World edges and barycentric interpolations missing in the mesh graph nets code.

[Puruaero]

Hi Team,

I am trying to work with reproducing the mesh graph nets idea with sphere dynamic dataset provided by you. Currently my 1 step RMSE losses are not converging at all even for a single trajectory. I have tried to have my own piece for bringing in the compute of world edges and some interpolations due to changing mesh vertices. I am just trying to train for RMSE in the sense that I dont need to work with remeshing (reason being for my problem statement does not need to remesh which I plan to apply mesh graph nets idea for). I am wondering if its possible to include world edge compute code and barycentric interpolations of mesh's world positions to target source positions. Currently, it seems like both of them are missing from the code.

Best Regards,
Purushottam

[Puruaero]

Just an update, the barycentric interpolation I was doing earlier, was all being done with target's mesh pos as the target. Changing it to current mesh's mesh pos i.e interpolating target world pos (use target mesh pos as source) and previous world pos (using previous mesh pos as source) as mentioned in the paper (missed earlier) to current mesh solved the problem of unconverging losses.

Still, looking forward for original implementations of world edges, interpolations and remeshing (included in code) if possible.