Stability of boundary measures

Transcription

1 Stability of boundary measures F. Chazal D. Cohen-Steiner Q. Mérigot INRIA Saclay - Ile de France LIX, January 2008

2 Point cloud geometry Given a set of points sampled near an unknown shape, can we infer the geometry of that shape?

3 Detecting singularities the volume of a cell is very sensitive to perturbation but if one consider the union of Voronoï cells whose site is contained in a given ball...

4 Detecting singularities the volume of a cell is very sensitive to perturbation but if one consider the union of Voronoï cells whose site is contained in a given ball...

5 Detecting singularities the volume of a cell is very sensitive to perturbation but if one consider the union of Voronoï cells whose site is contained in a given ball...

6 Detecting singularities the volume of a cell is very sensitive to perturbation but if one consider the union of Voronoï cells whose site is contained in a given ball...

7 Projection on a compact set Definition The projection p K : R n K R n maps any point x R n to its closest point in K. It is defined outside of the medial axis of K.

8 Boundary measure Definition For E R n, the boundary measure µ K,E is defined as follows : B K, µ K,E (B) = vol n ({x E p K (x) B}) that is, the n-volume of the part of E that projects on B. measure supported in K contains a lot of geometric information about K

9 Smooth object Let K R n be an n-dimensional object with smooth boundary. The smallest distance between K and its medial axis is called reach(k). Take E = K r {x R n ; d(x, K) r}, assuming r < reach(k).

10 Smooth object Tube formula (Steiner, Weyl, Federer) If r < reach(k) : vol n (K r ) = vol n (K) + n k=1 const(n, k)[ σ k 1 ] r k K

11 Smooth object Tube formula (Steiner, Weyl, Federer) If r < reach(k), for B K : n µ K,K r (B) = vol n (B) }{{} + k=1 const(n, k) [ σ k 1 ] B K }{{} r k Φ n K (B) Φn k K (B) The Φ i K are the (signed) curvature measures of K. If K is d-dimensional, they vanish identically for i > d. Φ 0 K (K) is the Euler characteristic of K. They are intrinsic, i.e. they do not depend on the embedding.

12 Convex polyhedron rda θ r 2 dl Ω r 3 the boundary measure of a convex polyhedron K can be decomposed as a sum : µ K,K r (B) = n k=0 const(n, k) Φ n k(b) r k. the curvature measure Φ i K is the i-dimensional measure supported on the i-skeleton of K whose density is the local external dihedral angle.

13 The boundary measure of a point cloud the boundary measure is a sum of weighted Dirac masses : µ C,C r = i vol n (Vor(x i ) C r )δ xi

14 Stability and computation of boundary measure 1 is it practically feasible to compute the boundary measure of a point cloud C R d? 2 if C is a good approximation of K (ie dense enough and without too much noise), does the boundary measure µ C,C r carry approximately the same geometric information as µ K,K r?

15 The boundary measure of a point cloud Here, the volumes of the Voronoï cells are evaluated using a Monte-Carlo method. Cost scales linearly with ambient dimension. Approximation error does not depend on the ambient dimension.

16 The boundary measure of a point cloud Here, the volumes of the Voronoï cells are evaluated using a Monte-Carlo method. Cost scales linearly with ambient dimension. Approximation error does not depend on the ambient dimension.

17 Stability Assume point cloud C samples compact K well, e.g. d H (K, C) ε. This means that C K ε and K C ε Are µ K,E and µ C,E close? In which sense?

18 Wasserstein distance Assume measures µ and ν are discrete : µ = i c iδ xi, ν = j d jδ yj we suppose that mass(µ) = mass(ν) a transport plan between is a set of nonnegative coefficients p ij specifying the amount of mass which is transported from x i to y j, with p ij = d j and i j the cost of a transport plan is C(p) = ij x i y j p ij W (µ, ν) = inf p C(p) p ij = c i

19 Kantorovich-Rubinstein theorem Definition Let µ and ν be two measures on R d having the same total mass µ(r d ) = ν(r d ). W (µ, ν) = inf E[d(X, Y )] X,Y where the infimum is taken on all pairs of R d -valued random variables X and Y whose law are µ and ν respectively. Theorem For two measures µ and ν with common finite mass and bounded support, W (µ, ν) = sup f f dµ f dν where the sup is taken over all 1-Lipschitz functions R n R.

20 Wasserstein distance between boundary measures dx E K p K (dx) K p K (dx) We consider the following transport plan : the element of mass p K (x)dx coming from an element of mass dx at x E will be transported to p K (x)dx. the total cost of this transport is : p K (x) p K (x) dx = p K p K L 1 (E) E W (µ K,E, µ K,E ) pk p K L 1 (E)

21 A L 1 stability theorem for projections Theorem If E is an open set of R n with rectifiable boundary, and K and K are two close enough compact subsets : p K p K L 1 (E) := p K p K E C(n)[vol n (E) + diam(k)vol n 1 ( E)] R K d H (K, K ) where R K = sup x E d(x, K). 1 close enough means that d H (K, K ) does not exceed min(r K, diam(k), diam(k) 2 /R K ) 2 C(n) = O( n)

22 Ingredients of proof Lemma Function v K : x x 2 d 2 K (x) is convex and v K = 2p K almost everywhere. Lemma If f, g : E R are convex and k = diam( f (E) g(e)) then f g L 1 (E) C(n) [vol n (E) + k vol n 1 ( E)] f g 1/2 + integral geometry arguments. + C(n) vol n 1 ( E) f g

23 Stability of boundary measures Theorem If K is a fixed compact set, and E an open set with smooth boundary, then W(µ K,E, µ K,E ) C(n, E, K) d H (K, K ) 1/2 as soon as K is close enough to K. A similar result holds for µ K,K r and µ K,K r.

24 Estimating curvature measures 1 for any K with positive reach, there exists measures Φ K,i such that for r < reach(k), µ K,r (B) = n (B)r i i=1 Φn i K 2 can be computed knowing only the boundary measures for n + 1 values r 0 <... < r n : denote the result by Φ (r) K,i. Corollary If reach(k) > r n and K is close to K, there is a constant C(K, n, (r)) such that ( ) d bl Φ K,i, Φ (r) K,i C(K, n, (r)) d H (K, K ) 1/2 C 0 (Hausdorff) closeness implies closeness of differential properties at a given scale.

25 Computing boundary measures algorithm Input : a point cloud C, a scalar r, a number N Output : an approximation of µ C,C r in the form 1 N n(pi )δ pi while k N I. Choose a random point X with probability distribution 1 H d (C r ) Hd C r II. Finds its closest point p i in the cloud C, add 1 to n(p i ) end while III. Multiply each n(p i ) by H d (C r ). Hoeffding s inequality : P [d bl (µ N, µ) ε] 2 exp ( ln(16/ε)n (K, ε/16) Nε 2 /2 )

26 Directions Replace volumes of Voronoï cells by their covariance matrices. This gives a tensor-valued measure. small/large eigenvalues tangent/normal space principal curvatures/directions?

27 Discussion 1 The boundary measure and its tensor version encode much of the geometry of a compact set. 2 these measures depend continuously on the compact set for the Hausdorff distance. 3 what can we do when the underlying shape has zero reach? 4 what happens if we replace nearest neighbors by approximate nearest neighbors? 5 how can we deal with outliers?