Segmentation Dataset alignment against `draw_segmentation_mask()`

[cregouby]

Current Situation

Current and future segmentation datasets

• coco_segmentation_dataset
• oxfordiiitpet_segmentation_dataset
• ...
  have default segmentation annotations file content that are not ready for draw_segmentation_mask() function.

In order to ease usage of draw_segmentation_mask, we currently apply a specific hidden target_transform within each dataset function :

torchvision/R/dataset-coco.R

Lines 130 to 145 in 8acd690

	masks <- lapply(seq_len(nrow(anns)), function(i) {
	seg <- anns$segmentation[[i]]
	if (is.list(seg) && length(seg) > 0) {
	mask <- coco_polygon_to_mask(seg, height = H, width = W)
	if (inherits(mask, "torch_tensor") && mask$ndim == 2) return(mask)
	}
	NULL
	})
	masks <- Filter(function(m) inherits(m, "torch_tensor") && m$ndim == 2, masks)
	if (length(masks) > 0) {
	masks_tensor <- torch::torch_stack(masks)
	} else {
	masks_tensor <- torch::torch_zeros(c(0, H, W), dtype = torch::torch_bool())
	}

torchvision/R/dataset-oxfordiiitpet.R

Lines 155 to 160 in 978ffa6

	masks <- png::readPNG(mask_path) * 255
	masks <- torch_tensor(masks)
	mask1 <- (masks==1)
	mask2 <- (masks==2)
	mask3 <- (masks==3)
	masks <- torch_stack(list(mask1, mask2, mask3))$to(dtype = torch_bool())

Design driver

• try to adhere to (py) torchvision guidelines "Torchvision datasets preserve the data structure and types as it was intended by the datasets authors. So by default, the output structure may not always be compatible with the models or the transforms."
• clean and easy-to-understand workflow for the end-user to load the dataset and draw a segmented image and to infer with a asegmentation model and draw the result.
• not having to perform hidden complex target-transformation of raw annotation into the image_with_segmentation_mask data format

Expected situation

• add two target transform functions :