MixUp and CutMix transforms for semantic segmentation

[vedantdalimkar]

Is there any way to use the MixUp and CutMix transforms for semantic segmentation masks? I could not find any documentation on it.

If this functionality does not exist, I will be happy to submit a PR for the same.

Motivation - CutMix is used in SOTA semi-supervised semantic segmentation methods such as UniMatch and MixUp is used in knowledge distillation methods such as "Knowledge distillation: A good teacher is patient and consistent"

[AntoineSimoulin]

@vedantdalimkar, after going through the codebase it seems MixUp and CutMix transformations only take tv_tensors.Image, tv_tensors.Video, and entry marked as "labels" as input. If those transformations are indeed used for segmentation masks in the literature, it would be a very nice addition to TorchVision. Feel free to send a PR and I will be happy to review!