Cage-Based Geometric Filtering

• Cage-based geometric filtering is a family of methods that use an auxiliary, coarser geometric scaffold to guide filtering and deformation on complex target shapes.
• It employs techniques like generalized barycentric coordinates and Jacobian-mediated transport to suppress artifacts and ensure smooth spatial signal transfer.
• Applications span texture transfer, 3D Gaussian splatting, NeRF editing, wild mesh learning, and soft-robot reconstruction, highlighting its versatility and efficiency.

Cage-based geometric filtering denotes a family of geometry-processing methods in which a cage—an auxiliary enclosing mesh, a user-specified boundary, or a sparse control lattice—serves as the spatial support for filtering operations on a finer geometric representation. In recent work, cages are used to suppress texture-transfer artifacts by excluding Non-Cosmetic Zones (NCZs), to define deformation fields for meshes, NeRFs, and 3D Gaussian Splatting, and to provide a clean proxy domain for learning on wild meshes (Zhou et al., 23 Jun 2026, Li et al., 2023, Edelstein et al., 24 May 2025, Xiao et al., 23 Jan 2025). Taken together, these works suggest that, in the cage-based setting, “filtering” is broader than classical denoising: it includes exclusion of invalid regions, coarse-to-fine signal transfer, Jacobian-mediated transport of local structure, and regularized deformation.

1. Cage representations and their roles

The literature uses several distinct cage constructions, but all of them place a lower-dimensional or structurally simpler geometric scaffold around a more complex target. In real-time texture transfer, the cage is an auxiliary source mesh already present in the pipeline, and it serves both as the source of texture or UV information and as the spatial reference for deciding which target vertices should be filtered out. In wild-mesh learning, the cage is a single-component manifold triangle mesh that envelops the input closely. In interactive implicit-scene editing, the representation uses nested proxy cages, with an inner cage defining the operation target and an outer cage defining the adjustment space. In high-curvature deformation, the cage itself becomes a higher-order surface, realized with Bézier patches. In tactile reconstruction of soft robots, the cage is a regular control lattice rather than a surface mesh (Zhou et al., 23 Jun 2026, Edelstein et al., 24 May 2025, Li et al., 2023, Xiao et al., 23 Jan 2025, Shou et al., 20 Mar 2026).

Setting	Cage form	Filtering role
Texture transfer (Zhou et al., 23 Jun 2026)	Source cage mesh	UV source and NCZ reference
3DGS / NeRF deformation (Tong et al., 17 Apr 2025, Xie et al., 2024, Li et al., 2023)	Enclosing or nested cages	Smooth deformation support
Wild-mesh learning (Edelstein et al., 24 May 2025)	Single-component manifold envelope	Proxy domain for signal transfer
Curved deformation (Xiao et al., 23 Jan 2025)	Bézier patch cage	High-curvature boundary control
Soft-robot reconstruction (Shou et al., 20 Mar 2026)	Regular $15 \times 15 \times 15$ cage grid	Sparse deformation basis

This variety of representations is significant because it shows that cage-based filtering is not tied to a single discretization. The common denominator is the use of a coarse geometric support that mediates operations on a denser or noisier target. A plausible implication is that the cage is best understood as a structural prior: it reduces the degrees of freedom of the filtering problem while preserving explicit geometric control.

2. Coordinate models, interpolation, and differential transport

Most cage-based methods express the target geometry or signal as an interpolation induced by cage degrees of freedom. In CAGE-GS, a point $p$ inside the source cage is written as

$p = \sum_i \omega_i(p)\, v_i^s,$

and after cage deformation the point becomes

$p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$

The same work uses the cage as the geometric filter that maps every enclosed point from source to target space. CageNet uses generalized barycentric coordinates and a transfer operator

$\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$

with $(C_{VW})_{ij} = \lambda_j(p_i)$ and $\sum_j \lambda_j = 1$. Flexible Bézier-patch cages extend Green coordinates to higher-order boundaries and write the deformed point as

$$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$

Sketch-guided Gaussian Splatting and related methods use a cage-induced deformation field $f_{S'}$ together with its Jacobian $J_f = \nabla f_{S'}$, while interactive NeRF editing uses trilinear interpolation inside the cage-defined adjustment region (Tong et al., 17 Apr 2025, Edelstein et al., 24 May 2025, Xiao et al., 23 Jan 2025, Xie et al., 2024, Li et al., 2023).

