Landmark shape¶

The equivalence class of corresponding Euclidean landmarks after location, uniform size, and proper rotation are removed.

Version: v1 · 2026-10-31 · History
Domain-specific #: 14528
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Landmark Shape Space → Experimental Design & Statistics
Aliases: Kendall landmark shape, Labelled landmark shape

Core Idea¶

A landmark shape is the geometric configuration of corresponding points after removing a common position, uniform size, and proper rotation.^[1] Formally, begin with an ordered list of at least two points in one Euclidean space, with at least two points distinct.^[1] Two such lists have the same landmark shape when one common translation, one positive uniform scale factor, and one proper rotation carry every point to its corresponding point in the other list.^[1] The shape is the resulting equivalence class, or orbit, rather than any particular drawing, coordinate array, or alignment procedure.^[1] Kendall defines this carrier and quotient explicitly for labelled point configurations.

The choice of transformation group matters. Kendall's base shape keeps reflection outside the equivalence: a reflected planar labelled triangle can be a different shape even when its side lengths match.^[1] Relabelling the points can also change the base class.^[1] An analysis may later identify mirrors or relabellings for its own task, but that produces a coarser class.^[1] It does not alter which configurations were equal under the original proper-similarity relation.

The same object occurs in unlike settings. Kendall constructs a shape space for labelled triangles in the plane.^[1] In an empirical human-body study, corresponding points from three-dimensional scans are treated as labelled configurations; the researchers explicitly define shape classes using translation, scale and rotations in the special orthogonal group, then analyze the observed body shapes.^[2] The former is a small formal configuration; the latter has many corresponding points sampled from standardized body meshes.^[1]^[2] In both, the class removes pose and size while retaining relative geometry.

How would you explain it like I'm…

Slide, Grow, Turn Match

Put stickers on the same spots of two faces, like the nose and both eyes. If you can slide one picture, make it bigger or smaller, and turn it so every sticker lands on its matching sticker, the two have the same landmark shape. Flipping it like a mirror doesn't count, and the stickers have to match up in order.

Shape Without Place or Size

Scientists sometimes mark matching points, called landmarks, on things like faces or bodies. To compare their shapes, they ignore where the points are, how big the whole thing is and which way it's turned. Two sets of points have the same landmark shape if you can slide, resize (evenly) and rotate one to sit exactly on the other, point for point. Flipping it in a mirror is not allowed, so a triangle and its mirror image can be different shapes. Swapping which point is which can also change the shape.

Shape Modulo Position, Scale, Rotation

A landmark shape is the geometry of a labelled set of corresponding points once you remove position, overall size, and rotation. Start with an ordered list of at least two points (not all the same) in ordinary space. Two lists have the same shape if one translation, one positive uniform scaling and one proper rotation map each point to its partner. The shape is the whole group of lists that match this way, an 'equivalence class', not any one drawing or alignment. Reflections are not included, so a mirror-image triangle with the same side lengths can be a different shape, and relabelling points can matter too. Researchers can later choose to ignore mirroring, but that makes a coarser grouping.

A landmark shape is the equivalence class of a labelled configuration of corresponding points under translation, positive uniform scaling and proper rotation. Formally, take an ordered list of at least two points in a Euclidean space with at least two distinct; two such lists have the same shape when a single common translation, scale factor and proper rotation carry each point to its correspondent. The shape is this orbit, not any drawing, coordinate array or alignment procedure; Kendall defines the carrier and quotient explicitly for labelled configurations. The choice of transformation group is essential: Kendall's base shape excludes reflections, so a reflected planar labelled triangle may be a distinct shape despite equal side lengths, and relabelling the points can change the class. An analysis may later identify mirror images or relabellings, but that yields a coarser class without altering equalities under proper similarity. The same object appears in formal settings, like Kendall's shape space of labelled planar triangles, and empirical ones, like corresponding points on standardized 3D human-body meshes analyzed with translation, scale and special-orthogonal rotations; in both, pose and size are removed while relative geometry is kept.

