[Insight-users] LBFGSB with Mattes MI

[Severino Fernandez]

I have to register two remote sensing images from different sensors.
I use OTB for it, and sometimes the registration is successful with
Mattes MI as metric and 1+1 evolutionary optimizer.
I wanted to try with LBFGSB, as it seems to be very popular, and
somewhere I even read that it would work OK for a "noisy" metric, like
MI.
The deformation I use is a simple translation (applied to small
windows, 200x200).
But LBFGSB, after fast convergence, gives very slow translation values
as optimized parameters (for instance less tna 1 pixel, when the
displacement is in the order of 20). I have set very different values
for LBFGSB parameters, but the result does not practically change.
I start basically from the same parameters as in DeformableRegistration12.
Is any other setting to be considered? In fact, the optimization seems
practically not to move from the starting value (0,0), as if each
optimization step were extremely small.
I appreciate any recomendation.
Severino