The executable (hj_skel) should be run as follows:

./hj_skel threshold method model.vtk
        with threshold: negative value to threshold flux/divergence
                       (e.g. in range  [-1e+3f, -1e-3f])
                method: 0-original H-J or 1-curvature-corrected.
             model.vtk: is some binarized model (as spit out by binvox).

Once the surface skeleton is computed, the program opens a window where
you can zoom/rotate/pan with the mouse. Finally, pressing 't' shows some
'importance' of the skeleton (i.e. the divergence of the velocity
field). I believe this is of less importance, in this context especially
sine it may not preserve the topology of the original model.

The most important parameter is threshold. For the original, Siddiqi
method, some small, negative value (e.g. -0.3 to -10.0) produces decent
surface skeletons. The curvature-corrected one, is quite unstable (due
to some numerical difficulties which aren't really documented in the
paper). Anyways, the threshold here could be/should be much higher, i.e.
negative values of up to -100, etc. should produce 'optimal' results.
>From my quick experiments, it seems though, that at least in 3D, the
original method is much more reliable. 
