[RFC] Implement transforms primitives for Segmentation Masks

[vfdev-5]

Related tracker issue: #5754

Functional transforms for segmentation masks

The idea is to implement the following low-level functional transforms for segmentation masks:
https://github.com/pytorch/vision/blob/main/torchvision/prototype/transforms/functional/_geometry.py

• affine_segmentation_mask (@vfdev-5) [proto] Added functional affine_segmentation_mask op #5613
• rotate_segmentation_mask (@vfdev-5) [proto] Added functional rotate_segmentation_mask op #5692
• perspective_segmentation_mask (@vfdev-5) [proto] Added functional perspective_bounding_box/segmentation_mask ops #5888
• center_crop_segmentation_mask (@federicopozzi33) feat: add functional center crop on mask #5961
• crop_segmentation_mask (@vfdev-5) Added crop_segmentation_mask op #5851
• resized_crop_segmentation_mask (@vfdev-5) Added resized_crop_segmentation_mask op #5855
• pad_segmentation_mask (@federicopozzi33) feat: add functional pad on segmentation mask #5866, need more tests: https://github.com/pytorch/vision/pull/5866/files#r859158083 -> [proto] Added tests for other padding modes #6104
• vertical_flip_segmentation_mask (@federicopozzi33) test: add functional vertical flip tests on segmentation mask #5860, refactor: port RandomVerticalFlip to prototype API (#5524) #5633

How to test the implementation ?

Put tests into test/test_prototype_transforms_functional.py:

• correctness test
• eager vs scripted
• scriptable
• check on cpu and cuda

cc @vfdev-5 @datumbox @bjuncek

[federicopozzi33]

Hi @vfdev-5,

I would like to work on vertical_flip_segmentation_mask primitive.

[vfdev-5]

@federicopozzi33 sure, go ahead!

EDIT: Currently, only tests are missing and you already coded the low-level primitive

[vadimkantorov]

Related: #4415

[federicopozzi33]

@vfdev-5,

I'd like to go on with pad_segmentation_mask primitive.

[vfdev-5]

Thanks @federicopozzi33 , sure go ahead !

[federicopozzi33]

I'd like to complete this issue doing center_crop_segmentation_mask.

[vfdev-5]

@pmeier in order to close this tracker as done we have to add more correctness tests for padding with other padding modes (https://github.com/pytorch/vision/pull/5866/files#r859158083)

pad_segmentation_mask uses pad_image_tensor and right now we does not have any tests for pad_image_tensor. I suggest to check correctness with one of suggested ways (from the link above). Later these tests could be reused for checking pad_image_tensor.

Any other suggestions from your side ?

[pmeier]

Nope, sounds like a good plan.

[vfdev-5]

I'm closing this tracker as done