Measured a 30% speedup on a 640x640 image. Fundamentally, this makes
fewer passes over the image; so it should never be worse that the
previous implementation.
This PR:
- speeds up SLIC while reducing memory usage,
- closes#717,
- adds support for a voxel spacing argument for anisotropic images,
- removes default gaussian blurring,
- removes "magic" guessing whether input was multichannel.
Implement @ahojnnes's comments on pull request. Use `np.double` as
dtype for arrays because in Cython, `float` is not `np.double`. And
add further clarification about the `spacing` parameter.
Previously, having a different `sigma` for different dimensions
required an array input. This allows the user to use a simple list,
which gets converted to an array internally.
Importantly, it removes a very unhelpful error:
```python
>>> im = np.random.rand(10, 20)
>>> from skimage import segmentation as seg
Exception AttributeError: "'UmfpackContext' object has no attribute '_symbolic'" in <bound method UmfpackContext.__del__ of <scipy.sparse.linalg.dsolve.umfpack.umfpack.UmfpackContext object at 0x1045ff5d0>> ignored
>>> s = seg.slic(im, 2, sigma=[2, 1])
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
<ipython-input-3-689b36a2f0ef> in <module>()
----> 1 s = seg.slic(im, 2, sigma=[2, 1])
/Users/nuneziglesiasj/venv/skimdev2/lib/python2.7/site-packages/scikit_image-0.9dev-py2.7-macosx-10.5-x86_64.egg/skimage/segmentation/slic_superpixels.pyc in slic(image, n_segments, compactness, max_iter, sigma, multichannel, convert2lab, ratio)
106 if not isinstance(sigma, coll.Iterable):
107 sigma = np.array([sigma, sigma, sigma])
--> 108 if (sigma > 0).any():
109 sigma = list(sigma) + [0]
110 image = ndimage.gaussian_filter(image, sigma)
AttributeError: 'bool' object has no attribute 'any'
```
The `ratio` keyword has been deprecated but it was still being used in
the example in the SLIC docstring. This replaces that usage by the new
`compactness` keyword.
`np.atleast_3d` will add a singleton dimension at the end of an array
if needed. This is not the correct thing to do if `multichannel=False`
based on the subsequent lines. If the input image was 2D with shape
`(40, 50)` and `multichannel=False`, then `np.atleast_3d` gives it
shape `(40, 50, 1)`, and then, because `multichannel=False`, the rest
of the code gives it shape `(40, 50, 1, 1)`. This results in the final
returned array having shape `(40, 50, 1)` instead of the desired
`(40, 50)`.
This commit fixes that and updates the test to detect this failure.
Luis Pedro Coelho