coordinateTransformations generated for 0.4 are scale-only

[d-v-b]

The most common methods of image downsampling result in a translation of the downsampled image, but this code for generating coordinateTransformations metadata only returns scale transformations, which will be incorrect for almost all multiscale pyramids.

ome-zarr-py/ome_zarr/format.py

Lines 260 to 271 in 56f72b0

	def generate_coordinate_transformations(
	self, shapes: List[tuple]
	) -> Optional[List[List[Dict[str, Any]]]]:
	data_shape = shapes[0]
	coordinate_transformations: List[List[Dict[str, Any]]] = []
	# calculate minimal 'scale' transform based on pyramid dims
	for shape in shapes:
	assert len(shape) == len(data_shape)
	scale = [full / level for full, level in zip(data_shape, shape)]
	coordinate_transformations.append([{"type": "scale", "scale": scale}])
	return coordinate_transformations

Suggested fix: generate translation transforms

[will-moore]

Thanks for this @d-v-b.
Let me just clarify on how to generate the translations here. Looking at the spec 0.4: "If translation is given it MUST be listed after scale to ensure that it is given in physical coordinates".

So, if we have an image that has a pixel size of 1 microns then the 'scale' for each resolution would look like this (given a zoom factor of 2 between resolutions):

dataset0: {"scale": [1, 1]}
dataset1: {"scale": [2, 2]}
dataset2: {"scale": [4, 4]}
dataset3: {"scale": [8, 8]}

When mapping to physical coordinates, I guess the translation needed after scaling depends on where the 'anchor' is when you're scaling. I'm assuming that this is the centre of the pixel at [0, 0]. So after scaling then the top-left of that pixel at 2 x its original size will be at -0.5, 0.5 microns. So then we need to translate 0.5, 0.5, and so on..

dataset0: {"translation": [0, 0]}
dataset1: {"translation": [0.5, 0.5]}
dataset2: {"translation": [1.5, 1.5]}
dataset3: {"translation": [3.5, 3.5]}

So the translation for each resolution will be:

(physical-pixel-size-at-resolution-N - physical-pixel-size-at-resolution-0) / 2.

That seems to be what https://github.com/thewtex/ngff-zarr is doing when it generates translations.

[d-v-b]

yep that's basically it for the most common types of downsampling methods.