This is to do with the maximum tuple size. http://blogs.msdn.com/b/vcblog/archive/2011/09/12/10209291.aspx
…ults, only potentially the accuracy.
…bjects from TooN for 2D and 3D
…culate the rotation and made them more robust against input with too little constraints as well
…point sets after Horn
… 3 perspective projected points
… TooN 1.0isms, different update implementation using Cholesky for better stability (keeps things symmetric) and no dependency on transformCovariance anymore (a small performance penality in predict ...)
- added drand48 implementation for Win32
…esults to the old code, due to differences in numerics, tolerances and etc. The nullspace is now found using Gauss-Jordan, instead of SVD (for a speedup of about 30x for this particular part). New hand written polynomial root finder, for a speedup of 4x and improved reliability (the old one would occasionally expect -1 roots due to numerical instabilities and fail with an alloc_error). The new root-finder will not attempt to find roots of p(x) for |x|>100, since it is not possible to find these root to high enough accuracy. Roots are bracketed using Sturm sequences (as before) but using the resursive implementation which is more efficient. The root polishing uses a hybrid Newton-Raphson/bisection rootfinder which usually gets roots to higher accuracy than the old bisection solver and almost always converges in < 8 iterations, and often in 4. It will give up after 10 iterations. A new tolerance feps is introduced, so that for a numerical root at x, if |p(x)| > feps, the root will not be recorded. The other tolerances (eps and zeps) follow the convention of NR in C.