SUMMARY I've stated to play around with automatic tagging using the maintenance tools and the result are underwhelming. I have collection of 100K+ photos and most keywords assigned are way off, specially when compared to the job done by Photoprism or Immich. I have tested with all available models and , while the level of detail found in the tags will change, the accuracy remains very poor in all cases. STEPS TO REPRODUCE 1. Maintenance tools 2. Auto-tag assignment 3. OBSERVED RESULT Tagging starts as expected and take a looooong time to complete. Tags are nested under the 'auto' tag. Checking the photos found under each tag using the filter option, most of the photos are completely unrelated to the word described by the tag. EXPECTED RESULT The same, but with better accuracy. I don't expect to select pictures tagged as 'airplane' and find shots of people walking on the streets. SOFTWARE/OS VERSIONS Windows: 10 macOS: (available in the Info Center app, or by running `kinfo` in a terminal window) Linux/KDE Plasma: KDE Plasma Version: KDE Frameworks Version: Qt Version: ADDITIONAL INFORMATION
Which DNN model did you use while your test ?
Hello, Yes, the current AI models for object detection for automatic tagging need to be updated. There are newer models that give better results. We're focusing on improving face detection and recognition in this release. We may be able to add some new auto-tagging AI models in the 8.5.0 release, but it may have to wait until 8.6.0. Please know we also think that auto-tagging could be better, and it's on our list of improvements. As Gilles asked, what are you using now? Cheers, Mike
(In reply to Michael Miller from comment #2) > Hello, > Yes, the current AI models for object detection for automatic tagging need > to be updated. There are newer models that give better results. We're > focusing on improving face detection and recognition in this release. We > may be able to add some new auto-tagging AI models in the 8.5.0 release, but > it may have to wait until 8.6.0. Please know we also think that > auto-tagging could be better, and it's on our list of improvements. > > As Gilles asked, what are you using now? > > Cheers, > Mike Thank you for your quick response. If you mean with DigiKam, I have tested with YOLO Nano v5 and ResNet50. If you mean to what I am comparing Digikam to, that is Photoprism,which adopts NASNet Mobile 224 , I think. Photoprism is way more conservative but yeld more reliable results overall. Both aplications are connected to the same set of photos.
Thank you. I'm going to keep this ticket open to remind me to look at NASNet when I start updating the auto-tagging recognition models. Cheers, Mike
*** Bug 496408 has been marked as a duplicate of this bug. ***
(In reply to cvinhaes from comment #3) > > Thank you for your quick response. If you mean with DigiKam, I have tested > with YOLO Nano v5 and ResNet50. If you mean to what I am comparing Digikam > to, that is Photoprism,which adopts NASNet Mobile 224 , I think. Photoprism > is way more conservative but yeld more reliable results overall. Both > aplications are connected to the same set of photos. I want to let you know I haven't forgotten about this ticket. I'm rewriting the auto-tagging pipelines now. I've tested several models including multiple versions of NasNet, and it seems YOLO v11 is the most accurate. It will probably be a few more weeks until the new code and models are merged into the daily builds. There will be an update to this bug entry when the new code is available if you'd like to test. Cheers, Mike
(In reply to Michael Miller from comment #6) > (In reply to cvinhaes from comment #3) > > > > Thank you for your quick response. If you mean with DigiKam, I have tested > > with YOLO Nano v5 and ResNet50. If you mean to what I am comparing Digikam > > to, that is Photoprism,which adopts NASNet Mobile 224 , I think. Photoprism > > is way more conservative but yeld more reliable results overall. Both > > aplications are connected to the same set of photos. > > I want to let you know I haven't forgotten about this ticket. I'm rewriting > the auto-tagging pipelines now. > > I've tested several models including multiple versions of NasNet, and it > seems YOLO v11 is the most accurate. It will probably be a few more weeks > until the new code and models are merged into the daily builds. There will > be an update to this bug entry when the new code is available if you'd like > to test. > > Cheers, > Mike Thank you for the updates, mate.
>Thank you for the updates, mate. The latest autotags code is in the nightly builds if you want to test it. Cheers, MIke
(In reply to Michael Miller from comment #8) > >Thank you for the updates, mate. > > The latest autotags code is in the nightly builds if you want to test it. > > Cheers, > MIke I'm afraid I was not able to find the nightly builds and it seems that the change was not yet incorporated to 8.5.0. Could you please point me in the right direction?
(In reply to cvinhaes from comment #9) > (In reply to Michael Miller from comment #8) > > >Thank you for the updates, mate. > > > > The latest autotags code is in the nightly builds if you want to test it. > > > > Cheers, > > MIke > > I'm afraid I was not able to find the nightly builds and it seems that the > change was not yet incorporated to 8.5.0. Could you please point me in the > right direction? The you can get the pre-release 8.6.0 version here: https://files.kde.org/digikam/ Cheers, Mike
There was a bug in the auto tags feature, especially when writing metadata to images. Waiting for packages with a date greater than March 12, 2025. Maik
Hi Maik, Windows build will be done this morning... Gilles
(In reply to caulier.gilles from comment #12) > Hi Maik, > > Windows build will be done this morning... > > Gilles I've tested the 8.6.0 release for Windows and the progress with both auto-tag and face detection is remarkable! Even though auto-tag is not 100% reliable in terms of finding their subjects in every photo (and it is not expected to be at this stage of the tech) the accuracy is much improved and on par, if not better than other tools. I've tested in DigiKam mostly with EfficientNet B7. Congratulations for the great work done.
> I've tested the 8.6.0 release for Windows and the progress with both > auto-tag and face detection is remarkable! Even though auto-tag is not 100% > reliable in terms of finding their subjects in every photo (and it is not > expected to be at this stage of the tech) the accuracy is much improved and > on par, if not better than other tools. I've tested in DigiKam mostly with > EfficientNet B7. Congratulations for the great work done. Thank you for the kind words. The team worked hard on the updates and refinements for 8.6.0. It's nice to know you're happy with the changes. Help us spread the word about digiKam. We think the improvements in the past few releases really puts digiKam at the top of the digital asset management application list. Cheers, Mike