity: | normal | CC: | caulier.gilles, massimo, metzpinguin, michael_miller |
| Priority: | NOR | ||
| Version First Reported In: | 8.1.0 | ||
| Target Milestone: | --- | ||
| Platform: | Microsoft Windows | ||
| OS: | Microsoft Windows | ||
| Latest Commit: | Version Fixed In: | ||
| Sentry Crash Report: | |||
| Attachments: | attachment-80579-0.html | ||
|
Description
Justin
2023-07-15 01:30:45 UTC
When importing images with face regions, no face matrices are created for face recognition, for performance reasons. In this case you need to use the Maintenance tool (Extra menu) and in the face detection section do the option "Clear training data and rebuild". Based on the confirmed faces in the database, new face matrices are created for all. Maik @Justin, Did you see the previous message from Maik ? Gilles Caulier Created attachment 162362 [details] attachment-80579-0.html Thanks for the reminder. I did try Maik's suggestion; there was some improvement but still not great. I then made a duplicate of my data and cleared any face exit data using exiftool. Starting fresh, the results were much better. The photo collection belonged to my father, who sometimes had 10 copies of the same photo scattered across multiple directories. I got distracted by the duplicates and didn't provide an update. I apologize. I think the accuracy is pretty good. In my opinion, the interface for accepting face tags is not as efficient as Picassa. On Sat, Oct 14, 2023, 11:20 PM <bugzilla_noreply@kde.org> wrote: > https://bugs.kde.org/show_bug.cgi?id=472266 > > --- Comment #2 from caulier.gilles@gmail.com --- > @Justin, > > Did you see the previous message from Maik ? > > Gilles Caulier > > -- > You are receiving this mail because: > You reported the bug. Hi Justin, We working first to integrate 2 new DNN models (YuNet + SFace) in the face management to improve performances of the detection and recognition. The recognition workflow will be changed/fixed for 8.5.0, but it's not yet complete... Questions: - Which extra program did you use uner Windows to tag faces in XMP metadata? - Can you share a file sample to double check? - Did you uses XMP sidecar file? Best Gilles Caulier (In reply to caulier.gilles from comment #4) > Hi Justin, > > We working first to integrate 2 new DNN models (YuNet + SFace) in the face > management to improve performances of the detection and recognition. The > recognition workflow will be changed/fixed for 8.5.0, but it's not yet > complete... > > Questions: > > - Which extra program did you use uner Windows to tag faces in XMP metadata? > - Can you share a file sample to double check? > - Did you uses XMP sidecar file? > > Best > > Gilles Caulier I used Picasa. When I bought a new computer, I began a search for a new photo management software and after trying many different options I found digikam. I believe Picasa has the option to export face region data to xmp. If I still have the old files I will upload a sample. I don't know if the files will be helpful because the face region data may have been edited by the multiple different face tagging software I tried. I have been thinking about face tagging in digikam vs Picasa. Picasa seemed to pre-group similar faces and then assign that grouping to a face tag. Picasa would keep suggesting groups to add to a face tag. It was an efficient way to build up a new database, but I also wonder if this interface hints at how Picasa was handling face detection internally. Digikam's documentation acknowledges that a person's face can vary in appearance for multiple reasons: "One reason can be that there are too many face tags assigned to a person which shows this person in a way that doesn’t really help the search algorithm, e.g. with sunglasses, blurred, unusual colors, carnival make up, dark shaded areas in the face, baby/kid/adult photographs mixed… Another reason to use that option can be false face recognition due to a wrong accuracy value in the Settings tab." What if digikam assigned multiple groups(clusters) to a single person tag? I'm assuming that a detected face is assigned a vector. The Face Accuracy slider determines how similar two vectors must be for the face to be considered the same. Initially the definition of a face results in a tight cluster but as more variations of the same person's face are added the cluster spreads out and may overlap with other people's face clusters. If a person's face is defined by multiple tighter clusters instead of a broad single cluster, the definition of a face wouldn't degrade with increasing number of detected faces assigned to a person. . Hi Justin, You're not too far off. You're off by only 127 :) . Each face is represented by 128 vectors. This applies to both OpenFace and SFace. The new SFace extraction model we're implementing is much more accurate at extracting face features (vectors), so matching is more accurate. We then use a KDTree for KNN classification of the unknown image. I should have something you can test with by the end of the week. I would love for you to give it a try and see what you think if you don't mind running a pre-release version of digiKam. First, I think you'll find the rate of false-positives is much, much lower. It also does a better job of accurately matching fuzzy/blurred/pixelated faces, too. In the future I plan on implementing clustering to further refine the KDTree which will better detect outliers. For the UI, face thumbnail images are grouped typically by the album in which they were found. You can easily accept multiple suggestions at once using shift-click or ctrl-click to select multiple images and then clicking confirm on one image to confirm the group. Hi Justin, Can you confirm this is fixed for you? Cheers, Mike Hi, The 8.7.0 pre-release Windows installer from today have been rebuilt from scratch with Qt 6.8.3, KDE 6.12, OpenCV 4.11 + CUDA support, Exiv2 0.28.5, ExifTool 13.27, ffmpeg 7, all image codecs updated to last version (jxl, avif, heif, aom, etc.). Please try with this version to see if your problem still reproducible... https://files.kde.org/digikam/ Thanks in advance Best regards Gilles Caulier *** Bug 505483 has been marked as a duplicate of this bug. *** Please check my bug reported as duplicate as to me happened I found hundreds of faces as confirmed under one name an nothing under unconfirmed or unknown. I'm now preparing to run a new import from clear data to confirm the issue... but I could not use the rebuild of the matrices as being the wrong faces confirmed it was not possible to train. M |