Structural Signature¶

Signature: corresponding labelled Euclidean points + a noncollapsed configuration + the fixed group of common translations, positive uniform scalings and proper rotations + the orbit of that group action.^[1]

Carrier. A configuration is an ordered \(k\)-tuple \(x=(p_1,\ldots,p_k)\) in \(\mathbb{R}^m\), with \(k\geq2\).^[1] Correspondence means \(p_i\) is compared with the same named landmark \(q_i\) in another configuration.^[1] Kendall allows partial degeneracy but excludes total collapse, where every point coincides.^[1]
Equivalence rule. \(x\sim y\) when there exist \(b\in\mathbb{R}^m\), \(a>0\), and \(R\in SO(m)\) such that \(q_i=aRp_i+b\) for every \(i\).^[1]^[2] One transformation must work for all landmarks. Independent repositioning of individual points changes the configuration.
Shape class. The set \([x]\) of all configurations reachable by those transformations is one landmark shape.^[1] Normalizing location and scale gives a pre-shape; quotienting by proper rotations yields a point in shape space.^[1]
Retained variation. The class preserves relative positions up to the declared group. Angle and length ratios of a planar triangle survive.^[1] For a non-collinear labelled triangle constrained to the plane, handedness also survives; this does not imply that every mirror pair in every dimension is distinct.^[1]
Boundary. Reflections, arbitrary relabelling, shears and unequal scaling along different axes are not allowed transformations.^[1] Changing the ambient dimension or landmark correspondence changes the comparison problem.

What It Is Not¶

It is not shape in every ordinary sense. The outline of a cup, the topology of a doughnut, and an observer's visual impression can each support a useful shape judgment without providing an ordered set of corresponding Euclidean landmarks or the fixed proper-similarity rule. The original screened article ranges across those senses; this entry states one narrower mathematical object.

It is not Procrustes transformation. A Procrustes fit chooses or estimates an alignment between configurations; a landmark shape is the class of configurations related by all permitted alignments.^[1]^[2] Fit residuals and statistical means are ways to compare or summarize classes, not the definition of the class itself.^[1]^[2]

It is not similarity by degree. Two configurations may be close under a shape metric yet belong to different exact orbits.^[1] Nor is it automatically a mirror-identified class. Kendall introduces mirror action separately from the base quotient, and later applications must disclose when they fold it into a coarser representation.^[1]

Scope of Application¶

Use this identity when a study has corresponding point landmarks in a declared Euclidean dimension and wants to compare their geometry independent of absolute location, orientation and uniform size. Kendall's construction supplies a formal planar-triangle case.^[1] Valero and colleagues apply the corresponding three-dimensional shape-space idea to standardized human-body scans.^[2] Their mesh preparation establishes corresponding surface points before the shape analysis; correspondence is an input to the mathematical class, not something the quotient creates.^[2]^[1]

The object can underlie statistical operations, including a shape distance, a mean, tangent-space approximation or principal component analysis.^[1]^[2] Those operations add assumptions and choices. In the human-body study, different approximation methods are compared for measured bodies, and the authors distinguish shape from overall size.^[2] The identity of each input class is narrower than the full inferential procedure.

A totally collapsed point list lies outside Kendall's ordinary carrier.^[1] An indexed tuple still has a mathematical orbit even when its points were measured poorly; without defensible correspondence across empirical specimens, however, a comparison of their orbits cannot be interpreted as a comparison of matching anatomical roles. Whether a noisy observation approximates a class well is a measurement question; mathematical equality of classes is exact.

Clarity¶

Think of photographing a labelled triangle while moving the page, turning it, or uniformly zooming in. Those changes alter coordinates and may alter the image, but they do not alter its landmark shape. Turning a planar page over reverses the triangle's handedness and is not a rotation within the plane.^[1] Under the stated planar SO(2) relation, that can change its class.

