[RFC] Extend the scope of ImageFolder for segmentation data

[glemaitre]

I am starting to work on segmentation problem.

ImageFolder is great but seems to be dedicated for classification.

Would it make sense to modify it such that we can load input/target which can be both list of images as in the case of CamVid dataset?
It is also true that segmentation can also be organized with a list of images and a list of annotations which does not fit the previous requirement.

So I was wondering what the core devs are thinking about it.

[glemaitre]

NB: the idea is to propose a potential PR if we can come to a proper design.

[fmassa]

Hi,

Sorry for the delay in replying.

We will be extending the datasets to better support detection / segmentation tasks.
This will also involve having transforms that nicely plug in them.
This is something I'm looking into now.

About CamVid, we could probably add support for it as a separate dataset? I'm not familiar on how the data is organized though.

[Choco31415]

I have a generalized suggestion. What if PyTorch got a new class that could be given multiple Datasets? The class would then load in tuples of data by using the same index for __getitem__ on all Datasets.

Considering CamVid, this new class would be input two Datasets. One Dataset would load in 'Raw', and another Dataset would load in 'Labeled'.

p.s. DatasetFolder seems like a weird name given its purpose. Maybe a better name would be ClassificationFolder? (Also why is DatasetFolder in vision when it's not restricted to images?)

[SebastienEske]

Like this pull request ?