GPU Acceleration || Face Naming

[olegred]

I am impressed with the work you've done. I installed it last night on my Linux server, and everything worked smoothly, even though I'm just an average user. However, I have a couple of questions. Is there a way to leverage the GPU to handle objects or conditions instead of relying on the CPU?

Secondly, is it possible to name faces, similar to how Nest cameras can register and identify faces? I would like the system to recognize and remember faces, enabling me to, for example, view all photographs of my son. I'm not aware of an easy way to do this. Thank you again for your excellent work.

Is it possible to train the system to identify similar images based on a specific photograph? For instance, can the system find all similar images to a sunset photograph and tag them as such? I apologize if my questions seem awkward, as I am not a technical expert.

Alex

[xemle]

Hi @olegred

thank you very much for your flowers. I am happy that you appreciate the gallery and the work.

Is there a way to leverage the GPU to handle objects or conditions instead of relying on the CPU?

Yes and no. IMHO the extraction is a one time task and does not need to scale. Since you are self host the gallery my assumption is that the full gallery does not need to be available immediately.

The API server - which does the AI stuff such as face extraction and similarity embeddings - can run TensorflowJS with a GPU backend using docker compose. So you can try to customize your docker compose setup.

But as I said I do not pay much attention to the performance of it. It is good enough for me. Newer AI capabilities would be more interesting to provide further integrations than polishing the current functionality. Something like ollama...

And since it is an open source software you can try to improve it by your own. If you have further questions to it, let me know.

Secondly, is it possible to name faces, similar to how Nest cameras can register and identify faces?

Not at the moment.

There are previous work from others but it is in progress. Well, more in a stale mode. Continue poking at #132.

Is it possible to train the system to identify similar images based on a specific photograph?

Yes. You can select your seed picture, multi tag best matching images and assign a tag to them.

Please have a look to the documentation for further details.