Shape: Invariance and Applications

• Shape is defined as the geometric, topological, or relational structure of an object, typically modeled as an equivalence class invariant to transformations such as translation, rotation, and scaling.
• Shape analysis leverages methods like Euclidean distance matrices, invariant descriptors, and group-equivariant networks to reconstruct, optimize, and interpret intrinsic geometry.
• Applications of shape span engineering, computer vision, biology, and urban design, demonstrating its role as both a reconstructive representation and a controllable functional state.

to=arxiv_search 诺果 {"query":"all:shape AND cat:cs.CV OR cat:math.AT OR cat:stat.ME OR cat:cond-mat.soft OR cat:cs.GR", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"} to=arxiv_search 天天中彩票qq 大发快三是不是 微信上的天天中彩票 to=arxiv_search 彩票总代理 {"query":"id:(Kiyono et al., 2021) OR id:(Romero et al., 1 Jul 2025) OR id:(Arguillère, 2015) OR id:(Anderson, 2018) OR id:(Lamberti, 2022)", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"} Shape denotes the geometric, topological, or relational structure of an object or configuration, and recent research treats it as an equivalence class under nuisance transformations, a coincidence structure on point constellations, a contour or mask descriptor, a response manifold for statistical modeling, and a design variable in engineering, biology, visualization, and urban form (Stöcker et al., 2021, Anderson, 2018, Romero et al., 1 Jul 2025). Across these settings, the central technical issue is invariance: shape analysis commonly seeks representations that ignore translation, rotation, scaling, reflection, indexing, or other transformations that preserve intrinsic geometry, while retaining enough structure for reconstruction, inference, optimization, or interpretation (Romero et al., 1 Jul 2025, Lamberti, 2022).

1. Quotient, manifold, and topological definitions

In geometric statistics, the shape of a planar curve and/or landmark configuration is its equivalence class under translation, rotation and scaling, whereas its form is its equivalence class under translation and rotation while scale is preserved (Stöcker et al., 2021). In the notation of planar shape analysis, shape is written as

$$[y]_{Trl\,Rot\,Scl} = \{\lambda u\,y+\gamma\,1:\lambda\in\mathbb{R}^+,\,u\in S^1,\,\gamma\in\mathbb{C}\},$$

while form is

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$

This quotient viewpoint places shape and form on non-Euclidean spaces, so regression, averaging, and distance are defined through tangent spaces, geodesics, and exponential and logarithm maps rather than through ordinary coordinate subtraction (Stöcker et al., 2021).

A broader abstract setting defines a shape space in an ambient manifold $M$ as a Banach manifold carrying a compatible action of a Sobolev diffeomorphism group of the ambient manifold (Arguillère, 2015). In this framework, landmarks, embeddings, products of shape spaces, tangent bundles, and ILH smooth shape spaces are unified by an ambient-action formalism, and the infinitesimal action

$\xi_q : \Gamma^s(TM)\to T_q\mathcal S$

encodes how an ambient vector field deforms a shape (Arguillère, 2015). This makes deformation-based shape analysis, especially LDDMM, naturally sub-Riemannian rather than merely Riemannian (Arguillère, 2015).

A far coarser notion appears in Topological Shape Theory, where shape is reduced to the topological content of $N$-point constellations in dimension $d$ (Anderson, 2018). Here, topological shape spaces are graphs: vertices are topologically distinct configurations, and edges encode topological adjacency, meaning one can pass from one configuration type to another by a single topological operation such as one coincidence forming or resolving (Anderson, 2018). In $d\ge 2$, the topological classes including maximal coincidence are exactly partitions of $N$, but the significant object is not merely the set of partitions; it is the graph of adjacencies among those partition-types (Anderson, 2018).

