Vectorize labels into polygons

[quentinblampey]

Converting a mask (2D Labels, multiscale or not) into shapes, i.e. a GeoDataFrame.

Speed test

1 min on 300k cells locally

Notes

• vectorization is applied on chunks
• for each chunks, scikit_image is used to find contours (for each label, a cropped view is extracted, then vectorized, and then translated according to the bounding box coords)
• polygons are also translated according to the chunk coords
• labels on multiple chunks are merged via the dissolve method

Future steps

• make sure shapes are always valid (see Invalid shapes during vectorization #583)

[codecov]

Codecov Report

Attention: Patch coverage is 90.65421% with 10 lines in your changes missing coverage. Please review.

Project coverage is 92.52%. Comparing base (3733514) to head (52404d7).
Report is 82 commits behind head on main.

Files	Patch %	Lines
src/spatialdata/_core/operations/vectorize.py	92.39%	7 Missing ⚠️
src/spatialdata/models/models.py	62.50%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #560      +/-   ##
==========================================
- Coverage   92.55%   92.52%   -0.03%     
==========================================
  Files          42       42              
  Lines        6003     6086      +83     
==========================================
+ Hits         5556     5631      +75     
- Misses        447      455       +8     

Files	Coverage Δ
src/spatialdata/__init__.py	100.00% <100.00%> (ø)
src/spatialdata/_core/operations/aggregate.py	94.05% <100.00%> (ø)
src/spatialdata/_core/query/_utils.py	95.65% <ø> (+1.53%)	⬆️
src/spatialdata/_core/query/spatial_query.py	95.19% <100.00%> (+0.01%)	⬆️
src/spatialdata/models/models.py	87.52% <62.50%> (-0.47%)	⬇️
src/spatialdata/_core/operations/vectorize.py	94.33% <92.39%> (-1.61%)	⬇️

[LucaMarconato]

I would add a comment explaining the choice of 4.

[quentinblampey]

I added a comment in the code.
When the number of vertices is 3, the shape is basically a line (2 points + repetition of the first point), which would raise a Shapely error. This is why I added this filtering.
Though, we should never have a shape with less than three coordinates: even a single pixel would be converted into a square. I just added this filtering for safety, but I'm not sure if skimage can produce a shape with less than 3 coordinates.

[LucaMarconato]

Excellent PR! I just added a minor comment, but otherwise looks great! I will check/implement the following minor cases (copying them from the hackmd):

small things

• to_circles() seems not to be implemented for points
• to_circles() seems not to be implemented for circles
• circles_to_polygons() should be more explicit: if values_geotypes == ['Point'], elif is.subset(['Polygon', 'MultiPolygon']) ... , else fail

other small things

• remove circles_to_polygons() and use to_polygons() instead
• to_polygons() should work for polygons, multipolygons, circles and points (or maybe fail on points)

I will merge after I complete them.