Gallery icon indicating copy to clipboard operation
Gallery copied to clipboard

[BUG] Classification pre-trained model can make use of some serious refinement...

Open derei opened this issue 1 year ago • 10 comments

It's quite obvious from the attachments what I mean. The worst is to not recognise Enterprise (blasphemy 😆).

SmartSelect_20241208_114926_Gallery SmartSelect_20241208_114914_Gallery SmartSelect_20241208_114859_Gallery

derei avatar Dec 08 '24 11:12 derei

This is the best I could find - which is the same as the Google Gallery Go uses.

The next update will be a major one where it would allow for manually creating and moving categories for the ones that do not match.

I'm not in any way an AI/Machine Learning expert nor I have any significant experience with it, but from my testing this is the best I could find that categories the media in a more general manner while being fast and efficient - which as you can see can definitely sacrifice the accuracy.

I have tried over 15 models - some were too specific (and very often wrong).

IacobIonut01 avatar Dec 08 '24 11:12 IacobIonut01

Also creating my own model is not a priority right now

IacobIonut01 avatar Dec 08 '24 11:12 IacobIonut01

@IacobIonut01 Just out of curiosity, what model are you using right now?

mon-jai avatar Dec 10 '24 18:12 mon-jai

@IacobIonut01 Just out of curiosity, what model are you using right now?

Mobile Ica 8bit with metadata

IacobIonut01 avatar Dec 10 '24 19:12 IacobIonut01

Have Ente's ML models been looked at? They work very well in my experience. You can search for a keyword and it will bring up all the image related to said keyword. It groups faces very well too.

Relevant links: Ente's GitHub: https://github.com/ente-io/ente Blog post: https://ente.io/blog/machine-learning/ Machine Learning Whitepaper: https://ente.io/ml

I could open a feature request in a seperate issue if you're interested. Good ML models really are the main reason I'm using Ente at the moment so this would be very beneficial.

ghost avatar Dec 20 '24 22:12 ghost

@IacobIonut01 So this is entirely loca? I was surprised that it immediately started classifying my images.

newhinton avatar Dec 22 '24 16:12 newhinton

@IacobIonut01 So this is entirely loca? I was surprised that it immediately started classifying my images.

Yes, it uses Tensorflow Lite to process them in the background locally

IacobIonut01 avatar Dec 22 '24 16:12 IacobIonut01

There are two major anachronistic dealbreakers here:

  1. You generate hundreds of useless categories, and I can't even search for them. So what's the point? There's also a lot of synonym categories like airplane aircraft aviation aerospace (btw 0 of the images actually have anything to do with any of those words)

  2. The accuracy is trash vs QuicPic or Tidy. I get one (1) swimwear image in Gallery, and and in picquery or tidy I have hundreds (accurate).

This will sound crass but, 'why don't you just use their code, it's right there and it's worked for years': https://github.com/greyovo/PicQuery https://github.com/slavabarkov/tidy

More importantly, you can actually SEARCH for the content instead of scroll through hundreds of uncalled-for categories spam. And search is instant.

tdbe avatar Mar 27 '25 21:03 tdbe

There are two major anachronistic dealbreakers here:

  1. You generate hundreds of useless categories, and I can't even search for them. So what's the point? There's also a lot of synonym categories like airplane aircraft aviation aerospace (btw 0 of the images actually have anything to do with any of those words)

  2. The accuracy is trash vs QuicPic or Tidy. I get one (1) swimwear image in Gallery, and and in picquery or tidy I have hundreds (accurate).

This will sound crass but, 'why don't you just use their code, it's right there and it's worked for years': https://github.com/greyovo/PicQuery https://github.com/slavabarkov/tidy

More importantly, you can actually SEARCH for the content instead of scroll through hundreds of uncalled-for categories spam. And search is instant.

Thanks for the suggestions.

"Why don't you just use their code?" - Simple, didn't know about those projects, but now I do thanks to you!

IacobIonut01 avatar Mar 27 '25 23:03 IacobIonut01

There are two major anachronistic dealbreakers here:

  1. You generate hundreds of useless categories, and I can't even search for them. So what's the point? There's also a lot of synonym categories like airplane aircraft aviation aerospace (btw 0 of the images actually have anything to do with any of those words)
  2. The accuracy is trash vs QuicPic or Tidy. I get one (1) swimwear image in Gallery, and and in picquery or tidy I have hundreds (accurate).

This will sound crass but, 'why don't you just use their code, it's right there and it's worked for years': https://github.com/greyovo/PicQuery https://github.com/slavabarkov/tidy

More importantly, you can actually SEARCH for the content instead of scroll through hundreds of uncalled-for categories spam. And search is instant.

Expect an update regarding this in the new Search overhaul update.

Currently I'm working on other features, but the Search overhaul will get prioritise right after that

IacobIonut01 avatar Mar 28 '25 10:03 IacobIonut01