Shape-Level Representation

• Shape-level representation is a formalism that encodes an entire object's geometry and semantics as a compact, single descriptor for global analysis.
• It employs methods like neural SDFs, linear models, and spectral descriptors to capture invariants and support pose-invariant comparisons.
• These representations are applied in vision, graphics, and scientific computing for tasks such as shape completion, 6D pose estimation, and semantic segmentation.

A shape-level representation is a mathematical, algorithmic, or learned formalism that encodes the entire geometry, semantics, or structure of an object or region as a single, often compact, entity—rather than as per-point, per-pixel, or local patch features. Such representations aim to capture essential invariants, global geometric properties, or functionally meaningful structures, supporting tasks such as analysis, manipulation, classification, correspondence, or downstream learning across domains including computer vision, graphics, scientific computing, and shape analysis.

1. Mathematical Foundations of Shape-Level Representations

Shape-level representations typically formalize a shape either as:

• An explicit collection of parameters encoding its entire geometry (e.g., point clouds, meshes, parameter vectors)
• The weights of a function (e.g., neural field parameters, SDF weights) encoding a signed distance or occupancy mapping whose level sets define the surface
• A fixed-size code or matrix derived intrinsically from the geometry (e.g., spectral descriptors, functional map embeddings, distance matrices)
• Algebraic or combinatorial constructs capturing mutual spatial/topological relationships (e.g., skeletons, symmetry axes, qualitative descriptors)

For implicit neural field models, a central paradigm is to encode the surface as the zero-level set of a function $f_\theta: \mathbb{R}^3 \to \mathbb{R}$, where $\theta$ is a vector of parameters—the shape-level representation—learned or optimized for each object instance (Lei et al., 2024).

For SDF-based models, $f_\theta(x)$ approximates the signed distance from $x$ to the shape's surface, and the totality of $\theta$ provides a continuous, resolution-agnostic descriptor encoding the entire object (Park et al., 2019).

Linear shape models encode a collection as a low-dimensional affine space: a prototype shape $v_0$, a basis $B$, and shape coordinates $c$ such that each instance is $u(c) = v_0 + Bc$, possibly aligned by an affine transformation for equivariance/separation (Loiseau et al., 2021).

Spectral signatures represent shapes by the spectrum or eigenfunctions of metric-induced operators (e.g., Laplace-Beltrami, quasi-geodesic matrices), yielding compact descriptors intrinsically tied to geometry and topology (Das et al., 2017, Huang et al., 2018).

Table: Key Mathematical Elements in Shape-Level Representations

Formalism	Associated Mathematical Object
Neural SDF/MLP	Parameter vector $\theta$
Linear shape model	$\theta$0
Functional map embedding	Matrix $\theta$1
Spectral descriptor	Eigenvalues/eigenvectors
Level-set/SDF parameters	Weights $\theta$2 in SDF-MLP

2. Invariance Properties and Canonicalization

Shape-level representations often seek invariance to global transformations (translation, rotation, scale) and stability across discretizations, resolutions, or sampling:

• Neural SDF parameters can be conditioned on pose (rotation, translation) via a hypernetwork, yielding representations equivariant to SE(3) and robust to sampling (Lei et al., 2024).
• Canonicalization schemes, such as SVD on distance matrices (Fujiwara et al., 2018), project a shape into a unique, rotation- and scale-invariant coordinate system before embedding.
• Spectral descriptors and latent shape operators (Huang et al., 2018, Das et al., 2017) are intrinsic and independent of embedding choice, enabling pose-invariant comparisons.
• Axis-based and qualitative relational schemas assign each shape an intrinsic coordinate frame based on geometry-invariant centers, axes, or topological features (Aslan et al., 2011, Dorr et al., 2014).

These invariance mechanisms are essential for shape retrieval, classification, and correspondence in unconstrained, unregistered data regimes.

3. Representation Learning: Neural Fields and Parameter Embeddings

Modern paradigms convert the shape-level representation learning task into one of inferring (or encoding) a vector of continuous parameters that define a neural field, implicit function, or structured composition of simple parts:

• Neural SDFs (DeepSDF, DualSDF): Learn an SDF decoder $\theta$3 mapping latent code $\theta$4 and spatial position $\theta$5 to the distance value; shape-level encoding is $\theta$6 (coordinate in latent manifold) or, in atlas or two-level models, a deformation of a global template (Park et al., 2019, Hao et al., 2020, Jiao et al., 2023).
• Level-set parameterization: The entire parameter tensor $\theta$7 of a trained neural SDF is treated as the shape-level descriptor, further normalized or decomposed for dataset-wide coherence; pose conditioning is realized with a hypernetwork generating only the first layer's weights/biases as a function of SE(3) (Lei et al., 2024).
• Alignment-aware linear models: Each shape is reconstructed via a low-dimensional linear basis, with affine alignment networks to allow separation of geometric deformations and global transformations (Loiseau et al., 2021).
• Sparse coding/dictionary-based approaches: Objects are represented by the weights on a learned dictionary (e.g., Local Probing Field atoms), with joint optimization over features, dictionary, and probe placements (Digne et al., 2016).

These approaches support semantic manipulation, interpolation, and novel applications such as few-shot part transfer and 6D pose estimation entirely at the shape level (Loiseau et al., 2021, Lei et al., 2024, Hao et al., 2020).

