Tutorial #9 Option 0 - Possibly incorrect bundle adjustment about gsoc2011sfm HOT 4 CLOSED

[deleted comment]

I get a whole heap of LAPACK errors, if that helps:

(for example)
LAPACK error: the leading minor of order 1 is not positive definite, the 
factorization could not be completed for dpotf2/dpotrf in sba_Axb_Chol()

Also, I might be crazy, but I swear I got a good result once (image attached). 
But every other time I run it I get the same result with cameras piled on top 
of each other.

Unfortunately this problem is still not solved... The lapack error is something 
I have too but this seems to be normal (I will ask Vincent if this is something 
he knows).

I never had a "good" result (except when using the whole sequence of the 312 
images), so this bundle adjustment starting to annoy me ;)

I set this issue as started, even if I have no clue how to fix this...

OK, so I now understand what's going on. This dataset is only given with P 
matrix so I used KRT_from_P function to get the different camera matrix. But 
this function wasn't working well as sometimes the rotation matrix was not a 
rotation matrix. So when I converted the rotation matrix into rotation 
quaternion, the bundle adjustment wasn't working...
Now it's working well because I enforce the determinant using this 
transformation:
      K.col(0) = -K.col(0);
      R.row(0) = -R.row(0);
      t(0)     = -t(0);
Thanks again to Julien Michot!