Shape-Adaptive Selection Methods

• Shape-Adaptive Selection is a design principle that customizes activation based on local geometric and topological features rather than fixed uniform rules.
• It encompasses techniques such as shape priors, medial ball selection, and adaptive octree tokenization to refine segmentation and reconstruction tasks.
• The approach enhances accuracy and efficiency by leveraging local descriptors and statistical measures to guide adaptive support and resource allocation.

Shape-adaptive selection denotes a family of methods in which the chosen primitives, neighborhoods, supports, updates, or activations are conditioned on shape rather than fixed uniformly. In the literature, this appears as shape priors selecting anatomically plausible segmentations at test time, top-down greedy abstraction with medial balls, interior/exterior decisions on a uniform CNN grid using local data resolution, shape statistically homogeneous pixel selection in DS-InSAR, and selective element activation on reconfigurable holographic surfaces (Bateson et al., 2022, Dou et al., 2019, Sharifi et al., 17 Feb 2026, Yao et al., 16 Sep 2025, Jalali et al., 27 Mar 2025). The common thread is not a single algorithmic template, but a recurrent design principle: local geometry, topology, morphology, or angular structure is used to decide what should be retained, merged, refined, emphasized, or suppressed.

1. Conceptual scope and recurring structure

Across domains, shape-adaptive selection is used to replace uniform rules with geometry-conditioned ones. In segmentation, the relevant choice is among low-entropy predictions; in geometric abstraction it is among candidate medial balls or contour groups; in volumetric learning it is among octree cells or resizing factors; in wireless systems it is among active surface elements; and in visualization it is among candidate shape palettes (Bateson et al., 2022, Dou et al., 2019, Liu et al., 2020, Jalali et al., 27 Mar 2025, Tseng et al., 2024).

Domain	Selected object	Selection signal
Test-time segmentation	plausible masks	shape priors, moments, entropy
3D abstraction	medial balls	local feature size, geometric constraints
CFD masking	grid nodes	nearest-neighbor distance, local resolution
Octree generation	cells or tokens	quadric error, occupancy
RHS deployment	active elements	predefined shape library, throughput gain

A central distinction in several papers is between shape and scale. In DS-InSAR, shape is the scale-invariant second-order structure captured by the normalized shape matrix $V = N\Sigma / \operatorname{tr}(\Sigma)$, whereas scale is overall backscatter strength (Yao et al., 16 Sep 2025). In CFD reconstruction, shape adaptation arises from enforcing interior/exterior decisions with signals that reflect the true geometry and local data resolution rather than from interpolation alone (Sharifi et al., 17 Feb 2026). In greedy pole selection, the balance is between large-ball preference and constraint satisfaction, so that the union of selected balls approximates the enclosed volume with minimal complexity (Dou et al., 2019). This suggests that “shape-adaptive selection” is best understood as a selection policy driven by invariants or descriptors that remain meaningful under nuisance variation in amplitude, scale, or sampling density.

2. Selection criteria: moments, persistence, and angular structure

A large class of methods implements shape-adaptive selection through explicit objective terms. In test-time segmentation, the shape-guided entropy framework minimizes predictive entropy while penalizing deviations from prior shape statistics. The entropy term is

$H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$

and the total objective is

$$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$

where $\mathcal{L}_{shape}$ is built from moments and descriptors such as area, compactness, eccentricity, centroid, and normalized moments (Bateson et al., 2022). The intended effect is that, among many confident segmentations, optimization favors anatomically plausible ones rather than empty masks or blob artifacts.

In multi-scale local shape analysis, the selected features are local PCA descriptors and persistent local homology classes computed at fixed radii. The method uses eigenvalues and eigenvectors from neighborhoods $X\cap B_R(z)$ together with persistence diagrams on $S_R(z)$, then keeps the top-$k$ persistent local homology classes per scale, where persistence is $\ell(u)=death-birth$ (Bendich et al., 2014). The selection is therefore both scale-indexed and topology-aware: singular neighborhoods retain persistent topological structure that would be suppressed by purely geometric summaries.

In Shape-to-Scale DS-InSAR, selection is explicitly statistical. Shape statistically homogeneous pixels share a common angular scattering structure, and the angular consistency adaptive filter evaluates the statistic

$$t(\widetilde{\mathbf z};\widehat{\Sigma}) = \widetilde{\mathbf z}^{\dagger}\widehat{\Sigma}^{-1}\widetilde{\mathbf z}$$

under a CACG model, with bootstrap-derived thresholds (Yao et al., 16 Sep 2025). Here, selection is driven by angular structure rather than amplitude similarity, and phase linking is then performed under a complex generalized Gaussian model. A plausible implication is that shape-adaptive selection often begins by isolating scale-invariant structure before a second stage models scale itself.

