Unused attribute voxel_origin in IntCoords

[kasparas-k]

I have noticed a currently unused attribute in IntCoords, that I was unsure what to do about, so instead of proposing a fix in a pull request, I'm raising an issue. Outside the file warpconvnet/geometry/coords/integer.py there are no references to voxel_origin, and the only use of it within the file would lead to an error:

WarpConvNet/warpconvnet/geometry/coords/integer.py

Lines 83 to 87 in 672653d

	def unique(self) -> "IntCoords":
	unique_indices, batch_offsets = voxel_downsample_random_indices(
	self.batched_tensor, self.offsets, self.voxel_size, self.voxel_origin
	)
	return self.__class__(self.batched_tensor[unique_indices], batch_offsets)

as the function voxel_downsample_random_indices takes only three arguments

WarpConvNet/warpconvnet/geometry/coords/ops/voxel.py

Lines 113 to 117 in 672653d

	def voxel_downsample_random_indices(
	batched_points: Float[Tensor, "N 3"], # noqa: F821
	offsets: Int[Tensor, "B + 1"], # noqa: F821
	voxel_size: Optional[float] = None,
	) -> Tuple[Int[Tensor, "M"], Int[Tensor, "B + 1"]]: # noqa: F821

Also, all info besides the coordinates (stride, voxel_origin, voxel_size) is lost on the return statement:

WarpConvNet/warpconvnet/geometry/coords/integer.py

Line 87 in 672653d

return self.__class__(self.batched_tensor[unique_indices], batch_offsets)

Same loss of information seems to happen in the sort method, where self.class is used to create a new instance of IntCoords with only the coordinates and offsets.

As it is now, this is fairly confusing as the attribute does nothing, for example when voxels are turned back into points by Voxels.to_point, there is no additional offset to the voxel origin, which I would expect if I provided such an argument.

WarpConvNet/warpconvnet/geometry/types/voxels.py

Line 243 in 672653d

batched_points = RealCoords(self.coordinate_tensor * voxel_size, self.offsets)

I think it may be useful to control the voxelization with voxel_origin, but of course the same control can be achieved with less book-keeping by just translating the whole point cloud to your desired voxel origin.