A Continuous 3-D Medial Shape Model with Branching

Transcription

1 A Continuous 3-D Medial Shape Model with Branching Timothy B. Terriberry Guido Gerig

2 Outline Introduction The Generic 3-D Medial Axis Review of Subdivision Surfaces Boundary Reconstruction Edge Curves Branch Curves Fin Points Conclusion

3 Introduction Computing the medial axis is unstable Lots of little branches, different topology for every shape Generating shapes from a fixed medial topology is stable Works great in 2-D In 3-D, have infinitely many singular points infinitely many constraints, but finite number of parameters

4 The Generic 3-D Medial Axis Most points belong to flat sheets with exactly two spokes (healthy left caudate)

5 The Generic 3-D Medial Axis (a) edge curve (b) branch curve (c) fin point (d) 6-junction

6 Catmull-Clark Subdivision Face points added with F-mask Edge points added with E-mask Old vertex positions updated with V-mask

7 Catmull-Clark Subdivision Subdivide once: quadrilaterals Subdivide twice: at most one extraordinary vertex Most patches are B-splines Evaluate rest with (Stam 1999)

8 Boundary Reconstruction Defined by a continuous medial axis m and radius function r. Boundary recovered via: B=m r U 2 U = r± 1 r n Here r is the Riemannian gradient: 1 r u r = [ mu mv ] I m rv [ ]

9 Edge Curves Challenge: Edge boundary condition r = 1 Yushkevich's first solution (23) Implicitly solve for a curve that satisfies it Different curve found for each object Makes statistics on groups of objects hard Yushkevich s second solution (25) Parametrize the model with a potential field ρ and solve a differential equation to recover r. Not enough free parameters for branches

10 Our Approach Start with a B-spline patch

11 Our Approach Interpolate in the v direction

12 Our Approach Convert to an interpolating basis in u

13 Our Approach Replace the curve on the left with a control curve

14 Our Approach Finish interpolating in the u direction

15 Edge Curves Interpolation is given by the B-spline: T 2 3 T 2 3 m, r =[ 1 u u u ] B P B [ 1 v v v ] Change to Catmull-Rom basis C on left: T 2 3 T [ ] [ ] m, r = 1 u u u C P' B 1 v v v, P' =C BP Now r, v and r v, v are independent of P ', j Hold mv, v, r v, v and mu, v fixed and solve for r u, v

16 Edge Curves Split r into two components: r v rv = Gm r v = r u G m r v F m G m E m G m F m 2 Solving r = 1 for r u yields: r u= r v F m± G m r v E m G m F m Gm Note r is always pointing inwards with the + solution v

17 Edge Curves Replace P ', jwith control curve r v Interpolate remaining columns in v r i v =[ P ' i, P ' i, 1 P ' i, 2 P ' i,3 ] BT [ 1 v v 2 v 3 ], i=1 3 Solve for r v r v = 1 r u, v C 11 r 1 v C 12 r 2 v C 13 r 3 v C 1 Finish the interpolation 2 3 r u, v =[ 1 u u u ] C [ r v r 1 v r 2 v r 3 v ] T

18 Edge Curves Final Result

19 Edge Curves Interpolation is asymmetric in u and v Therefore the mesh must have no corners Trivial example: regular grid with corners deleted Extraordinary vertices on the edge are also disallowed All edge vertices will have valence 3, except at fin points

20 Edge Curves Complete model (healthy left ventricle)

21 Edge Curves Complete model (healthy left ventricle)

22 Branch Curves If we can get a succinct expression of the condition to enforce, can use exactly the same method Assume all 3 patches meet on the u= curve (share control points)

23 Branch Curves The tips of - r and U lie in the same plane in all three patches

24 Branch Curves Project everything into this plane Need 2 1 = And i i i 1 =

25 Branch Curves Using right triangles shows r v, i = 1 r 2v Gm cos i (i) u Solving for r produces 1 i i 2 i i 2 r = r v F m cos G m r v E m G m F m Gm i u

26 Fin Points Edge curve meets branch curve Enforce edge condition on patch () as normal, take θ as fixed (1) Use a control curve on r to set θ (2) and θ so they all sum to π ( v,1) and Use another on m to set mu mu( v,2) to bisect them

27 Fin Points 3 more constraints required at fin point itself (detailed in paper) But only 3! Can enforce by adjusting control points using the masks: [ r ' r ' ] [ m' m ' ][ 2 r '' r '' 2 r '' ] Adjusts ru()(,) Adjusts mu()(,) Adjusts ruu()(,)

28 Fin Points Final result

29 Conclusion Demonstrated a new, continuously defined, generative 3-D medial model First such model to support branches Restricted to a mesh without corners Bad for synthetic objects, good for biological Don t explicitly enforce several additional inequality constraints needed to prevent overfolding, etc. Enforce during model fitting process

30 Questions?

31 Extra Slides

32 Catmull-Clark Subdivision Handle extraordinary vertices via eigendecomposition (Stam 1999) P1=A P 1= A P P n Pn =A Pn 1=A P n 1 n= A P n 1= A A P P

33 Catmull Clark Subdivision ¾ of domain is now a B-spline P1=A P P P1= A n Pn =A Pn 1=A P n 1 n= A P n 1= A A P P

34 Catmull-Clark Subdivision No explicit subdivision required Diagonalize A: 1 A=V Λ V n 1 n 1 1 n= A A P = A V Λ V P P 1 V P can be computed once per patch V can be precomputed A

35 Edge Curves Sympathetic overfolding sometimes appears on the right side C2 continuity does not help Choosing P = 12 P P, P = 118 P 38 P helps slow down r without changing the edge Choosing P =P, P = 32 P 12 P gives infinite slowdown, but ill-defined tangent plane, j, j 1, j 2, j 2, j 1, j 1, j 1, j 1, j 2, j 2, j

36 Sympathetic Overfolding Using a 4th-order spline lets us add an extra condition: a well-defined tangent plane via reparameterization