Add depth estimation pipeline

[NielsRogge]

Feature request

We currently have 2 monocular depth estimation models in the library, namely DPT and GLPN.

It would be great to have a pipeline for this task, with the following API:

from transformers import pipeline

pipe = pipeline("depth-estimation")
pipe("cats.png")

This pipeline could default to the https://huggingface.co/Intel/dpt-large checkpoint. Also check out the Space that showcases the model.

This can be implemented similar to other pipelines. For an example PR that added a pipeline, see #11598.

Motivation

Pipelines are a great way to quickly perform inference with a model for a given task, abstracting away all the complexity.

Your contribution

I can assist with this, together with @Narsil.

[Narsil]

What would be the output like @NielsRogge ?

My understanding is that depth is just a gray scale image (black = infinitely far, white = infinitely close).

If that's the case It seems really close to image-segmentation in the sense that it's generating a new image from the original image, so we should try and reuse as much as possible.

Also maybe we could have something like image-generation to try and keep the name generic ? (And have an alias for depth-estimation for instance ?)

[nandwalritik]

Hi @NielsRogge I would like to add this pipeline.

[NielsRogge]

Hi @Narsil,

I'm not sure whether we should add this to the existing image-segmentation pipeline. Depth estimation is basically pixel regression, rather than pixel classification (the latter is image segmentation). It would be quite confusing to add it there.

Depth estimation is quite a different field, see e.g. https://paperswithcode.com/task/depth-estimation

And hi @nandwalritik, thanks for your interest in this. Feel free to start a draft PR.

[nandwalritik]

Thanks I will start working on it.

[Narsil]

I'm not sure whether we should add this to the existing image-segmentation pipeline.

I said we should inspire from it, not reuse it, but I suggested using an image-generationone. (Just to be slightly more general)
The output is a grayscale image, right ?