We have two ideas for how to do the regularization.

  1. Use approach from the PolyVector paper, Section 4.2 Angle-bounded fields
  2. Use Misha’s approach with iterative re-weighting

I’ve tried 1., using fitting to a corrected field, but there are problems

  • The local (correction) step breaks the alignment to Sobel tangents – we don’t want to rotate both directions away from the bisector, but rather only the one that is ‘further’ from the tangent direction (makes sense?)
  • For higher fitting weights, the result is over-regularized and not really aligned to the corrected field – seems like a bug? (screeenshots) The result is the same if I disable tangent alignment, and set the smoothness weight to almost zero.

Approach 2. does not work well either yet.

The annoying thing is that we can have additional problems when interpolating the frame field from vertices to faces: