Abstract
In this article, we introduce a family of elastic metrics on the space of parametrized surfaces in 3D space using a corresponding family of metrics on the space of vector-valued one-forms. We provide a numerical framework for the computation of geodesics with respect to these metrics. The family of metrics is invariant under rigid motions and reparametrizations; hence, it induces a metric on the “shape space” of surfaces. This new class of metrics generalizes a previously studied family of elastic metrics and includes in particular the Square Root Normal Field (SRNF) metric, which has been proven successful in various applications. We demonstrate our framework by showing several examples of geodesics and compare our results with earlier results obtained from the SRNF framework.
Similar content being viewed by others
Notes
See the article [20] for an example of a path of closed surfaces that connects two distinct shapes, such that the whole path has the same SRNF.
References
Abe, K., Erbacher, J.: Isometric immersions with the same gauss map. Math. Ann. 215(3), 197–201 (1975)
Allen, B., Curless, B., Popović, Z.: The space of human body shapes: reconstruction and parameterization from range scans. ACM Trans. Graph. 22(3), 587–594 (2003)
Bauer, M., Bruveris, M., Charon, N., Møller-Andersen, J.: A relaxed approach for curve matching with elastic metrics. To appear in ESAIM: COCV (2018)
Bauer, M., Bruveris, M., Michor, P.W.: Overview of the geometries of shape spaces and diffeomorphism groups. J. Math. Imaging Vis. 50(1–2), 60–97 (2014)
Bauer, M., Harms, P., Michor, P.W.: Sobolev metrics on shape space of surfaces. J. Geom. Mech. 3(4), 389–438 (2011)
Bauer, M., Harms, P., Michor, P.W.: Sobolev metrics on shape space, ii: weighted sobolev metrics and almost local metrics. J. Geom. Mech. 4(4), 365–383 (2012)
Bauer, M., Klassen, E., Preston, S.C., Su, Z.: A diffeomorphism-invariant metric on the space of vector-valued one-forms. arXiv:1812.10867 (2018)
Bronstein, A.M., Bronstein, M.M., Kimmel, R.: Numerical Geometry of Non-Rigid Shapes. Springer, Berlin (2008)
Bruveris, M., Michor, P.W., Mumford, D.: Geodesic completeness for sobolev metrics on the space of immersed plane curves. In: Guralnick, R. (ed.) Forum of Mathematics, Sigma, vol. 2. Cambridge University Press (2014)
Celledoni, E., Eslitzbichler, M., Schmeding, A.: Shape analysis on Lie groups with applications in computer animation. J. Geom. Mech. 8(3), 273–304 (2015)
Cervera, V., Mascaro, F., Michor, P.W.: The action of the diffeomorphism group on the space of immersions. Differ. Geom. Appl. 1(4), 391–401 (1991)
Cury, C., Glaunes, J.A., Colliot, O.: Template estimation for large database: a diffeomorphic iterative centroid method using currents. In: International Conference on Geometric Science of Information, pp. 103–111. Springer (2013)
Fletcher, R.: Practical Methods of Optimization. Wiley, Amsterdam (2013)
Grenander, U., Miller, M.I.: Computational anatomy: an emerging discipline. Q. Appl. Math. 56(4), 617–694 (1998)
Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., Seidel, H.P.: A statistical model of human pose and body shape. In: Hauser, H., Benes, B (eds.) Computer Graphics Forum, vol. 28, pp. 337–346. Wiley Online Library (2009)
Heeren, B., Rumpf, M., Wardetzky, M., Wirth, B.: Time-discrete geodesics in the space of shells. In: Hauser, H., Benes, B (eds.) Computer Graphics Forum, vol. 31, pp. 1755–1764. Wiley Online Library (2012)
Jermyn, I., Kurtek, S., Laga, H., Srivastava, A.: Elastic shape analysis of three-dimensional objects. Synth. Lect. Comput. Vis. 7, 1–185 (2017)
