Reduce normal tolerance for Weld

[rotu]

Fix #1028
Chose 0.01 as a default tolerance which corresponds to ±arcsin(0.01)/2 = ±0.286°.

Chose 0.05 as a default tolerance which corresponds to ±arcsin(0.05)/2 = ±1.43°.

[rotu]

Oops! I had forgotten to submit this PR.

I think it's a pretty good threshold, but I'm not sure if there's a more principled way to set these values!

[donmccurdy]

Would you be able to share a model that reproduces the bug described in #1028?

One test case I've used before is this:

https://sketchfab.com/3d-models/sculpt-january-day-19-lovecraftian-34ad2501108e4fceb9394f5b816b9f42

gltf-transform optimize in.glb out.glb --verbose

The largest mesh in that scene has just under 900K vertices, and the normal tolerance makes a big difference when welding:

• before: 873,244 → 228,453 (–73.84%) vertices
• after: 873,244 → 802,173 (–8.14%) vertices

At least in this case, the more aggressive weld is preferable.

Aggressive welding is also an important step before simplifying a model. I'm hoping there's a "happy medium" value that can give us a better default here, but perhaps this option just needs to be exposed in the CLI too.

[rotu]

https://wormhole.app/YO9Mr#-YSuDqBOBOySwiBYmNR7tA

The visible weld issue on my model goes away at a normal tolerance of about 0.3. So maybe 0.1 is a good threshold?

[rotu]

Here's a pathological example that's wrecked by optimize, though not by the weld step alone: https://sketchfab.com/3d-models/scifi-hexsphere-cb364832b9994b768dba6245e6b3f51b

edit: a weld threshold of 0.1 is almost enough to prevent the problem.

[donmccurdy]

Thanks! I think 0.05 seems about right. I'd recommend higher tolerance, closer to 0.5, when weld() is used as a pre-process before simplification for LODs. Simplification (for now, at least) cannot do much with disjoint topology.