The same words “remove orientation” can hide two different choices. A proper rotation changes pose without reversing handedness. A reflection reverses handedness and belongs to a different component of the orthogonal group. Kendall explicitly uses the proper rotation group \(SO(m)\) and treats reflection as a separate operation.^[1] The rule must be stated before deciding whether two measured objects have the same shape.

Similarly, a plotted shape statistic need not show every distinction that the underlying classes preserve. In Kendall's quasar study, labelled triangle classes feed a spherical construction, but the empirical figures use a further blackboard quotient that folds vertex relabelling and mirror images.^[1] A point in that display can therefore stand for several distinct base landmark shapes.

Manages Complexity¶

The orbit packages many possible placements, rotations and sizes into one object. Instead of comparing every photograph or coordinate frame separately, an analyst can ask which relative geometric information survives the declared group. Kendall's pre-shape normalization makes that compression explicit; Valero and colleagues use it to situate measured human bodies in a three-dimensional shape space.^[1]^[2]

This simplification is controlled rather than unlimited. The analyst still records the landmark correspondence, ambient dimension, allowable transformations, degeneracies and any later statistical quotient. If those choices are omitted, two reports can use the same word “shape” while making different equality claims. The underlying group action provides a checkable test for the word.

Abstract Reasoning¶

Let \(A=(0,0)\), \(B=(1,0)\), and \(C=(0,1)\) be labelled planar landmarks. Translating all three by the same vector, multiplying all coordinates by a positive number, or rotating all three by one angle gives another representative of the same orbit. The ratios among their side lengths and their labelled orientation do not change.

Now replace only \(C\) by \((0,-1)\) while leaving \(A\) and \(B\) fixed. The two triangles have the same unlabeled side lengths, but the new triangle is a mirror of the original across the \(AB\) line. A proper planar rotation that fixes both distinct labelled points \(A\) and \(B\) is the identity, so it cannot carry \(C=(0,1)\) to \(C=(0,-1)\). They are different labelled landmark shapes under \(SO(2)\).^[1] If the application explicitly adds reflection to its equivalence group, the classes would merge. The difference is in the rule, not in numerical imprecision.

For measured bodies, the same test is conceptual: standardize correspondence, ask whether one allowed common transform maps every landmark, and treat any remaining relative geometry as shape variation. A small fitting residual suggests similarity but does not prove equality of exact classes.^[2]

Knowledge Transfer¶

What transfers from a planar triangle to a scanned body is the orbit rule: a labelled point configuration, one transformation group acting on every point, and a class of all equivalent configurations. The number of landmarks and ambient dimension can change while those roles remain literal. Kendall's formula \(S/SO(m)\), after centering and size normalization, is the general form of that construction.^[1]

What does not transfer automatically is the convenient geometry of a two-dimensional triangle sphere, its coordinates, or an empirical study's uncertainty model. The human-body paper treats a much higher-dimensional shape space and compares intrinsic and approximate statistical methods; a triangle plot cannot supply those conclusions.^[2]^[1] Conversely, anthropometric findings do not determine how a quasar blackboard should fold labels or mirrors.

Examples¶

Canonical: labelled planar triangles¶

Kendall constructs the space of shapes of three labelled points in a plane.^[1] Each noncollapsed triangle configuration has a class after translation, positive common scale and proper planar rotation are removed.^[1] The full labelled triangle shape space has the structure of a sphere; mirror action is a further map on that sphere rather than part of the defining orbit.^[1]

Mapped back: ordered vertices are the carrier; a common proper similarity is the relation; the orbit is the shape; side and angle ratios remain; mirror and vertex relabelling are separate choices.^[1] The construction makes the exact boundary visible because it specifies the group.

Applied: measured three-dimensional human bodies¶

Valero and colleagues use standardized human-body scans with corresponding points in three dimensions.^[2] They define a body shape through labelled point configurations and quotient by translation, scale and special-orthogonal rotations.^[2] They then compare observed shapes and statistical summaries, including mean body shapes.^[2] Their empirical data vary in human proportions rather than in the three-vertex geometry of the canonical example.^[2]

Mapped back: correspondingly indexed mesh points are the carrier; translation, scale and SO(3) are the nuisance group; each person's body configuration supplies a class; relative 3D form remains.^[2] Mesh preparation, Procrustes fitting and statistical comparison are additional operations around that class, not parts of its definition.