Jermyn, I.H., Kurtek, S., Klassen, E., Srivastava, A.: Elastic shape matching of parameterized surfaces using square root normal fields. Comput. Vis. ECCV 2012, 804–817 (2012)
Kilian, M., Mitra, N.J., Pottmann, H.: Geometric modeling in shape space. In: ACM Transactions on Graphics (TOG), vol. 26, p. 64. ACM (2007)
Klassen, E., Michor, P.W.: Closed surfaces with different shapes that are indistinguishable by the srnf. arXiv:1910.10804. (2019)
Kühnel, L., Sommer, S., Arnaudon, A.: Differential geometry and stochastic dynamics with deep learning numerics. Appl. Math. Comput. 356, 411–437 (2019)
Kurtek, S., Klassen, E., Ding, Z., Jacobson, W., Jacobson, S., Avison, J.L.J.M., Srivastava, A.: Parameterization-invariant shape comparisons of anatomical surfaces. IEEE Trans. Med. Imaging 30(3), 849–858 (2011)
Kurtek, S., Needham, T.: Simplifying transforms for general elastic metrics on the space of plane curves. arXiv preprint arXiv:1803.10894 (2018)
Kurtek, S., Srivastava, A., Klassen, E., Laga, H.: Landmark-guided elastic shape analysis of spherically-parameterized surfaces. In: Hauser, H., Benes, B (eds.) Computer Graphics Forum, vol. 32, pp. 429–438. Wiley Online Library (2013)
Laga, H., Xie, Q., Jermyn, I.H., Srivastava, A.: Numerical inversion of srnf maps for elastic shape analysis of genus-zero surfaces. IEEE Trans. Pattern Anal. Mach. Intell. 39(12), 2451–2464 (2017)
Mio, W., Srivastava, A., Joshi, S.: On shape of plane elastic curves. Int. J. Comput. Vis. 73(3), 307–324 (2007)
Praun, E., Hoppe, H.: Spherical parametrization and remeshing. In: ACM Transactions on Graphics (TOG), vol. 22, pp. 340–349. ACM (2003)
Srivastava, A., Klassen, E., Joshi, S.H., Jermyn, I.H.: Shape analysis of elastic curves in euclidean spaces. IEEE Trans. Pattern Anal. Mach. Intell. 33(7), 1415–1428 (2011)
Srivastava, A., Klassen, E.P.: Functional and Shape Data Analysis. Springer, Berlin (2016)
Su, J., Kurtek, S., Klassen, E., Srivastava, A.: Statistical analysis of trajectories on Riemannian manifolds: bird migration, hurricane tracking and video surveillance. Ann. Appl. Stat. 8(1), 530–552 (2014)
Su, Z., Klassen, E., Bauer, M.: Comparing curves in homogeneous spaces. In: Slovák, J. (ed.) Differential Geometry and its Applications, vol. 60, pp. 9–32. Elsevier (2018)
Tumpach, A.B.: Gauge invariance of degenerate riemannian metrics. Notices of the AMS 63(4), 342–350 (2016)
Tumpach, A.B., Drira, H., Daoudi, M., Srivastava, A.: Gauge invariant framework for shape analysis of surfaces. IEEE Trans. Pattern Anal. Mach. Intell. 38(1), 46–59 (2015)
Younes, L.: Computable elastic distances between shapes. SIAM J. Appl. Math. 58(2), 565–586 (1998)
Younes, L., Michor, P.W., Shah, J.M., Mumford, D.B.: A metric on shape space with explicit geodesics. Rendiconti Lincei-Matematica e Applicazioni 19(1), 25–57 (2008)
Zhang, Z., Su, J., Klassen, E., Le, H., Srivastava, A.: Video-based action recognition using rateinvariant analysis of covariance trajectories. arXiv:1503.06699 (2015)
Acknowledgements
The authors thank Anuj Srivastava and all the members of the Florida State statistical shape analysis group for helpful discussions during the preparation of this manuscript. In addition, we are grateful to Sebastian Kurtek and Alice Barbara Tumpach for discussion about the implementation of the minimization over the diffeomorphism group.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
M. Bauer was partially supported by NSF-Grant 1912037 (collaborative research in connection with NSF-Grant 1912030). S. C. Preston was partially supported by Simons Foundation Collaboration Grant for Mathematicians No. 318969. E. Klassen was partially supported by Simons Foundation Collaboration Grant for Mathematicians No. 317865.