3. Geometric primitive selection and correspondence search

In 3D shape abstraction, shape-adaptive selection is literal primitive selection. The medial-axis-based approach begins with inside poles and polar balls obtained from sampled surfaces, then greedily enlarges the covered region by selecting balls subject to application-dependent constraints. For penetration-free ball-stick modeling, the chosen ball is

$$p_{new} = \arg\max_{p_j\in P^-}\{r_j : d_{ij}\ge \epsilon,\ \forall p_i\in P^+\},$$

where $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$0 is the signed gap between balls (Dou et al., 2019). For porous structures, the score becomes

$H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$1

which allows controlled penetration. The same paper ties acceleration to a geometric theorem: if two selected balls intersect, then their power cells are adjacent, so intersection checks can be pruned to local adjacency.

Adaptive locally affine-invariant shape matching treats selection as a dynamic-programming search over contour segmentations. Contours are broken at curvature maxima, opposite points, and max-size points; groups-of-segments are then matched under local affine normalization, with skip transitions handling missing or extraneous contour parts. The objective combines unary matching cost, binary consistency terms, and skip penalties, and the recurrence

$H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$2

searches over valid segment groups ending at indices $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$3 (Marvaniya et al., 2015). Here, shape-adaptive selection is neither feature weighting nor statistical testing, but a combinatorial choice of which contour portions should be grouped, locally corrected, or left unmatched.

These geometric formulations share a key property: the selected support is variable-length and structure-dependent. Large stable regions are covered by coarse primitives; articulated or ambiguous regions are split more finely; and unmatched regions are explicitly representable. That pattern recurs in later neural formulations.

4. Adaptive representations, octrees, and learnable network shape

Several neural methods convert shape-adaptive selection into adaptive allocation of spatial resolution or token budget. Shape Adaptor replaces fixed resizing with a learnable scale selector:

$H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$4

with $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$5 (Liu et al., 2020). This makes the network’s spatial shape trainable end-to-end, with local and global rounding schemes, memory control via a penalty factor $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$6, and extensions to compression and transfer learning. The paper’s empirical point is narrow but important: the schedule of down-sampling itself is a selection problem.

Adaptive octree methods move the same principle into 3D. Dual Octree Graph Networks represent volumetric fields with an adaptive octree and a dual-graph convolution that fuses irregular cross-level neighbors into a regular directional stencil:

$H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$7

and decode the field through Neural MPU blending (Wang et al., 2022). Octree-Based Adaptive Tokenization makes the octree itself the selected support: a cell is subdivided only if it is non-empty and its averaged minimized quadric error exceeds a threshold $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$8, with $H(p) = - \sum_x \sum_{c=1}^{C} p_c(x)\log p_c(x),$9 and token counts reported as approximately $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$0, $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$1, $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$2, and $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$3 for $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$4 (Deng et al., 3 Apr 2025). The resulting variable-length latent is then serialized in breadth-first order for autoregressive generation.

The consequence is a shift from fixed-capacity shape encoding to complexity-conditioned encoding. In OAT, the adaptive representation uses about 50% fewer tokens than fixed-size methods at matched visual quality, while at similar token length it yields higher-quality shapes (Deng et al., 3 Apr 2025). This suggests that shape-adaptive selection in representation learning is fundamentally an allocation problem: where geometry is simple, the support can remain coarse; where geometry is detailed, refinement is justified.

5. Anatomical, contextual, and set-conditioned selection

In image segmentation, shape-adaptive selection is often mediated by context. ACSNet combines Local Context Attention, a Global Context Module, and an Adaptive Selection Module. The hard-region attention map is computed from the previous decoder prediction,

$$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$5

and local features are modulated as

$$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$6

Channel-wise fusion in ASM then selects and aggregates local, global, and decoder features (Zhang et al., 2023). The intended shape adaptivity is explicit: small or boundary-challenging polyps depend more on precise local contrast, whereas large lesions benefit from broad global context.

AdaCoSeg makes the selection set-conditioned rather than anatomy-conditioned. A pre-trained part prior proposes plausible per-shape parts; a co-segmentation network then optimizes labelings over a set using a rank-based group consistency loss,

$$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$7

where $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$8 stacks descriptors of the $$\mathcal{L} = \mathcal{L}_{entropy} + \lambda\,\mathcal{L}_{shape},$$9-th part across shapes (Zhu et al., 2019). The same shape can therefore receive different segmentations in different sets. The selected decomposition is not fixed semantics, but the decomposition that yields low-rank within-label consistency and inter-label separation for the current group.

Adaptive particle-based shape modeling pushes the same idea into correspondence. Particles are encouraged to move toward regions with high signed-distance residuals, but a neighborhood correspondence loss compares normalized local neighborhoods after removing translation, rotation, and scale, and a geodesic correspondence algorithm periodically regularizes optimization by geodesic neighborhood consistency (Xu et al., 10 Jul 2025). The paper makes the trade-off explicit: increasing the adaptivity coefficient increases feature sensitivity, but excessive adaptivity can harm correspondence.

