Cage-Based Deformations

• Cage-based deformation is a geometric processing paradigm that embeds a target shape within a coarse cage, using barycentric coordinates for smooth, shape-preserving transformations.
• It leverages advanced methods such as Green, harmonic, and Poisson coordinates, and incorporates Bézier patches to accurately model high-curvature boundaries.
• Efficient computation is achieved through adaptive quadrature and global projection techniques, ensuring exact linear reproduction and robust deformation fidelity.

Cage-based deformation is a geometric processing paradigm in which a target shape (mesh, point cloud, or volume) is embedded within a coarse surrounding boundary called a "cage," and the positions and/or normals of the cage boundary are manipulated to induce a smooth, often globally coherent, deformation of the interior. The interior deformation is typically realized via generalized barycentric coordinate functions that interpolate the influence of cage handles throughout the enclosed domain. Cage-based deformation is fundamental in computer graphics, vision, mesh editing, numerical PDEs, and more recently, adversarial and neural shape optimization.

1. Fundamentals of Cage-Based Deformation

Cage-based deformation frameworks characterize each interior point by a set of coordinate functions $\{\phi_i,\psi_j\}$ with respect to cage vertices $v_i$ and (optionally) cage face normals $n_j$. The deformation map is given by

$$\eta \;=\; \sum_{i=1}^N \phi_i(\eta)v_i \;+\; \sum_{j=1}^M \psi_j(\eta)n_j,$$

and upon cage displacement, the target configuration uses deformed vertices and normals $({\tilde v}_i,{\tilde n}_j)$ (Xiao et al., 23 Jan 2025). Key coordinate systems include Mean Value Coordinates (MVC), harmonic and Poisson coordinates, and Green coordinates—the latter introduced to incorporate normals for improved conformality and shape preservation (Lipman et al. 2008).

Classical cages use triangle or quad meshes, but these planar structures have limited ability to compactly describe curved or high-curvature boundaries without excessive subdivision, since quad-based ruled surfaces yield straight isoparametric curves and rapidly lose local control for curved deformations (Xiao et al., 23 Jan 2025).

2. High-Order and Bézier-Patch Cages

To remedy the deficiencies of linear cages, recent works introduce cages built from high-order polynomial or tensor-product Bézier patches. A tensor-product Bézier patch of degree $(m,n)$ is parameterized by

$$b(u,v) = \sum_{i=0}^m \sum_{j=0}^n B^m_i(u)B^n_j(v)\,b_{ij},\quad 0 \leq u,v \leq 1,$$

with $B^m_i, B^n_j$ Bernstein basis terms and $b_{ij}$ forming the control net (Xiao et al., 23 Jan 2025). Normals at each control point $N_{ij}$ are constructed from local cross-products and then smoothly interpolated across each patch. This curved boundary construction allows the cage to match high-curvature boundaries using only a compact set of control points.

Green coordinates for Bézier-patch cages are derived by applying Green’s third identity over each smooth patch, producing for every interior point $\eta$ a set of coefficients $\phi^{ij}_Q(\eta)$ (vertex influence) and $\psi^{ij}_Q(\eta)$ (normal influence) for each control point pair of each patch $Q$: $$\phi_Q^{ij}(\eta) = \iint_{[0,1]^2} \frac{\lambda^{ij}(u,v)\,(b(u,v)-\eta)\cdot N(u,v)}{4\pi\|b(u,v)-\eta\|^3} du\,dv,$$

$$\psi_Q^{ij}(\eta) = \iint_{[0,1]^2} \frac{\lambda^{ij}(u,v)}{4\pi\|b(u,v)-\eta\|} du\,dv,$$

with $\lambda^{ij}$ the local Bernstein weights (Xiao et al., 23 Jan 2025). This achieves shape-preserving, quasi-conformal deformation with reproducibility of affine motions.

3. Coordinate Computation and Linear Reproduction

No closed-form exists for most Green coordinate integrals over curved patches. Practical computation relies on discretized Riemann summation—adaptive triangulation in $(u,v)$ is performed with greater density near the projection of the interior sample point, and Dirichlet and Neumann terms are summed over triangles using solid angle and closed-form Gaussian integrals (Xiao et al., 23 Jan 2025).

