Codecov Report

Patch coverage:98.78% and project coverage change:+0.04% 🎉

Comparison is base(5f37398) 92.16% compared to head(5ded851) 92.21%.
Report is 7 commits behind head on master.

@@            Coverage Diff             @@##           master    #1251      +/-   ##==========================================+ Coverage   92.16%   92.21%   +0.04%==========================================  Files          98       99       +1       Lines       12371    12453      +82       Branches     2542     2559      +17     ==========================================+ Hits        11402    11483      +81  Misses        646      646- Partials      323      324       +1

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report?Share it here.

@oesteban What do you think about adding anaffine directly toPointset?

classPointset:def__init__(self,coordinates,affine,is_homogeneous): ...def__rmatmul__(self,affine):"""Transform coordinates without evaluating"""assertis_affine(affine)returnreplace(self,affine @self.affine)defas_homogeneous(self):ifself.is_homogeneous:returnself.coordinatesreturnnp.hstack(self.coordinates,np.ones(...))defget_coords(self,homogeneous=False):"""Evaluate coordinates"""coords=self.affine @self.as_homogeneous()ifnothomogeneous:coords=coords[:, :-1]returncoords

Then transforming can simply be:

transformed=xfm @pointset

fromdataclassesimportdataclass,replaceimportnumpyasnp@dataclassclassPointset:coordinates:np.ndarrayaffine:np.ndarrayis_homogeneous:booldef__rmatmul__(self,affine):"""Transform coordinates without evaluating"""assertis_affine(affine)returnreplace(self,affine=affine @self.affine)defas_homogeneous(self):ifself.is_homogeneous:returnself.coordinatesreturnnp.hstack(self.coordinates,np.ones((self.coordinates.shape[0],1)))defget_coords(self,homogeneous=False):"""Evaluate coordinates"""coords=self.affine @self.as_homogeneous()ifnothomogeneous:coords=coords[:, :-1]returncoords# Trick numpy into using our __rmatmul__ndim=2__array_priority__=99defis_affine(arr):returnarr.shape== (4,4)andnp.array_equal(arr[3], [0,0,0,1])

This would work for clouds, meshes or grids and would let us add format-defined affines without applying them prematurely. And for grids, it would allow us to wait to generate the voxel indices.

I would not attempt to apply warps with__matmul__, so a final call toget_coords() would be necessary.

I've gone ahead and pushed forward on bundling affines withPointsets, defaulting to identity.Grid is now a simple wrapper that provides factory functions for creating aPointset from a spatial image (all coordinates) or a mask (selected coordinates only).

Updated comparisons toImageGrid:

• ImageGrid().ndindex corresponds toNDGrid().coordinates (untransformed)
• ImageGrid().ndcoords corresponds toNDGrid().get_coords() (transformed)

LMK what you think. It can be reverted, if this doesn't make sense to anybody else.

I've removed the triangular meshes from this for now. The existing architecture depended too much on the previousPointset implementation, and I don't want to hold this up.