Problem with Transform.SimilarityTransform

[Jean K]

Hi everyone,

I'm currently trying to use skleanr.Transform.SimilarityTransform to remove 
scaling translation and rotation from one set of points to align it to the 
other.
However, if I centre the sets around the origin first, there seems to be 
frequently a problem (which doesn't occur if the points are all positives), 
the output being NaN.

I tried to write a small reproducible code:
In [77]:

#I fixed the seed here for reproducibility but this happens often

np.random.seed(4)

 

#Two random set of points

a = np.random.randn(10, 2)

b = np.random.randn(10, 2)

 

# Center the points arount the origin

a -= np.mean(a, axis=0)[np.newaxis, :]

b -= np.mean(b, axis=0)[np.newaxis, :]

 

tform = SimilarityTransform()

tform.estimate(a, b)

tform(a)

Out[77]:

array([[ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan],
       [ nan,  nan]])


Note that if I don't centre the point there is no problem:

In [89]:

#I fixed the seed here for reproducibility but this happens often

np.random.seed(4)

 

#Two random set of points

a = np.random.randn(10, 2)

b = np.random.randn(10, 2)

 

# Center the points arount the origin

#a -= np.mean(a, axis=0)[np.newaxis, :]

#b -= np.mean(b, axis=0)[np.newaxis, :]

 

tform = SimilarityTransform()

tform.estimate(a, b)

tform(a)

Out[89]:

array([[ 3.76870886, -0.35152078],
       [ 1.83453334, -1.25080725],
       [ 5.42428044, -4.30088121],
       [ 2.51364241, -1.00154154],
       [ 6.14244682, -2.71511189],
       [ 5.37956586, -0.65190768],
       [ 4.5752074 , -0.19039746],
       [ 1.96968262, -1.99729896],
       [ 1.47865106,  0.59493455],
       [ 5.39473376, -0.31125435]])


Sometime it also tells me that the *SVD doesn't converge*.

Any idea what is going on?


Thanks,


Jean 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scikit-image/attachments/20131204/28b504b2/attachment.html>