I’m Tobi, new in this community and four weeks ago I switched from Lightroom to darktable. There are multiple reasons but I would say the main reason is that darktable is Open Source and runs on Linux.
I’m very happy with darktable and thanks to the great videos of @Bruce_Williams I could find my way around very quickly.
Anyway, mainly I work as researcher/developer and because of this, I started working on an auto-tagging plugin in Lua for darktable, which I would like to present here to get feedback. It uses a small command-line application, called photils-cli, to extract features from an image. These features will be used to detect keywords/tags (at no time your image leaves your device, only a representation).
I would be very happy if it is tested by some of you and you could give me some feedback.
BTW I didn’t know about photils but I like this “… The usage of the feature vector has this advantage that at no time your selected image will leave the PC.”
From privacy-wise is really good and for bandwith too.
I tried to install it, but the instructions are not clear to me. Where should I put the photils AppImage? It needs to be found by the lua plugin I guess? The darktable lua instructions were not clear. I did put the lua script in the darktable lua/contrib directory, and added ‘require “contrib/photils”’ to the luarc file. But nothing shows up in darktable under tagging.
I remember someone did a similar Lua script for face recognition + tagging with Tensor Flow. I wonder if we shouldn’t ship these officially at some point, at the condition that the machine learning is done locally and not outsourced to a server/third party.
@scheckmedia That’s awesome. Would it be possible to include support for hierarchical tags ?
I think I can implement something in the photils-cli application but this will blow up the file size to ~500 MB. The reason is that I have to provide the reference features and the corresponding tags what causes the file size.
From that side the hosted approach is more resource friendly. But sure I can think about some optional flag to provide an offline option.
I think this is possible as well. If someone can provide me some hierarchical structure I can improve the tagging in the near future.
I should mention that the machine learning part is done locally! It uses TensorFlow lite for the feature extraction. Only the request to find images and tags with the same feature of a given image is done online.
Looks like a useful script, but I’ve tried to install it and the get tags panel doesn’t appear after copying the two files to the suggested folder in the ~/.confi/darktable directory. Hopefully you can clarify the installation. Pretty sure lua is working, I use the export to GIMP script frequently.
Update: have installed the lua and linked it to the executable, but nothing happens when I selec an image and hit “Get Tags” . Is the location of the AppImage important?
Thanks for your script, which looks really nice !!
I installed it without issues, but I have an issue when running it … it seems the script is allergic to filepathes containing spaces !
Do you confirm ?
Do you want/need me to book an issue on GH ?
Thanks again !
M_LL_G
Thanks for this, meanwhile I solved the problem and I will update the script. I’ve not used quotation marks while I passed arguments to photils-cli. This works until spaces appear.
Thanks for the work.
I’m totally new to neural networks, so how does it learn and show more specific tags the next run? From my local tags?
I selected an image from Sugar Loaf mountain, a more a less iconic mountain, and it didn’t show any tag related to this country (Brazil).
Here’s the image: