Adapt the spatial relations of a multi-body key pose to bodies of other
proportions by an interaction mesh (Ho, Komura and Tai 2010, "Spatial
relationship preserving character motion adaptation", ACM TOG 29(4),
article 33): a volumetric mesh over every joint and contact point of all
bodies at the key pose records how they lie among each other, and the
adapted pose is the one that deforms that mesh least, by its Laplacian
coordinates, while each body keeps its own bone lengths and every contact
its pairing, two points coinciding or held at a fixed offset.
Mesh. The Delaunay tetrahedralization of the points
(tetrahedralizeAutoMovieDelaunay, Bowyer-Watson on Qhull's joggled
input). Its edges are the points' neighbourhoods.
Laplacian coordinates. Each point's offset from the inverse-distance
weighted mean of its neighbours, excluding bone edges so deformation
does not oppose the bone lengths being changed (Ho et al. 2010,
equation 3 and section 5.1; HumanInteractionMeshLaplacian).
Solve.min sum_i |L(p)_i - d_i|^2 / 2 subject to each bone's
length and each contact's pairing, by the l1 exact penalty's sequential
convex optimisation with a trust region and a steered penalty
(HumanSequentialConvex; Nocedal and Wright 2006, "Numerical
Optimization", 2nd ed., sections 17.2 and 18.5; Byrd, Nocedal and Waltz
2008): each convex model linearizes a bone |p_i - p_j| = l along its
current direction and holds the Laplacian's quadratic exactly, solved by
HumanQuadraticProgram within the region; a step is taken only where
the merit falls, so a large change of proportions is approached in
steps the linearization holds over.
Sides. Each tetrahedron not flat on the key pose keeps its
orientation (HumanInteractionMeshSide), an inequality on its signed
volume linearised as a hard row in each model: a point that crossed the face across from it would have passed
through the bodies the mesh spans, the interpenetration the paper's
mesh is built to reduce. The strictly positive determinant is the
nonlinear solve's domain: a trial outside it has infinite merit and is
rejected by the trust region, as an inversion barrier excludes flips
(Schueller et al. 2013, "Locally Injective Mappings", section 3).
Ho et al.'s full text (author upload: https://www.researchgate.net/publication/220183715_Spatial_Relationship_Preserving_Character_Motion_Adaptation)
uses collision constraints on bounding volumes during iterative
morphing. This key-pose adaptation instead keeps every recorded
tetrahedron's orientation as a hard domain, with the physical contact
accuracy of the whole-body fit (HumanBodyFitMerit.TOLERANCE), stricter
than #3438's 10 mm bone and contact acceptance.
Adapt the spatial relations of a multi-body key pose to bodies of other proportions by an interaction mesh (Ho, Komura and Tai 2010, "Spatial relationship preserving character motion adaptation", ACM TOG 29(4), article 33): a volumetric mesh over every joint and contact point of all bodies at the key pose records how they lie among each other, and the adapted pose is the one that deforms that mesh least, by its Laplacian coordinates, while each body keeps its own bone lengths and every contact its pairing, two points coinciding or held at a fixed offset.
Mesh. The Delaunay tetrahedralization of the points (
tetrahedralizeAutoMovieDelaunay, Bowyer-Watson on Qhull's joggled input). Its edges are the points' neighbourhoods.Laplacian coordinates. Each point's offset from the inverse-distance weighted mean of its neighbours, excluding bone edges so deformation does not oppose the bone lengths being changed (Ho et al. 2010, equation 3 and section 5.1;
HumanInteractionMeshLaplacian).Solve.
min sum_i |L(p)_i - d_i|^2 / 2subject to each bone's length and each contact's pairing, by the l1 exact penalty's sequential convex optimisation with a trust region and a steered penalty (HumanSequentialConvex; Nocedal and Wright 2006, "Numerical Optimization", 2nd ed., sections 17.2 and 18.5; Byrd, Nocedal and Waltz 2008): each convex model linearizes a bone|p_i - p_j| = lalong its current direction and holds the Laplacian's quadratic exactly, solved byHumanQuadraticProgramwithin the region; a step is taken only where the merit falls, so a large change of proportions is approached in steps the linearization holds over.Sides. Each tetrahedron not flat on the key pose keeps its orientation (
HumanInteractionMeshSide), an inequality on its signed volume linearised as a hard row in each model: a point that crossed the face across from it would have passed through the bodies the mesh spans, the interpenetration the paper's mesh is built to reduce. The strictly positive determinant is the nonlinear solve's domain: a trial outside it has infinite merit and is rejected by the trust region, as an inversion barrier excludes flips (Schueller et al. 2013, "Locally Injective Mappings", section 3).Ho et al.'s full text (author upload: https://www.researchgate.net/publication/220183715_Spatial_Relationship_Preserving_Character_Motion_Adaptation) uses collision constraints on bounding volumes during iterative morphing. This key-pose adaptation instead keeps every recorded tetrahedron's orientation as a hard domain, with the physical contact accuracy of the whole-body fit (
HumanBodyFitMerit.TOLERANCE), stricter than #3438's 10 mm bone and contact acceptance.