Qualified variant: quasar-triangle display¶

Kendall also applies triangle-shape geometry to the planar positions of named Arp–Hazard quasars.^[1] His figures plot triangle shapes on a spherical blackboard that additionally identifies vertex permutations and mirrors.^[1] Mathematically, a labelled base triangle class maps to that coarser display class.^[1] The plotted point should not be described as though it still preserved all labels and handedness.

Mapped back: observed quasar triples instantiate the point carrier and base proper-similarity class, while a later quotient changes what the display distinguishes.^[1] This variant is useful for recognizing a boundary between an object and one way of presenting it.

Structural Tensions¶

Removing nuisance variation versus preserving information. Quotienting by translation, size and proper rotation makes configurations comparable across poses. Adding reflection or arbitrary relabelling may simplify some tasks further, but it can erase distinctions that a handedness or correspondence question needs. Diagnostic: Which exact group was used, and which distinctions must the task preserve?^[1]

Exact class versus approximate fit. The mathematical orbit asserts exact equality. Measured anatomy has noise, imperfect landmarks and preprocessing choices. A Procrustes residual or shape distance can report nearness, while the equality rule stays fixed. Diagnostic: Is the claim about exact membership, measured similarity, or a population summary?^[2]

Stable correspondence versus flexible surfaces. A body surface can be scanned densely, but independent point clouds do not automatically align point-for-point. Standardized homologous meshes and sampled corresponding vertices supply the cross-specimen correspondence needed for an interpretable empirical comparison. Diagnostic: What justifies treating landmark \(i\) in one body as the same role in another?^[2]

Base object versus analysis display. The same base classes may be mapped into a coarser plot or into a tangent-space approximation. These outputs can aid inference while losing some information. Diagnostic: Which conclusions are about the original class, and which only about the later representation?^[1]^[2]

Structural–Framed Character¶

Evaluative weight: membership is a mathematical claim about corresponding points and an allowed transformation group, not a judgment that a body is attractive, a quasar alignment is important, or a statistical estimate is useful. A poor measurement or unhelpful application can still instantiate the same well-defined shape class.

Human-practice dependence: people choose landmarks, ambient dimension and a reflection convention before applying the construction. Once those inputs are fixed, the orbit relation and its consequences are mathematical. Mesh registration in the human-body study is a practical means to supply correspondence; it does not make all body outlines automatically identical classes.^[2]

Institutional origin: Kendall's 1984 paper gives the labelled mathematical construction; Valero and colleagues explicitly use Kendall's 3D shape space for empirical anthropometry.^[1]^[2] No institution, software package or single study owns the identity. Vocabulary travel: words such as shape, equivalence and landmark occur in many fields, but the named object travels literally only where corresponding Euclidean points are quotiented by common translation, positive uniform scale and proper rotation.

Import versus recognition: recognizing that object in measured human bodies follows the authors' explicit SO(3) shape-space definition. Calling an arbitrary silhouette, a topological deformation class or a mirror-folded display point this same labelled orbit would import the term without its defining relation. Its character: a structural, mathematically defined domain-specific equivalence-class object whose application depends on declared landmark and group conventions; it is neither a visual taste category nor a statistical method.^[1]^[2]

Structural Core vs. Domain Accent¶

The portable skeletal move is to specify a carrier, an equivalence relation and its classes. The domain-specific core adds ordered corresponding points in one Euclidean space, noncollapse, and the exact group translation + positive uniform scale + SO(m). Removing cross-specimen point correspondence leaves each formal orbit intact but makes the empirical comparison ambiguous; allowing arbitrary affine deformation, reflection or relabelling changes the class; retaining absolute size yields a different geometric object. The 2D triangle sphere and the 3D body-study shape space instantiate this same core with different numbers of points, dimensions and tasks.^[1]^[2]