4. Global, Intrinsic, and Topology-Preserving Representations

Intrinsic, non-local representations capture both the full metric and structure of shapes beyond simple point sets or meshes:

• Spectral geometric signatures: E.g., all-pairs quasi-geodesic distance matrices whose eigendecomposition yields an invariant, discriminative signature capturing both local (curvature) and global (topology, part layout) geometry (Das et al., 2017).
• Functional map latent spaces: Embedding shapes into a shared latent frame via canonicalized functional maps, enabling commutative, algebraic operations, and robust analogy, clustering, and analysis (Huang et al., 2018).
• Qualitative and axis-based structures: Symmetry skeletons (Aslan et al., 2011) or qualitative spatial relation matrices (Dorr et al., 2014) encode the mutual arrangement and structural properties tied to the entire object.

Such methods are particularly able to address isometric invariance, detection of stable regions and self-symmetries, and support algebraic shape analogies and correspondences.

5. Optimization and Numerical Algorithms for Shape-Level Models

Advanced optimization techniques underpin the construction of shape-level representations with geometric priors or constraints:

• Convexity-Constrained Level-Set Methods: Necessary and sufficient convexity is imposed by Hessian or Laplacian constraints on the level-set/SDF function; variational energy minimization is solved via ADMM with penalty or splitting variables, enabling numerically efficient enforcement of global shape properties (Luo et al., 2018, Li et al., 2020).
• Augmented Lagrangian and constraint projection: Strategy for enforcing unit-gradient, convexity, or other hard constraints in the SDF parameter space.
• ADMM for high-dimensional shape optimization: Scales to optimization over N-dimensional domains and supports versatile applications from segmentation to robust convex hull computation (Li et al., 2020).

These numerical frameworks are generalizable, supporting the extension of learned or variational representations to arbitrary geometric functionals and high-dimensional contexts.

6. Experimental Validation and Practical Impact

Shape-level representations have demonstrated favorable properties in a range of applications:

• Classification and retrieval: Level-set parameter vectors outperform point cloud baselines and SO(3)-equivariant networks in both upright and arbitrary pose classification and retrieval tasks on ShapeNet, Manifold40, and similar datasets, achieving accuracies above 90% (Lei et al., 2024).
• 6D pose estimation: Registration is performed directly in level-set parameter space, attaining sub-degree accuracy even under occlusion or noise, surpassing conventional ICP and registration methods (Lei et al., 2024).
• Shape completion and segmentation: DeepSDF, DualSDF, and related SDF-based models enable high-quality completion and segmentation, with few-shot transfer and strong semantic interpretability (Park et al., 2019, Loiseau et al., 2021, Hao et al., 2020).
• Global correspondence and part matching: Hierarchical neural semantic representations (HNSR) leverage shape-level features for training-free, robust global-to-local matching, supporting cross-category transfer (Du et al., 22 Sep 2025).
• Applications in object abstraction and primitive-based representation: Unsupervised shape abstraction and segmentation can be achieved by aligning instance-semantic sparse codes in shape-level space, supporting repeatable primitive discovery (Li et al., 10 Mar 2025).

Table: Selected Empirical Outcomes (with reference to original work)

Application	Best Shape-Level Representation	Key Performance Metric	Reference
Shape classification	Level-set parameter vector	Acc. 91.5–93.5% (SO(3)/upright)	(Lei et al., 2024)
Shape retrieval	Level-set parameter vector	mAP top-1/5/10 >83%	(Lei et al., 2024)
6D pose estimation	SDF-level-set parameters	RRE ≈ 0.06°, RTE ≈ 0.12	(Lei et al., 2024)
Few-shot segmentation	Linear shape model atlas	Top shape-part transfer IoU	(Loiseau et al., 2021)

7. Extensions, Limitations, and Open Directions

• The continuity and resolution independence of field-parameter representations (level-set, SDF, ODF) enable geometric analyses at sub-voxel scales and overcome discretization issues, but present challenges for direct integration with traditional discrete methods (Lei et al., 2024, Houchens et al., 2022).
• Current field parameterizations achieve invariance or pose-conditionality primarily by parameter or hypernetwork transformation of early layers; future work may generalize this to other classes of neural fields and richer transformation groups.
• Despite strong empirical performance, computational and memory costs for storing high-dimensional parameters (e.g., full SDF networks per shape) and for large-scale optimization may require further advances in compression, sharing, or joint learning.
• Integration of explicit hierarchy or semantic part structure into continuous shape-level representations remains a key open challenge—recent developments in two-level (proxy+detail) models and repeatable primitive segmentation frameworks represent partial progress (Hao et al., 2020, Li et al., 10 Mar 2025).
• Joint global/local or multi-scale representations, and the explicit encoding of higher-order geometric or functional invariants (e.g., curvature, correspondence fields, local symmetry) are promising directions for increased expressivity and robustness (Huang et al., 2022, Das et al., 2017).

In summary, shape-level representation encompasses a spectrum of methodologies that encode full-object information in compact, numerically stable, and semantically rich forms, providing a rigorous foundation for geometric learning, analysis, correspondence, and application across the full pipeline of 2D/3D shape understanding.