These formulations differ in mathematical detail, but they perform closely related tasks. Mean value coordinates, harmonic coordinates, biharmonic coordinates, Green coordinates, and trilinear interpolation all define mappings from cage variables to interior geometry. The main distinction lies in what additional structure is preserved. Green coordinates include both positions and normals. Jacobian-based methods preserve local anisotropy by transporting covariance or local differential structure. Proxy-domain methods on wild meshes preserve spatial smoothness and geometric locality through interpolation matrices rather than direct processing on disrupted connectivity.

3. Geometric filtering for artifact suppression in texture transfer

The most explicit use of cage-based geometric filtering as a filtering pipeline appears in “Cage-based Texture Transfer with Geometric Filtering” (Zhou et al., 23 Jun 2026). The method targets projection artifacts in low-latency texture transfer, including bleeding onto internal geometry, unintended projection into excluded parts, projection through holes or gaps, black speckling or stray texture leakage, and misprojection onto regions the artist wants untouched. The central construct is the NCZ, covering regions such as internal geometry and artist-intended exclusions like eyes, teeth, and tongue.

The input consists of a source cage mesh, a target mesh, and UV mappings or texture information associated with the cage. For each target vertex $p$0, the method finds the nearest cage triangle and uses that triangle’s normal $p$1 to cast a ray from the vertex position $p$2: $p$3 If this ray intersects the target mesh itself, the vertex is marked as NCZ because it likely lies on internal or occluded geometry. For vertices that survive that test, the same ray is tested against the cage mesh. If the ray intersects the cage, the vertex is considered valid; if it misses the cage, for example by passing through a hole, the vertex is marked as NCZ.

The method then elevates the decision from individual vertices to connected components. The target mesh is partitioned into connected components

$p$4

and a segment-level transfer ratio is computed as

$p$5

where $p$6 is the accumulated surface area of a segment after self-intersection filtering and $p$7 is the remaining surface area after cage-intersection filtering. If $p$8 is below a threshold, the entire segment is culled and all its vertices are labeled NCZ. This segment-level culling is the paper’s key geometric filtering stage.

The implementation is explicitly lightweight. Spatial queries on both target and cage meshes use KD-Trees, and the computational complexity is reported as

$p$9

with memory

$p = \sum_i \omega_i(p)\, v_i^s,$0

The reported runtime is approximately $p = \sum_i \omega_i(p)\, v_i^s,$1 ms on an Android Samsung Tablet S6 Lite for a 4,782-triangle lizard head, with roughly $p = \sum_i \omega_i(p)\, v_i^s,$2 MB across hardware tiers.

Method	Runtime	Artifact suppression
Naive Texture Transfer	$p = \sum_i \omega_i(p)\, v_i^s,$3–$p = \sum_i \omega_i(p)\, v_i^s,$4 ms / also described as $p = \sum_i \omega_i(p)\, v_i^s,$5 ms	Low
Manual Authoring	$p = \sum_i \omega_i(p)\, v_i^s,$6–$p = \sum_i \omega_i(p)\, v_i^s,$7 minutes	Gold standard
Proposed framework	$p = \sum_i \omega_i(p)\, v_i^s,$8–$p = \sum_i \omega_i(p)\, v_i^s,$9 ms	High

The significance of this pipeline is its positioning between naive proximity-based transfer and large learned models. The paper states that robust learned alternatives are resource-intensive, require multiple days of training on manually annotated datasets, and incur large memory costs, whereas the cage-based method requires no training, no second UV sets, and no manual NCZ labeling. That makes it directly suitable for digital character cosmetics, interactive avatar personalization, automotive texturing, game assets, and other resource-constrained interactive media.

4. Deformation-centered filtering: Gaussian splatting, NeRFs, and high-order cages

A second major interpretation of cage-based geometric filtering treats the cage as the support of a deformation field. In “CAGE-GS: High-fidelity Cage Based 3D Gaussian Splatting Deformation,” the cage is learned from the target and used to guide geometric transformation of a source 3DGS scene (Tong et al., 17 Apr 2025). Gaussian centers are deformed through mean value coordinates, while covariance is updated through a Jacobian matrix: $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$0 The paper’s central claim is that geometry is transferred by the cage and local appearance structure is transferred by the Jacobian. Quantitatively, it reports CD $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$1, DINO $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$2, and user votes $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$3, with roughly $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$4 minutes for a source model with about $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$5k Gaussians. The sampling strategy for Jacobian estimation reduces chair runtime from $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$6 minutes to $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$7 minutes and car runtime from $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$8 minutes to $p' = \sum_i \omega_i(p)\, v_i^{s\to t}.$9 minutes.

