[jira] [Created] (MATH-1400) Rotation(double[][], double) constructs inverse quaternion due to bug in mat2quat

["Robert Huitl (JIRA)" <ji...@apache.org>]

Robert Huitl created MATH-1400:
----------------------------------

             Summary: Rotation(double[][], double) constructs inverse quaternion due to bug in mat2quat
                 Key: MATH-1400
                 URL: https://issues.apache.org/jira/browse/MATH-1400
             Project: Commons Math
          Issue Type: Bug
    Affects Versions: 3.6.1
            Reporter: Robert Huitl


Constructing a Rotation object from a rotation matrix and extracting the quaternion gives the *inverse* of the expected quaternion.

E.g. something like this:
{code}
Rotation rot = new Rotation(Array2DRowRealMatrix(matrixData).getData())
Quaternion q = new Quaternion(rot.getQ0(), rot.getQ1(), rot.getQ2(), rot.getQ3());
{code}

results in q being the inverse of what is expected.

I tracked this down to Rotation#mat2quat(final double[][]) which seems to access the matrix elements as if they were stored transposed. E.g. compare with Quat4f#set(Matrix3f):

{code:title=Rotation.java}
quat[1] = inv * (ort[1][2] - ort[2][1]);
quat[2] = inv * (ort[2][0] - ort[0][2]);
quat[3] = inv * (ort[0][1] - ort[1][0]); // <-- m01 - m10
{code}

{code:title=Quat4f.java}
this.x = (m1.m21 - m1.m12) * ww;
this.y = (m1.m02 - m1.m20) * ww;
this.z = (m1.m10 - m1.m01) * ww; // <-- m10 - m01
{code}

I compared the result from Commons Math with JavaFX, JavaX Vecmath and NumPy + http://www.lfd.uci.edu/~gohlke/code/transformations.py.html. All but Commons Math agree on the result.

You can find my test program here: http://pastebin.com/jxwFi9mt
It prints the following output (Python results added manually):

{noformat}
[ 0.7  0.0  0.0 -0.7] (Commons Math)
[ 0.7  0.0  0.0  0.7] (JavaFX)
[ 0.7  0.0  0.0  0.7] (JavaX Vecmath)
[ 0.7  0.0  0.0  0.7] (NumPy + transformations.py)

[-0.2  1.0  0.0  0.0] (Commons Math)
[ 0.2  1.0  0.0  0.0] (JavaFX)
[ 0.2  1.0  0.0  0.0] (JavaX Vecmath)
[ 0.2  1.0  0.0  0.0] (NumPy + transformations.py)

[ 0.2  0.0  1.0  0.0] (Commons Math)
[ 0.2  0.0 -1.0  0.0] (JavaFX)
[ 0.2  0.0 -1.0  0.0] (JavaX Vecmath)
[-0.2  0.0  1.0  0.0] (NumPy + transformations.py)

[-0.2  0.0  0.0  1.0] (Commons Math)
[ 0.2  0.0  0.0  1.0] (JavaFX)
[ 0.2  0.0  0.0  1.0] (JavaX Vecmath)
[ 0.2  0.0  0.0  1.0] (NumPy + transformations.py)
{noformat}

The other constructor using mat2quat() is probably also affected although I did not verify this.




--
This message was sent by Atlassian JIRA
(v6.3.4#6332)