NeuForm addresses another tension: instance-specific detail versus editability. It blends an overfitted and a generalizable neural shape representation by a spatially varying gate $\mathcal{L}_{shape}$0, using the overfitted model where reliable data is available and the generalizable model near altered joints or edited regions (Lin et al., 2022). The distinctive claim is that blending in parameter space and selected feature layers avoids seams more effectively than output-space blending. In all four cases, selection is spatially localized, but the locality may be defined by uncertainty, group consistency, geodesic neighborhoods, or edit-aware joint regions rather than by Euclidean distance alone.

6. Extensions, trade-offs, and unresolved questions

Outside geometry and segmentation, the same principle appears in reconstruction, visualization, function approximation, wireless control, and even multi-parameter statistical learning. In CNN-ready CFD domain recovery, distance-based masking classifies grid nodes by the rule $\mathcal{L}_{shape}$1 with $\mathcal{L}_{shape}$2, runs in $\mathcal{L}_{shape}$3–$\mathcal{L}_{shape}$4 ms per $\mathcal{L}_{shape}$5 mask, and achieves $\mathcal{L}_{shape}$6–$\mathcal{L}_{shape}$7 speedups over classical alpha-shapes; the adaptive alpha-shape variant remains stable at $\mathcal{L}_{shape}$8 and is $\mathcal{L}_{shape}$9–$X\cap B_R(z)$0 faster than the classical version (Sharifi et al., 17 Feb 2026). Here, shape-adaptive selection suppresses unsupported activation while preserving connectivity.

In Adaptive RBF-KAN, the selected quantity is the kernel type and the global shape parameter $X\cap B_R(z)$1. A LOOCV criterion initializes $X\cap B_R(z)$2, then training refines it jointly with edge coefficients. The reported pattern is task-dependent: Gaussian is best on the smooth Franke surface, Matérn $X\cap B_R(z)$3 and Wendland $X\cap B_R(z)$4 handle discontinuities, Wendland $X\cap B_R(z)$5 are strong on oscillatory targets, and Wendland $X\cap B_R(z)$6 is best for localized singularities (Cavoretto et al., 20 May 2026). This is not geometric shape in the morphological sense, but it is still shape-adaptive selection in the sense of choosing basis locality to match function structure.

In visualization, shape-adaptive selection becomes empirical palette design. The pairwise discriminability score

$X\cap B_R(z)$7

is aggregated into a palette objective $X\cap B_R(z)$8 over selected shape sets, and the paper reports that performance “does not map well to classical features of shape such as angles, fill, or convex hull” (Tseng et al., 2024). This is an important corrective to a common misconception: shape-adaptive selection does not require hand-crafted geometric features to be effective, and in some settings those features are poor predictors of actual performance.

In reconfigurable holographic surfaces, the selected object is the active region of the aperture. Binary masks are drawn from a finite shape library and optimized jointly with AP beamforming and RHS phase shifts by alternating optimization. With $X\cap B_R(z)$9 panels, $S_R(z)$0 elements, and two predefined shapes of $S_R(z)$1 active elements each, the shape-adaptive strategy increases throughput from $S_R(z)$2 to $S_R(z)$3 bits/s/Hz at $S_R(z)$4 dBm relative to the best fixed-shape baseline (Jalali et al., 27 Mar 2025). The paper also identifies a limitation: library-based selection is tractable, but continuous shape optimization over the aperture remains open.

A distinct but related use appears in boosting for GAMLSS, where adaptive step-length selection prevents imbalance among location, scale, and shape predictors. The line search

$S_R(z)$5

makes per-parameter comparison fair, and in the Gaussian case the optimal step for $S_R(z)$6 scales with variance while $S_R(z)$7 approaches $S_R(z)$8 late in boosting (Zhang et al., 2021). This broadens the term’s scope: “shape-adaptive” can refer either to geometric morphology or to distributional shape parameters, but in both cases the central issue is preventing fixed update rules from obscuring structurally meaningful variation.

Taken together, these papers define a coherent research pattern. Shape-adaptive selection replaces uniform supports, thresholds, or update schedules with choices conditioned on morphology, topology, local feature size, angular structure, or context. Its benefits are consistent—better fidelity, fewer degenerate solutions, improved efficiency, or better robustness—but so are its trade-offs. Priors can hurt under severe pathology, adaptive refinement can overfit noise, finite shape libraries can limit optimality, and stronger adaptivity can weaken correspondence. The enduring technical question is not whether shape should guide selection, but how to encode that guidance so that it remains stable, computable, and faithful to the structure actually present in the data.