“Sketch-guided Cage-based 3D Gaussian Splatting Deformation” uses a closed manifold triangular cage extracted from a GS model and optimizes that cage with a variant of Neural Jacobian Fields rather than direct vertex manipulation (Xie et al., 2024). Gaussian centroids are moved by $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$0, and covariance is updated by

$\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$1

The system combines silhouette loss

$\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$2

with 3D-aware SDS. The ablation reports about a $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$3 average decrease in $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$4 in two examples relative to the stronger constrained setup, and relative CLIP-IQA changes by less than $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$5, whereas the MLP+HexPlane baseline loses $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$6 on average. Here the cage acts simultaneously as a low-dimensional optimizer state and as a regularizer against entanglement and fuzzy renderings.

“Interactive Geometry Editing of Neural Radiance Fields” applies an analogous idea directly to an implicit radiance field (Li et al., 2023). Two proxy cages are used: an inner cage defines the target region, and an outer cage defines the adjustment space. Edits operate by warping sampling coordinates back into canonical NeRF space rather than by extracting and editing explicit geometry. The discrete mode changes only the inner region; the continuous mode interpolates deformation through the space between inner and outer cages. The paper explicitly notes that a larger outer cage generally produces smoother results, showing how cage extent functions as a support radius for deformation filtering.

“Flexible 3D Cage-based Deformation via Green Coordinates on Bézier Patches” generalizes the cage itself from piecewise-linear boundaries to curved, high-order boundaries (Xiao et al., 23 Jan 2025). The Bézier patch

$\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$7

defines a compact curved cage, and the method derives Green coordinates from Green’s third identity, then corrects approximate coordinates by a global projection

$\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$8

The reported advantage is smoother high-curvature deformation with a more compact cage than dense polygonal approximations.

A sensor-driven version appears in “Zero Shot Deformation Reconstruction for Soft Robots Using a Flexible Sensor Array and Cage Based 3D Gaussian Modeling” (Shou et al., 20 Mar 2026). The undeformed STL proxy is enclosed by a regular $\mathcal{P}(M, W, \tilde{y}) = C_{VW}\,\tilde{y},$9 cage with $(C_{VW})_{ij} = \lambda_j(p_i)$0 nodes. A graph attention network with $(C_{VW})_{ij} = \lambda_j(p_i)$1 GAT layers, $(C_{VW})_{ij} = \lambda_j(p_i)$2 attention heads, and $(C_{VW})_{ij} = \lambda_j(p_i)$3 hidden units regresses cage displacements from tactile input, and dense Gaussian centers are updated through inverse-distance weighting: $(C_{VW})_{ij} = \lambda_j(p_i)$4 The full model reports IoU $(C_{VW})_{ij} = \lambda_j(p_i)$5, SSIM $(C_{VW})_{ij} = \lambda_j(p_i)$6, and Chamfer distance $(C_{VW})_{ij} = \lambda_j(p_i)$7 mm, with about $(C_{VW})_{ij} = \lambda_j(p_i)$8 FPS for coarse Gaussians and about $(C_{VW})_{ij} = \lambda_j(p_i)$9 FPS for high-resolution Gaussians. The ablation against direct dense regression indicates that the cage functions as a deformation prior that suppresses drift and enforces globally consistent geometry.

5. Proxy-domain filtering and learning on wild meshes

Cage-based geometric filtering is also used as a proxy-domain strategy for data with inconvenient topology. “CageNet: A Meta-Framework for Learning on Wild Meshes” defines a cage as a single-component manifold triangle mesh that envelopes the input mesh closely (Edelstein et al., 24 May 2025). The cage is constructed automatically by computing the unsigned distance field of the input shape, extracting an $\sum_j \lambda_j = 1$0-level set using marching cubes, removing internal components using winding numbers, increasing $\sum_j \lambda_j = 1$1 if the surface is disconnected, and simplifying with quadric error edge collapse. In experiments the cage typically has about $\sum_j \lambda_j = 1$2K faces, and sometimes up to $\sum_j \lambda_j = 1$3K.

