FastLSM: Fast Lattice Shape Matching for Robust Real-Time Deformation Alec R. Rivers and Doug L. James Cornell University Presenter: 이성호
Prior work: Meshless Deformations Based on Shape Matching
Best fit Rigid Transformation Q: What can be precomputed?
Best fit Rigid Transformation Q: Which is the generalized one, between R and A? Q: Prove the solution of A
Extracting Rotation
Particles position and velocities update
Linear shape matching
Linear shape matching
Quadratic shape matching
Best fit quadratic transformation Q: Could it be precomputed Apq and/or Aqq, and what dimensions they are?
Cluster Based Deformation
FastLSM
Approach
Assumptions Construct regular lattice of cubic cells containing mesh [James et al. 2004]
Computational cost
Naive sum
Bar-plate-cube sum
Constant-time sum
Center of mass
Rotations
Goal positions Q: Prove this. (Recall in [Mueller et al. 2005], p6)
Pseudocode
Fast polar decomposition Cold start (V=I) 1.9 Jacobi sweeps/solution 2500ns/decomposition Warm start (V=V from the last timestep) 0.4 Jacobi sweeps/solution 450ns/decomposition (Refer to p5)
Damping From [Mueller et al. 2006] Apply damping per-region basis (See demo)
Fracture Break by distance [Terzopoulos and Fleischer 1988]
Hardware-accelerated rendering
Per-vertex normals Precompute per each vertex
Constant memory restirction Construct triangle batches
Statistics
Conclusion and Discussion Lattice Shape Matching Fast summation algorithm Allows large deformation Maintaining speed and simplicity Orientation sensitive smoothing Not physically accurate But reasonably plausible and fast Future works Try different particle frameworks Tetrahedral, irregular samplings Adaptive particle resolution