liminchen / OptCuts

OptCuts, a new parameterization algorithm, jointly optimizes arbitrary embeddings for seam quality and distortion. OptCuts requires no parameter tuning; automatically generating mappings that minimize seam-lengths while satisfying user-requested distortion bounds.
http://www.cs.ubc.ca/labs/imager/tr/2018/OptCuts/
MIT License
283 stars 47 forks source link

Hessian is semi-definite #8

Open liminchen opened 5 years ago

liminchen commented 5 years ago

Sometimes the rounding error might cause a linear solver that requires strict positive definiteness to crash. However the Eigen::SimplicialLDLT we use by default works just fine.

A simple trick that one can do is to add an epsilon * identitiyMatrix to each of the projected element Hessian, so that the rounding error won't make the global Hessian indefinite. Similarly, one can also blend the global Hessian with an epsilon scaled lumped mass to help with this issue.