The central mechanism is generalized barycentric transfer. A function is predicted on the cage and then mapped back to the target through

$\sum_j \lambda_j = 1$4

The paper discusses mean value coordinates, harmonic coordinates, and biharmonic coordinates, and uses DiffusionNet as the mesh network operating on the clean cage. This is not presented as denoising in the classical signal-processing sense, but the structure is unmistakably filter-like: define a signal on the coarse support, process it there, and prolong it back to the fine domain through smooth interpolation.

Empirically, the proxy-domain design preserves performance on clean data while improving robustness on broken data. On the Human Segmentation benchmark, DiffusionNet and CageNet both report $\sum_j \lambda_j = 1$5 on clean inputs, while CageNet generalizes better to broken or wild meshes. For skinning weights on $\sum_j \lambda_j = 1$6 artist-created Roblox biped characters, with $\sum_j \lambda_j = 1$7 test meshes held out, the method reports Avg $\sum_j \lambda_j = 1$8, Precision $\sum_j \lambda_j = 1$9, Recall $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$0, $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$1, Avg vertex distance $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$2, and Max distance $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$3. The cage offset augmentation values $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$4, $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$5, $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$6, and $$\tilde{\eta} = \sum_Q\sum_{i=0}^{m}\sum_{j=0}^{n} \left( \phi_Q^{ij}(\eta)\,\tilde b_{ij}^Q + \psi_Q^{ij}(\eta)\,\tilde N_{ij}^Q \right).$$7 further improve robustness to topological variations in the cage itself.

This proxy-domain viewpoint is important because it decouples filtering from the original topology. Disconnected components, non-manifold elements, and triangle soups become secondary once the learning or diffusion operator runs on a single clean enclosure. A plausible implication is that cage-based filtering can be viewed as a form of geometry-aware domain transfer: the cage provides a stable computational manifold even when the target does not.

6. Scope, misconceptions, and reported limitations

A recurrent misconception is to treat every paper containing the words “cage” or “geometry filtering” as part of the same technical lineage. Several nearby literatures are explicitly distinct. “Low Rank Matrix Approximation for Geometry Filtering” is a normal-estimation and position-update framework based on non-local self-similarity and weighted nuclear norm minimization; it does not use control cages, deformation cages, or cage-like embedded control structures (Lu et al., 2018). “A General Framework for Bilateral and Mean Shift Filtering” provides domain-agnostic feature-preserving smoothing on images, meshes, and other domains, but it does not use cage deformation or cage coordinates (Solomon et al., 2014). “Cage Breaking Far from Equilibrium” studies active-matter escape on an entropic landscape rather than geometry processing (Popowski et al., 16 Mar 2026). “An interaction network approach predicts protein cage architectures in bionanotechnology” concerns protein cage polymorphism via interaction-network tilings, not deformable control cages (Fatehi et al., 2023). “V-CAGE” uses a dynamic map of prohibited spatial areas or prohibited volumes during context-aware scene instantiation, and the paper explicitly notes that the “cage” is not a literal cage module in the formal sense (Liu et al., 21 Jan 2026).

Within actual cage-based geometry-processing papers, limitations are reported consistently. The NCZ-based texture-transfer pipeline depends on segmentation to express intended NCZ structure well and is sensitive to cage quality and fit; without segmentation, results may be less aligned with artist intent, and poor cages can produce nonsensical filtering outcomes (Zhou et al., 23 Jun 2026). CAGE-GS assumes that the target shape can be converted into a point-cloud proxy and cannot guarantee preservation of straight lines, planes, and parallel structures, especially for artificial or man-made shapes (Tong et al., 17 Apr 2025). CageNet is problematic for self-intersections if intersecting regions should have different labels or features, and cage generation can remain topologically unstable in difficult cases (Edelstein et al., 24 May 2025). The sketch-guided GS system depends on ControlNet and on extracting a mesh from the GS with SuGaR, which can take several minutes (Xie et al., 2024). Bézier-patch Green coordinates improve curvature handling but do not guarantee containment of the deformed object inside the cage (Xiao et al., 23 Jan 2025).

These reported limits indicate that cage-based geometric filtering is strongest when a reliable enclosure or scaffold exists, when the target benefits from coarse spatial regularization, and when explicit geometric control matters more than unrestricted dense prediction. This suggests a continuing convergence between cages, learned priors, and task-specific verification: not the elimination of geometric structure, but its increasingly deliberate use as a controllable filtering substrate.