The construct does not meet a Prime bar merely because orbit quotients occur throughout mathematics. Those abstract quotient roles are already captured by the generic Equivalence Relation and Invariance primes; the residual here is the Euclidean landmark carrier and proper-similarity group. The present sources establish a domain-specific shape object, not a substrate-independent pattern that still identifies the same thing after the geometric carrier is removed. A future Prime proposal would need multiple unlike nongeometric substrates in which every role of the proposed general mechanism remains literal, and it would need to show an autonomous residual beyond Equivalence Relation, Invariance and other existing primes.

The typed DAG gives this identity two strict compositional parents. Equivalence Relation is a prerequisite: without exact same-shape equivalence there is no orbit class, yet generic equivalence relations occur without landmarks. Group is an internal constituent: proper similarities form reversible composable transformations acting on every admitted configuration, and that action generates the orbit. A shape class is neither a binary relation nor a group, so neither edge is subsumption. Invariance follows from the quotient; Procrustes Transformation, Morphometrics and digital Shape Analysis are neighboring methods or practices rather than all-instance parents.

This entry includes Group and presupposes Equivalence Relation.

The construction exemplifies a specific equivalence relation: reflexivity, symmetry and transitivity follow from a transformation group acting on configurations. It also makes invariance concrete: the class is unchanged under each allowed transform. Those generic ideas help explain the entry, while the labelled carrier and fixed proper-similarity group make it domain-specific. The two typed parents were separately tested for all-instance necessity; the source-specific identity remains the labelled Euclidean carrier and proper-similarity action.

It is related to Procrustes transformation, an alignment operation, and to Morphometrics and digital Shape Analysis, practices that may use landmark classes. Topical proximity alone does not make those practices constitutive parents.

Relationships to Other Abstractions¶

Current abstraction Landmark shape Domain-specific

Parents (2) — outgoing typed relations

Landmark shape presupposes Equivalence Relation Prime

A landmark shape is an orbit class only because proper-similarity sameness is an equivalence relation on labelled configurations.
Landmark shape includes Group Prime

The reversible proper-similarity group and its action are internal to every labelled landmark-shape orbit.

Neighborhood in Abstraction Space¶

Landmark shape sits in a sparse region of the domain-specific corpus (87^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Manifolds, Stacks & Classifying Spaces (19 abstractions)

Nearest neighbors

Fixed points of isometry groups in Euclidean space — 0.81
Reach (Mathematics) — 0.81
Brandt matrix — 0.80
Bundle of Principal Parts — 0.80
Deligne–Lusztig theory — 0.80

Computed from structural-signature embeddings · 2026-10-31

Not to Be Confused With¶

A visual outline or silhouette: may lack labelled point correspondence.
Congruence: usually retains size; this identity removes uniform size.
Mirror-identifying similarity: additionally folds reflection and may merge different labelled planar classes.
Topological shape: can identify deformations no proper similarity permits.
Procrustes analysis: computes fits, distances or summaries rather than being the class itself.
Statistical conclusions about a population: depend on data and inference beyond the equality relation.

References¶

[1] D. G. Kendall, “Shape Manifolds, Procrustean Metrics, and Complex Projective Spaces,” Bulletin of the London Mathematical Society 16 (1984), 81–121, especially §§2, 5–6 and 8.3. https://doi.org/10.1112/blms/16.2.81. Full author article. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y ↩z ↩27 ↩28 ↩29 ↩30 ↩31 ↩32 ↩33 ↩34 ↩35 ↩36 ↩37 ↩38 ↩39 ↩40 ↩41 ↩42 ↩43 ↩44 ↩45 ↩46 ↩47

[2] Jorge Valero et al., “Statistical Shape Analysis of Human Bodies,” Statistical Analysis and Data Mining: The ASA Data Science Journal (2025), especially §§2, 3.1 and 4.2. https://doi.org/10.1002/sam.70024. Publisher full text. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x