Appendices
Appendix A: The Geodesic Equation
In the following, we give the geodesic equation on the space of immersions \({\text {Imm}}(M, {\mathbb {R}}^3)\) with respect to the pullback of the metric (3) on the space of 1-forms. In this “Appendix,” we will assume that the domain M is a compact orientable surface without boundary, because we will need to use the Hodge decomposition. We will view \(\alpha \in \varOmega ^1_+(M,{\mathbb {R}}^3)\) as a vector-valued 1-form with components \(( \alpha ^1, \alpha ^2,\alpha ^3)\), where each \(\alpha ^i\) is a 1-form on M in the usual sense. Then, metric (3) can be rewritten as
where \(\xi = (\xi ^1, \xi ^2,\xi ^3)\in T_{\alpha }\varOmega ^1_+(M,{\mathbb {R}}^3)\), \(\varLambda _{\alpha } = (\alpha ^T\alpha )^{-1}\) is the induced Riemannian metric on 1-forms on M, and \(\varphi _{\alpha } = \sqrt{\det {(\alpha ^T\alpha )}}\) is the induced volume form on M. As such, all computations can be done one component at a time.
If \(F = (f^1, f^2, f^3)\) is a vector-valued function with each \(f^i:M\rightarrow {\mathbb {R}}\) real-valued, then \(\beta = dF\) is a vector-valued 1-form with \(\beta ^i = df^i\). The Hodge decomposition tells us that every 1-form \(\xi \) may be written as
where \(\delta \gamma = 0\) and \(\delta :\varOmega ^1(M, {\mathbb {R}})\rightarrow C^{\infty }(M)\) is the codifferential operator.
The space \({\text {Imm}}(M,{\mathbb {R}}^3)\) is formally a submanifold of \(\varOmega ^1_+(M,{\mathbb {R}}^3)\), and thus by general submanifold geometry, we know that the geodesic equation on \({\text {Imm}}(M,{\mathbb {R}}^3)\) will be given by
Since \(\delta \gamma = 0\), we know that \(\star \gamma \) is an exact form, where \(\star \) denotes the Hodge star operator. Then, there is a function p, unique up to a constant, such that \(dp = \star \gamma \). We obtain
In coordinates (u, v) on M, the operator \(\star d\) is given by:
where \(\varphi \) is the volume form on M. From the geodesic equation on \(\varOmega ^1_+(M,{\mathbb {R}}^3)\) with respect to the metric (3), as calculated in our previous paper [7], we know that the covariant derivative is given by
Since \(d\alpha _{tt} = 0\), we obtain
Let \(L = \alpha _t \alpha ^+\). Applying the Hodge star operator to both sides, we see that \(\varPhi \) is a geodesic on \({\text {Imm}}(M, {\mathbb {R}}^3)\) if and only if we have
where \(\varOmega = LL^T + L^2 - L^T L +\frac{1}{2}{{\,\mathrm{tr}\,}}(L^TL) - {{\,\mathrm{tr}\,}}(L)L\). Here, we emphasize that p and \(\varPhi \) are actually vector-valued functions, so these computations are done componentwise for each \(i\in \{1,2,3\}\). In other words, we have
Appendix B: Proofs
Proof of Theorem 2
In the following, we prove the correspondence between our split metric on the space \(\varOmega _+^1(M,{\mathbb {R}}^3)\) and the SRNF metric on the space of surfaces. Let \(M_+(3,2)\) be the space of \(3\times 2\) matrices with rank 2. Using the pointwise property of our metric, we will focus on the corresponding split metric on the matrix space \(M_+(3,2)\). For \(a\in M_+(3,2)\) and \(v\in T_mM_+(3,2)\), we decompose v into four parts
where
The corresponding split metric on \(M_+(3,2)\) is then of the form:
Now consider the projection \(\pi :M_+(3,2)\rightarrow {\text {Sym}}_+(2), \ a\mapsto a^Ta\). This projection is a Riemannian submersion, where \(M_+(3,2)\) carries metric (13) with choices of constants (1, 1, 1, 1) and the space \({\text {Sym}}_+(2)\) is equipped with the following metric:
The horizontal bundle with respect to the projection \(\pi \) is given by
and the differential \(d\pi \) induces an isometry
It is easy to check that \(v_m\) and \(\frac{1}{2}{{\,\mathrm{tr}\,}}(a^+v)a\) are horizontal vectors.