A constructive alternative defines infinite families of qualitatively similar shapes from a finite ordered set of landmarks

$$\mathbf r_0,\mathbf r_1,\ldots,\mathbf r_{N-1}\in\mathbb R^D$$

and a continuous parameter $\kappa\in(0,\infty)$ (García-Morales, 2021). The core construction replaces discrete landmark selection by a smooth $$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$0-embedding and yields open or closed shape families whose small-$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$1 limit interpolates the landmarks and whose large-$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$2 limit collapses to the centroid (García-Morales, 2021). A key theorem states that every member of the family lies in the convex hull of the landmarks (García-Morales, 2021). This suggests a persistent duality in the literature: some theories define shape by quotienting away nuisance structure, while others retain a parameterized representative because parameterization itself is analytically useful.

2. Shape descriptors, invariance, and statistical representation

A recurrent requirement in shape quantification is a descriptor that is invariant to transformations that preserve intrinsic geometry while still remaining discriminative (Romero et al., 1 Jul 2025). In ShapeEmbed, a simply connected 2D contour is sampled as

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$3

and encoded through its Euclidean distance matrix $$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$4, with entries

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$5

Translation and rotation invariance come directly from pairwise distances, and scale invariance is obtained by Frobenius normalization,

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$6

while point indexing and traversal direction are handled by an encoder with circular padding and a reconstruction loss that minimizes over all $$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$7 equivalent reindexings (Romero et al., 1 Jul 2025). The resulting latent code is intended to capture intrinsic contour geometry only, while remaining reconstructive enough that outlines can be recovered by multidimensional scaling (Romero et al., 1 Jul 2025).

Few-shot shape recognition adopts a different learned representation. FSSD treats shape as the geometric structure of an object, expressed mainly through its silhouette and boundary, and uses a group-equivariant CNN together with a dual attention mechanism and learnable shape primitives (Shi et al., 2023). The primitive-based reconstruction is written as

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$8

so that each sample’s shape feature is a linear combination of learnable primitives (Shi et al., 2023). The method is explicitly designed to counter texture bias and to generalize to unseen shapes under few-shot conditions (Shi et al., 2023).

A single-number summary of high-dimensional shape is pursued in hyper-Shape Proportion and hyper-Sphericity (Lamberti, 2022). The paper defines

$$[y]_{Trl\,Rot} = \{u\,y+\gamma\,1:u\in S^1,\,\gamma\in\mathbb{C}\}.$$9

as occupancy relative to the minimum encompassing $M$0-ball, and

$M$1

as an $M$2-dimensional generalization of compactness or sphere-likeness (Lamberti, 2022). For an $M$3-ball, both metrics attain the calibration value $M$4, and the paper derives closed forms for $M$5-simplexes, $M$6-cubes, and $M$7-orthoplexes (Lamberti, 2022). This suggests that invariant shape analysis often alternates between full reconstructive representations and deliberately compressed scalar summaries.

3. Learning, perceiving, and reconstructing shape

In computer vision, shape may be predicted directly rather than through bounding boxes or class labels alone. “Straight to Shapes” extends a YOLO-style detector so that each detection predicts not only box coordinates and category probabilities, but also a compact shape code that can be decoded into an object mask (Jetley et al., 2016). The preferred representation is a learned shape encoding produced by a denoising convolutional auto-encoder, trained with a binary cross-entropy reconstruction loss, and then regressed jointly with localization and classification (Jetley et al., 2016). The paper reports what it describes as the first real-time shape prediction network, running at about 35 FPS, and argues that the learned embedding supports higher-order concepts such as viewpoint similarity, pose variation, and occlusion reasoning (Jetley et al., 2016).

Multimodal reconstruction pushes the idea further by treating shape as an inference problem over a learned prior. In monocular vision and touch, a robot first predicts a voxelized 3D shape from a single RGB image and then refines that estimate by touching regions of high uncertainty with a GelSight sensor (Melikhov, 2018). Uncertainty is defined from voxel occupancies by

$M$8

and the robot actively explores regions minimizing aggregate confidence, while tactile observations are converted into occupancy constraints that update the latent code of the 3D shape estimator (Melikhov, 2018). The method combines monocular RGB vision, high-resolution tactile sensing, and priors learned from ShapeNet, and the paper reports that common objects can be reconstructed from a color image and a small number of tactile explorations, around $M$9 (Melikhov, 2018).

