The registration from hemibrain to FAFB is squished

[SridharJagannathan]

Hi,
When I try to download a mesh or a skeleton to transform to FAFB like below:

neuron_mesh = neu.fetch_mesh_neuron(1324365879, lod=2, with_synapses=False)
flywireneuron_mesh=navis.xform_brain(neuron_mesh,source='JRCFIB2018Fraw', target='FAFB',verbose=True)

The actual hemibrain mesh looks like below:

However, the transformed stuff looks like squished like below:

[schlegelp]

The overall transform looks good as you can see from this co-visualization with the FAFB neuropil mesh.

The transform path is this

1. JRCFIB2018Fraw (voxel) to JRCFIB2018F (nm)
2. JRCFIB2018F (nm) to JRCFIB2018Fum (um)
3. JRCFIB2018Fum (um) to JRC2018F (um)
4. JRC2018F (um) to FAFB14um (um) 
5. FAFB14um (um) to FAFB14 (nm)

This (local) squishing seems to be introduced in step 3. I checked if it had anything to do with my implementation of the Saalfeld lab's Hdf5 transforms but I also tried using the original Java implementation via nat and that produced the same results. I guess we could ask John Bogovic - maybe he'd be able to improve the transform? @jefferis any thoughts?