Auto-tag models are poor

Help
1

Hi,

I have tried to use Auto-tagging for my photos, but it seems so pure…
And 99% tags are bad and not accurate. I have tried 3 model that we have in latest digikam and all are very poor…

In comparison I have tried some online free services in the internet, and they give me nice bunch of tags, 95% of them are accurate and useful !

Could you plz update the models? Or add some good one or maybe there are some secrets configuration for use of existing models to have nice result?

Tnx in advance

1 Like
2

Hi @ArtemOstapov ,
We updated the models for digiKam 8.6.0. They are the best ones we’ve found that run locally on your computer. Yes, there are better models, but they require you to share your images. They also require more compute power than what is typically found on a personal computer.

I’m happy to look at other models. Please let me know what you suggest.

Cheers,
Mike

3

for me I cant use it as I want, cause for me it generates small amount and invalid / bad tags.
Mb for someone else it is good?

For me is better to spend x10 time for each photo for autotaging in comparison not to have auto tags at all )

I just found one example, and it seems produce nice bunch of good tags:

Do you need model for C-language? Python? or any Language?

Do you have any constraints in size etc?
If model huge - mb do it as additional downloaded content

4

Hi @ArtemOstapov ,
Thank you for sharing the repo. That looks very interesting.

digiKam uses c/c++, and there isn’t a way to include Python code. It’s still possible to convert to code to c++, though. Also, the repo uses pillow for loading the image, and digiKam uses OpenCV.
The biggest issue is the speed of the model. Based on the screenshots in the repo, average processing time is over 4 seconds per image. I also have to assume that the demo screenshots were created on a machine with a fairly new GPU. On older machines this means it could be almost double the amount of time to process each image. Even a medium library like mine (~25,000 images) it would take almost a 1.5 days to process all the images at 4 seconds/image, and on larger libraries it could take almost a week.

I’m going to continue to look at this and try to come up with some way to include this in digiKam. I love the idea and really like the tags created by the model. I think this could be very useful. I just need to figure out a way to do it so it fits in the digiKam structure.

Cheers,
Mike

1 Like
5

if the model is hard, I think nice idea is to add it as additional one to digikam and lets user to choose newHard/oldQuick.
As for me, (4sec per image) I m gonna to use it for some folders each by each )

Do you have any stat about using current auto-tag models ?
Does anyone actually use current models ?

6

Hi @ArtemOstapov,
digiKam does not collect any telemetry from the end user, so the only way we know people are using the existing models is through feedback. I do know there are many users who run the existing models to get tags. I don’t know how many, or what the percentage of total digiKam users is.

Cheers,
Mike

1 Like
7

Hi @ArtemOstapov,
I want to thank you for letting me know about ImageIndexer. It’s a fabulous tool. I’m talking with the rest of the digiKam team about we can incorporate this functionality into digiKam. I highly doubt it will make it into the 8.7.0 release, which should be 30-45 days away, but maybe we can have something solid for 8.8.0, which would be in the autumn.

I already have an experimental pipeline developed, and I’ll begin testing it tonight.

Cheers,
Mike

3 Likes
8

wow, tnx for such good info!