Let \(g = \pi (a) =a^Ta\). By computation, we have
and
Therefore, the first term in (13) becomes
and the second term becomes
For the third term in (13), we consider the corresponding unit normal map on the space of matrices given by
where \(a_1\) and \(a_2\) are the first and the second columns of a, respectively. For any tangent vector \(u = \begin{pmatrix} u_1&u_2 \end{pmatrix}\) at a, the differential of n at a is
It is easy to check that \(aa^+v\) is in the kernel of the differential \(dn_a\), i.e.,
Note that \({{\,\mathrm{tr}\,}}(a^+v^{\perp }) = 0\), \(aa^+a_1 = a_1\) and \(aa^+a_2 = a_2\). Using the following identity for three-dimensional vectors b, c, d, e:
and the formula for the inverse of \(a^Ta\):
we have
where \(v^{\perp }_1, v^{\perp }_2\) are the first and the second columns of \(v^{\perp }\) and \(v_1, v_2\) are the first and the second columns of v, respectively. It follows that
that is,
Therefore, the split metric (13) on \(M_+(3,2)\) can be rewritten as
Now it is easy to see that the first three terms give rise to the formula of the full elastic metric on the space of surfaces for \(A = {\mathfrak {a}}/4, B = ({\mathfrak {b}}-{\mathfrak {a}})/8, C = {\mathfrak {c}}\) and the SRNF metric corresponds to the split metric (4) with constants \((0 ,\frac{1}{2}, 1, 0)\). \(\square \)
Proof of Theorem 3
We first perform the computation in spherical coordinates \((\theta , \phi )\in [0, 2\pi ]\times [0, \pi ]\). Denote the usual spherical coordinate orthonormal basis by
We have the following formulas for the partial derivatives:
We also note that the covariant derivatives are given by
Write
For a real parameter t, we consider the following map \(W:S^2 \rightarrow {\mathbb {R}}^3\) given in coordinates by
Then, \(\eta = W/|W|\).
Note that in order for \(\eta \) to be a diffeomorphism, we require that the Jacobian determinant be nonzero; it is given by
Observe that
Since \(\eta _{\phi }\) and \(\eta _{\theta }\) are both perpendicular to W, we know that \(\eta _{\phi }\times \eta _{\theta }\) is parallel to W; thus, we obtain the formula
using the cyclic invariance of the scalar triple product and the fact that \(W\times P_{W^{\perp }}(V) = W\times V\) for any vector V.
Since \(W = e_1+tU\) for the vector field \(U = ue_2+ve_3\), it is straightforward to compute using (14)-(15) that
Let \(m_1 = u_{\phi }, m_2 = v_{\phi }, m_3 = \frac{u_{\theta }-v\cos {\phi }}{\sin {\phi }}, m_4= \frac{v_{\theta }+u\cos {\phi }}{\sin {\phi }}\). We have by (15) that
which we abbreviate by
Thus, the Jacobian is nonzero if and only if the following determinant is nonzero:
Expanding the determinant (16) along the first column, then it is given by
where \(J = \begin{pmatrix} 0 &{} -1\\ 1 &{} 0 \end{pmatrix}\).
Let \({\overline{M}} = \tfrac{1}{2} (M+M^T)\) denote the symmetrization of M, and let \(\lambda _1\le \lambda _2\) denote the real eigenvalues of \({\overline{M}}\). Then, \({{\,\mathrm{tr}\,}}{M} = {{\,\mathrm{tr}\,}}{{\overline{M}}}\) and \(\det {M} = \det {{\overline{M}}} + \tfrac{1}{4} (m_2-m_3)^2\), so that
Since J is a rotation, we have
Thus,
For sufficiently small t, we know \((1+\lambda _1t)\) is positive, and since \(\lambda _1\le \lambda _2\), we obtain
Thus, \(1+\lambda _1t>0\) is a sufficient condition for positivity of D, and this happens as long as \(|t|< \frac{1}{|\lambda _1|}\). It is easy to compute that
In particular, \(m_1+m_4 = {{\,\mathrm{tr}\,}}{(\nabla U)} = {\text {div}}{U}\), and by the divergence theorem, we know the integral of \(m_1+m_4\) over \(S^2\) is zero, and in particular \(m_1+m_4\) is either identically zero or changes sign on \(S^2\). Since t is nonnegative, we therefore are concerned about the most negative that \(\lambda _1(x)\) can be:
which is equivalent to
This is clearly (9). \(\square \)
Rights and permissions
About this article
Cite this article
Su, Z., Bauer, M., Preston, S.C. et al. Shape Analysis of Surfaces Using General Elastic Metrics. J Math Imaging Vis 62, 1087–1106 (2020). https://doi.org/10.1007/s10851-020-00959-4
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10851-020-00959-4