Approximate numerics may violate linear reproduction (i.e., the property that undeformed cages map interior points to their original positions). This is corrected via a global projection technique: the coordinate vector $\Phi$ is projected onto the subspace defined by the linear system $A\Phi = \eta$, enforcing exact reproduction,

$$\Phi = \overline\Phi + A^T(AA^T)^{-1}(\eta - A \overline\Phi),$$

where $\overline\Phi$ is the initial Riemann sum and $A$ collects current control positions and normals. This global step is computationally efficient as $AA^T$ is low-dimensional.

4. Algorithmic Pipeline and Practical Complexity

A typical pipeline for high-order cage deformation using Bézier patches proceeds as follows (Xiao et al., 23 Jan 2025):

1. Precompute Bézier patch control points and per-control-point normals.
2. For each interior mesh vertex, invert to $(u,v)$ coordinates within each patch, tessellate the patch domain adaptively, compute $\{\phi,\psi\}$ via Riemann sums.
3. Project coordinates to ensure linear reproduction.
4. Deform interior points via $\tilde\eta = \sum(\phi\,\tilde b + \psi\,\tilde N)$. The dominant per-point cost is in coordinate computation via adaptive quadrature. For tens of thousands of mesh points and a cage of $30-50$ patches, deformation precomputation is typically performed in seconds to tens of seconds on a multi-core CPU.

5. Shape-Preservation and High-Curvature Fidelity

Including normals via the Neumann term in Green coordinates yields quasi-conformal, detail-preserving deformation even at large curvatures. Curved Bézier cages accurately trace complex shape boundaries with minimal control points, whereas traditional quad/triangle cages would require dense subdivision, leading to facetization and local errors (Xiao et al., 23 Jan 2025).

Comparative experiments show that Bézier Green cage deformation produces smooth, segmentation-free results in high-curvature scenarios (vase handle, cactus, hydrant) while linear or quad methods require $10\times$ more elements to avoid visual artifacts. For instance, a $3\times3$ Bézier patch can match a complex curve which would require upwards of $18$ triangles per patch for equivalent quality in Lipman et al. (2008), or $9$ quads per patch in Thiery et al. (2018/2022).

Method	Cactus (98k vertices, 34 patches, time)	Bar (230k vertices, 6 patches, time)
GC-18	5.00 s	2.12 s
QMVC-9	102.9 s	41.6 s
QGC-9	124.6 s	50.0 s
Bézier-GC	27.8 s	8.5 s

Although bicubic patches involve more integral computations, the cost remains competitive with GC methods using heavy triangulation, while offering significantly improved fidelity with compact control nets (Xiao et al., 23 Jan 2025).

6. Theoretical Properties and Generalizations

Green coordinate deformation using high-order polynomial or Bézier cages exhibits the following properties:

• Exact affine invariance: any affine motion of the cage yields the same affine motion of all interior points.
• Harmonicity and conformality: interior mapping is locally conformal in $\mathbb R^2$ and quasi-conformal in $\mathbb R^3$.
• Maximum principle: deformation error in the interior is controlled by boundary error.
• Compactness: curved patches need far fewer control points for equivalent smoothness in high-curvature domains.
• Smoothness: mapping is $C^\infty$ in the interior except at cage boundaries (Xiao et al., 23 Jan 2025, Liu et al., 2024).

7. Applications and Significance

Flexible high-order cage-based deformation is central to geometry processing, character animation, mesh editing, and emerging fields such as adversarial shape perturbation and 3D neural optimization. The Bézier-Green framework achieves fast, shape-preserving deformation with compact boundary control—enabling robust manipulation even in complex, high-curvature environments and facilitating efficient integration into interactive or large-scale systems.

Recent works extend Green coordinates to arbitrary polynomial cages in $\mathbb R^2$, yielding closed-form, conformal harmonic deformation for any order polynomial boundaries, where interior deformation follows directly by manipulating Bézier control points (Liu et al., 2024). Such methodologies generalize cage-based deformation beyond traditional mesh domains to a wider range of shape and topology descriptors, broadening the impact in computer graphics and vision.

In summary, cage-based deformation with Bézier-patch Green coordinates represents a comprehensive solution for compact, high-quality, shape-preserving deformation of complex models, reconciling the demands of local flexibility and global fidelity within a mathematically principled interpolation framework (Xiao et al., 23 Jan 2025).