Deformable registration in D

Robert Jacques sandford at jhu.edu
Fri Aug 3 05:26:33 PDT 2007


> What cutoff are you using? Denormals are only generated for x around  
> -11300 for 80-bit reals, but for floats they'll happen for x between  
> about -85 and -105. IIRC, denormals are ~100 times slower than normals  
> on x87, so could easily be performance killers.

I first started with -5 and and used -3 in the release, as the method is  
multi-resolution/sigma. Matlab, if I remember, used a sigma of 2.5 as its  
default cutoff for Gaussian window. And yes, due to their slow speed, I  
included denormal detection in order to error early. (infinity or nan are  
invalid outputs and happen only when the algorithm becomes unstable)



More information about the Digitalmars-d-announce mailing list