Photoprism is able to use TensorFlow on Raspberry PI 3/4

[taophp]

You wrote in your README.md:

Raspberry Pi can be used but microservices container has to be comment out in docker-compose since TensorFlow has not been supported in Docker image on arm64v7 yet

I currently use Photoprism on my Raspberry Pi 3 and object / face recognition is working quite well. TensorFlow works. I just followed the instructions found here. So I think you should inspire you from it to make Immich able to use TensorFlow on Raspberry Pi 3/4. What do you think ?

[alextran1502]

Thank you, I will take a look how they build their image for Arm64 and find a way to incorporate that into Immich 👍

[taophp]

@alextran1502 So, is Photoprism able to use TensorFlow on Raspberry Pi now ?

[alextran1502]

I don't know, you have to ask the Photoprism team

[taophp]

@alextran1502 Sorry, I'm quite tired those days. I meant : is Immish able to use TensorFlow on Raspberry Pi now ?

[Eirikr70]

@alextran1502 Thanks. I hadn't been looking for it in that place.

[boboche]

with 1.79.1 I don't see any tagging on images, I see the exif, I see that the face detection works, but no tags or classification even if I change the model in the config. Same issue with Rpi 4/4GB with SDimage or on a fresh ubuntu 22 VM. Am I missing something or its broken?

[alextran1502]

@boboche We don't use image classification's results for anything actually, and we don't use traditional object detection methods. We use CLIP search in order to search for things and the information is encoded as an array of floating number.

[boboche]

Thanks for the reply Alex, I am trying to figure out the intended workflow and see how to fit my needs within what's available. I see object tagging / image classification in this context: I have over a million images, I want to dump them on an offline solution that can help search/segmentation based off what is in the image or its context (i.e. industrial, beach, buildings, sky, clouds, and what not). I want to see the labels/tags or whatnot that are applied to the image in order to assess the tagging performance (i.e. calibrate it to 0.9, or 0.95/0.8 depending on the signal to noise ratio) and finally, be able to see those tags will help me search relevant pictures and/or further filtering down to specifics.

That on top of all classification by date, ISO, lens and exif whatnots.

The closest I've seen doing it properly is librephotos so far but immich is more mature and polished, My use case might be a niche scenario but aside from the very large quantity of images, I don't see how people wouldn't benefit from this if using a good AI model.

Another approach would be to tag the image and embed this in the jpg info but I am dealing with raw images as well, and there are no out of the box solutions for such a workflow, made simple, that I am aware of.

Nice work for the overall project btw :)

[jrasm91]

The model we use now is pretty inaccurate. You bring up some good ideas related to search and filtering though. We're in the process of refactoring search, so hopefully it will improve in the future. We do have plans to enable user supplied tags in the future as well.