Skan returns pixel 0,0 under all conditions

[chrisdonlan]

Perhaps this is not a bug. However, it was a trip wire.

>>> foo = numpy.zeros((50,50))
>>> skeleton_to_csgraph(foo)
ValueError: ...
>>> foo[0,0] = 1
>>> skeleton_to_csgraph(foo)
ValueError: ...
>>> foo[-1,:] = 1
>>> adj, coords, degs = skeleton_to_csgraph(foo)
>>> coords
[[0. 0.], [0. 0.], ...

This issue summarizes as:

• Two errors above are for all zeros, and an array with only isolated points.
• The final call has two [ 0. 0. ] coords. At least 1 [0. 0.] is always present
• The coords are all floating point and must be cast to integers for indexing

To fix it, I wrapped it like:

def get_cs_graph(skeleton: numpy.ndarray):
    """
    Get compressed sparse adjacency matrix of a skeleton
    :param skeleton: a skeletonized image
    :return: adj:csv_matrix, nonzero_pixel_coords: ndarray, degree_mask:ndarray
    """
    # Skan: https://jni.github.io/skan/getting_started.html

    # Skan, the source of 'skeleton_to_csgraph' always returns the origin in the pixel list, even if the
    # matrix is mostly zeros, so calculate this component with numpy
    coords = numpy.vstack(numpy.where(skeleton > 0)).T

    # Skan yields errors under some numerical conditions related to all zero arrays, and arrays with isolated pixels
    try:
        adj, _, degrees = skeleton_to_csgraph(skeleton)
    except ValueError as e:
        warnings.warn(str(e))
        adj = csr_matrix()
        degrees = numpy.zeros_like(skeleton)
    return adj, coords, degrees

[jni]

Thanks for raising @chrisdonlan! It's definitely a trip wire and you're not the first to raise it (see #107 and #93). I might be able to set some time aside next week to address this. I think the ideal fix is to add a layer of indirection between the pixel ids and the coordinates array.

Note that the float coordinates are by design, because some junction pixels get collapsed to their centroid. Perhaps a utility function that does tuple(np.round(coords).astype(int))) would be a useful addition also. It might even belong in scikit-image itself.