In industrial CAD analysis, Shape is presented as a self-supervised 3D geometry foundation model that converts surface meshes into dense per-token embeddings on a fixed latent 3D grid (Mounmo et al., 19 Apr 2026). Given sampled surface points, normals, and curvature, the backbone produces

$\xi_q : \Gamma^s(TM)\to T_q\mathcal S$0

with $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$1 latent tokens in the released model (Mounmo et al., 19 Apr 2026). Pretraining combines masked-token reconstruction of 28-dimensional geometry statistics with multi-resolution contrastive consistency, and per-token squared reconstruction residuals serve as attribution scores for explainability (Mounmo et al., 19 Apr 2026). A 10.9M-parameter model pretrained on 61,052 CAD meshes achieves $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$2 on masked geometric reconstruction and $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$3 top-1 retrieval on held-out meshes, while a $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$4 ablation identifies per-dimension normalization of heterogeneous geometric targets as critical (Mounmo et al., 19 Apr 2026).

4. Shape in biological systems and active materials

Cell biology uses shape not as a nuisance-invariant equivalence class, but as a physically informative state variable. In confluent epithelial monolayers, cell shape is quantified by the dimensionless parameter

$\xi_q : \Gamma^s(TM)\to T_q\mathcal S$5

where $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$6 is cell perimeter and $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$7 is area (Gopinath et al., 18 Nov 2025). Low-density mobile MDCK and HaCaT monolayers show a broad, positively skewed distribution $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$8, peaked near $\xi_q : \Gamma^s(TM)\to T_q\mathcal S$9 and centered above the static-reference value $N$0, with mean $N$1, normalized variance about $N$2, and skewness about $N$3 (Gopinath et al., 18 Nov 2025). The paper argues that this distribution is not a heterogeneous mixture of cells with fixed intrinsic shapes and cannot plausibly arise from small elastic fluctuations around a fixed preferred shape (Gopinath et al., 18 Nov 2025). Instead, a deformable particle model in which the preferred perimeter adapts to local forces during motion reproduces the experimental distribution to within $N$4, leading to the claim that in fluidized confluent tissues, cell shape is an emergent consequence of active motion and force adaptation, not a fixed input parameter (Gopinath et al., 18 Nov 2025).

Shape also functions as a programmable material state in magnetic shape memory polymers (Ze et al., 2019). The reported composite embeds Fe$N$5O$N$6 microparticles for inductive heating and NdFeB microparticles for magnetic actuation in an amorphous shape memory polymer matrix (Ze et al., 2019). Above the glass transition temperature, the matrix softens and stored magnetization profiles drive reversible bending or folding under an actuation field; after cooling, the deformed shape is locked (Ze et al., 2019). For the representative composite P15-15, the storage modulus drops from $N$7 at $N$8 to $N$9 at $d$0, and shape-memory testing reports a shape fixity ratio of $d$1 and a shape recovery ratio of $d$2 (Ze et al., 2019). Here shape is neither purely geometric nor purely descriptive; it is a controllable functional state linked to actuation, locking, and reprogrammability.

These biological and materials examples indicate that shape can be modeled either as a statistical observable generated by dynamics or as a mechanically addressable state variable. A plausible implication is that the distinction between “shape as representation” and “shape as physics” is domain-dependent rather than absolute.

