MRG: Recreate our helmet graphic

[larsoner]

I tried to recreate our helmet:

Using Mayavi I get this output for the new example:

If I switch to PyVista, I get:

Problems thus far:

1. The coincident topology is not resolved for mayavi or PyVista (if you rotate the scene, PyVista also cuts off parts of the yellow). Perhaps if plot_evoked_field could set the polygon_offset it could be fixed?
2. The coincident topology for the field lines is wrong for PyVista -- they should be on top.
3. The camera systems are not equivalent between PyVista and Mayavi.

@GuillaumeFavelier can you look into these when you get a chance? Feel free to push directly here

[GuillaumeFavelier]

The camera settings are better IMO. It's good enough for me but I let you try @larsoner

I use a fixed width for the tubes, it would be better to find a suitable value automatically.

The coincident topology still needs to be fixed.

[larsoner]

Indeed it is getting closer:

[larsoner]

For coindicent topology, can we have something like polygon_offset=0 in some mesh call? Then it sets the appropriate Mayavi and PyVista properties under the hood (use polygon offset resolution + set polygon offset)? Assuming it fixes the problem then it's okay to make use of this in the example, and we can use it in _Brain, too. (I've noticed some coincident topology problems there, too, from time to time when rotating a brain.) We probably need to enable the polygon offset for all meshes for it to work (?).

[GuillaumeFavelier]

I added polygon_offset in mesh() and surface() and updated the docstrings. It also replaces resolve_coincident_topology in _Brain. I'm not sure what is the best strategy to resolve those issues. For now, I just set some fixed values that give the best result based on what I observe.

We probably need to enable the polygon offset for all meshes for it to work (?)

That I don't know. I did some experiments locally and it does not seem necessary but it may be in the future once we have a definitive solution.

Also about Mayavi, I did not find any alternative to SetRelativeCoincidentTopologyPolygonOffsetParameters. Here are some relevant default values that I found:

...
'resolve_coincident_topology': 'polygon_offset',
'resolve_coincident_topology_': 1,
'resolve_coincident_topology_polygon_offset_faces': 1,
'resolve_coincident_topology_z_shift': 0.01,
...

[GuillaumeFavelier]

This is the helmet I have with the new changes:

[GuillaumeFavelier]

By letting the contour filter detect the range of scalar values, I get the same result (talking about the lines):

PyVista	Mayavi

[larsoner]

• Can you try playing with the Mayavi parameters window (leftmost icon in the toolbar) to see if you can find the right button to tweak there to fix the coincident topology? If you can't get it to work I can try.
• Should we add a parameter for the line width of the contours? It looks like mayavi defaults to 1 and PyVista doesn't

[GuillaumeFavelier]

1. I can try
2. For PyVista these are tubes now, I will render again using just the lines. To tweak that, there is the width parameter.

This is what I use:

        # And the field lines on top
        renderer.contour(surface=surf, scalars=data, contours=n_contours,
                         vmin=-vmax, vmax=vmax, opacity=alpha,
                         colormap=colormap_lines, kind='tube',
                         width=0.0005)

[larsoner]

Do the lines also require a polygon_offset tweak?

[larsoner]

(The tubes presumably do not because they actually protrude from the surface; the lines OTOH will be coincident)

[GuillaumeFavelier]

Seems like it does not matter it it's just a screenshot. The difference is more visible when you interact with the scene:

line	tube

Do the lines also require a polygon_offset tweak?

Apparently they don't need the modification.

[larsoner]

Okay, feel free to choose either one based on which one you think looks better / more like the original. I guess the only remaining issue is the Mayavi coincident topology, right?

[GuillaumeFavelier]

I guess the only remaining issue is the Mayavi coincident topology, right?

Yes indeed

[GuillaumeFavelier]

I have pretty much the same mapper options:

I tried z-shift but there is some noise remaining near the cell edges:

The best solution I found was to "Force translucent" the white surface (that way, the colors blend):

Result:

[larsoner]

pip-pre errors/segfaults are numpy/numpy#17198 and can be ignored for now

[GuillaumeFavelier]

Update after #8182 :

Mayavi	PyVista

[GuillaumeFavelier]

Also we should consider making PyVista the default backend. It's an item of #7162

[agramfort]

Also we should consider making PyVista the default backend. It's an item of #7162 <#7162>

+10 ! can you start a PR for this?

[GuillaumeFavelier]

@agramfort I opened #8220 :)

[agramfort]

21 was the mayavi default?

[GuillaumeFavelier]

yes, IIRC

[GuillaumeFavelier]

My bad, it's actually 10 but 21 was the default number of contours used in the function. I will change that

[agramfort]

ok merge when green then !

…

[GuillaumeFavelier]

I pushed a commit to tweak the distance too:

Before	After

WDYT @larsoner ?

[agramfort]

perfect on my side !

[larsoner]

Thanks @GuillaumeFavelier !