Advice on a photo manager with Facial / object recognition and Google Coral Support

[Tomb_of_ash]

Qnap refugee here and one thing that I'm missing is Qmagie app. I have a USB Google Coral tensorflow processor that I would love to continue using for facial / object recognition. Could anyone suggest an app that does that AND supports the Google Coral usb stick? I am eyeing photoprism but I cant find any mention of tensorflow support.

[pidg30n]

What are your thoughts on Photoprism?
See : https://hub.docker.com/r/photoprism/photoprism 
and : https://github.com/photoprism/photoprism

Looks like facial recognition is done via tensorflow go. I haven't seen much about it in their documentation though... just a curosry search of the repo revealed that..

[Tomb_of_ash]

17 hours ago, pidg30n said:

What are your thoughts on Photoprism?
See : https://hub.docker.com/r/photoprism/photoprism 
and : https://github.com/photoprism/photoprism

Looks like facial recognition is done via tensorflow go. I haven't seen much about it in their documentation though... just a curosry search of the repo revealed that..

Yeah that's the one I went with. I couldnt get tensorflow to work and I'm not a fan of the interface at all but it's better than nothing

[pidg30n]

I also just found LibrePhotos.
https://docs.librephotos.com/docs/installation/
Looks like it integrates with nextcloud, and uses https://github.com/ageitgey/face_recognition for facial recognition... which can use a GPU. Uses cuda instead of tensorflow.

[tigga69]

On 4/19/2023 at 1:38 PM, pidg30n said:

which can use a GPU. Uses cuda instead of tensorflow.

So rather than a $40 coral tpu, I should put in a $1000 gpu into my unraid server???  hmmm not sure I agree with that approach!

 

Using desktops for tensorflow is a different requirement to a server, and like the OP (and many others) I have a requirement to take advantage of a photo library dating back 25 years of family, friends, holidays etc.  and would love to use AI to find old photos in a different way.

 

I've just started using photoprism because I cant find an alternative.  When it works, it is amazing... find photos with person A and person B on a beach.  and it gives you a selection of photos.   The issue I have is I have over 50K photos, but not all are detecting faces and matching.

 

However, I just read today a note on one page saying

Quote

Face recognition was developed and tested under the assumption that the background worker runs every 15 minutes, unless the backend is busy with other tasks like indexing.

 

I read the section on the background worker to imply that you have to wait a long time for it to  process all the faces.

The Point the OP is really asking is that CPU tensorflow can process maybe 3 frames a second whereas coral TPU would process 100.  This means you could manage my library in hours rather than days.  and I can increase the people and object I want to recognise regular.

Without Coral, I could spend a year sorting the library to do facial recognition.

 

 

[pidg30n]

Whatever works for you. I know a lot of folks have graphics cards already, for graphics. So for them it's not an extra expense. Although I'm not sure how well it integrates with docker. 

I'm totally with you on using AI to parse older data in a different way and add value to old collections. Happy to hear about progress as you make it.

[bagican]

Immich doesn't support Coral? Immich is very powerful Google photo alternative

[martabal]

Immich uses onnxuntime for machine-learning and it does not support coral. Currently immich (as of v1.94.1) only supports CUDA (nvidia) and OPENVINO (intel)