5. Shape in visualization, theory, and urban morphology

Visualization research treats shape both as a perceptual encoding channel and as a conceptual scaffold. In theory-figure research, “theory figures” are figures that depict a theory’s components and relationships, and the paper’s central claim is that “theory is shapes” (Varona et al., 1 Oct 2025). Cartesian planes, matrices, networks, and set diagrams are described as conventional theory-figure shapes with distinct diagrammatic affordances: dimensional mapping and quadrants for planes, categorical combination for matrices, path-following and branching for networks, and overlap or containment for set diagrams (Varona et al., 1 Oct 2025). The paper then argues for more expressive shapes such as horseshoes, icebergs, Möbius strips, and BLT sandwiches, not as decoration but as devices that can shape theorizing itself (Varona et al., 1 Oct 2025).

As a perceptual encoding, shape is studied in multiclass scatterplots through tasks such as relative mean judgment and correlation estimation (Tseng et al., 2024). The paper evaluates 39 shapes across four experiments and reports that shape palette design is not well explained by broad descriptors such as filled, unfilled, open, angle count, or convex hull (Tseng et al., 2024). Performance varies substantially across specific pairs, category count matters strongly, expert-designed palettes differ widely in effectiveness, and expert selections themselves show low consensus, with average pairwise cosine similarity $d$3 and $d$4 in an expert-choice study (Tseng et al., 2024). The paper therefore builds a model from pairwise performance data and category-count-dependent matrices, together with a design tool for recommending shape palettes (Tseng et al., 2024).

Urban morphology uses shape in yet another sense. A parametric city model separates elongation $d$5, sprawl $d$6, average building height $d$7, and radial vertical profile $d$8, and evaluates average commuting distance as a proxy for transportation energy demand (Batik et al., 30 Jun 2025). In this model, compactness beats elongation, low sprawl beats dispersed urbanization, and centrally dense pyramid or needle profiles outperform bowl and ring profiles (Batik et al., 30 Jun 2025). The paper explicitly states that compact and centrally dense cities minimise the total travel distance in cities, and that a compact and round city reduces travel distance, while elongated, sprawling, and peripherally concentrated cities perform worst (Batik et al., 30 Jun 2025). It also emphasizes that this “best shape” is defined under a narrow but important criterion—minimizing average travel distance as a proxy for mobility energy—and does not automatically optimize daylight, open space, congestion, or social outcomes (Batik et al., 30 Jun 2025).

6. Shape optimization, pre-shape calculus, and named computational systems

Engineering shape optimization often distinguishes the geometric optimum from the parameterization used to compute it. In pre-shape calculus, the pre-shape space is

$d$9

while the corresponding shape space is the quotient

$d\ge 2$0

(Luft et al., 2021). This distinction is exploited to optimize shape and mesh quality simultaneously: instead of modifying the metric used to represent the shape gradient, the paper adds pre-shape derivatives of parameterization-tracking functionals only to the right-hand side of the gradient system (Luft et al., 2021). The fully regularized system combines the original shape derivative, a tangential surface-tracking term, and a projected volume-tracking term, while leaving the optimal shapes to the original problem invariant under regularization and avoiding larger additional linear or nonlinear systems solely for mesh regularization (Luft et al., 2021).

The term SHAPE also appears as a proper noun in machine learning, though here it denotes shift invariance in sequence models rather than geometric form. “Shifted Absolute Position Embedding for Transformers” replaces $d\ge 2$1 during training by

$d\ge 2$2

thereby inducing shift invariance while preserving the simplicity of absolute positional encoding (Kiyono et al., 2021). The method reduces to standard APE when $d\ge 2$3, and the paper sets $d\ge 2$4 during inference, so the shift is purely a training mechanism (Kiyono et al., 2021). Empirically, SHAPE is reported as comparable to relative position embedding on WMT16 English–German while remaining much closer to APE in speed and architectural simplicity (Kiyono et al., 2021).

Across these engineering and computational uses, shape becomes a locus where invariance, parameterization, and optimization meet. This suggests a unifying pattern across otherwise distant literatures: whether the object is a contour, a mesh, a CAD surface, a city, or a diagram, shape is repeatedly defined by what should be preserved, what may vary, and which geometry is treated as the meaningful signal.