GeneralizedIterativeClosestPoint bad results when data isn't centered

The Generalized ICP implementation gives much worse results when data isn't centered around the mean of the target pointcloud (i.e. subtracting mean of target pointcloud from all points) prior to registration.
With my data i consistently get a fitness score of ~90 when not centering data, while centered data gives a score of 0.
Point to Point, Point to Plane and Nonlinear ICP all work fine when data isn't centered.

I'm using a Matlab-Wrapper with Matlab data and i'm not really supposed to give out that data so i can't attach it here. If you have trouble reproducing this let me know and i'll figure something out.
I'm working on PCL 1.7.2 in Visual Studio 2010 64. I couldn't find a fix for this in the trunk version.

Start with the Generalized ICP implementation in PCL 1.7.2 and compare registration on centered and uncentered target point clouds using the reported Matlab-wrapper setup. Reproduction is the main first step because the original data cannot be shared; done means uncentered input no longer produces substantially worse results than centered input.