- Change regularization weight to 1.0 by default, and add the regularization term which was missing when computing the functional.
- The 'metric' returned earlier was not divided by the number of time steps. It was simply a sum of the alpha norm over all time steps. Now, divide the sum by number of time steps to get an approximation to the integral of the norm over time. As a consequence the metric returned will be smaller by a factor of sqrt(number of time steps). That means for default parameters people will see metrics which are smaller by a factor of sqrt(10).
- In CurrentMatching, we were scaling the target normals by the ratio (surface area of template)/(surface area of target). Removed this scaling. The code is still there (lines 368-377 of CurrentMatching.hpp), but commented out.
- Updated DblMatrix to remove the cblas calls and call the blas routines directly and correctly. This should be of no consequence to anyone though, unless one has written some code which uses Matrix-Vector or Matrix-Matrix multiplication involving non-symmetric matrices.
- Changed the initial step size for Gradient Descent to 1e-4 from 1.0. Again this should have no consequence for the user. It just avoids a few extra compute steps in the beginning because the algorithm tries to make the step size smaller since 1.0 is too big.
Started using Gnu Triangulated Surface Lib 0.6.7. New validation data added to website. Version 184.108.40.2066.07.28
Modified lddmm-surface to write cost file.2005.09.02
Initial version of Surface Matching module configured and documented.
Last Modified: Monday, 12th December, 2011 @ 09:23am