Must shape be represented by landmarks?

No. Curves, surfaces, and other representations are possible, with their own correspondence and statistical requirements.

Are differently sized objects always different shapes?

Not under the common shape definition that removes uniform scale, although their physical size and form can differ.

Does alignment remove every imaging effect?

No. For example, perspective distortion and occlusion may require additional modeling or better acquisition rather than simple rigid alignment.

AI Glossary / Computer vision and multimodal AI / Intermediate

Statistical shape analysis

Definition

Statistical shape analysis studies variation in the geometry of objects using statistical methods. It commonly separates shape from differences in position, orientation, and size, then analyzes the remaining geometric variation. Objects may be represented by corresponding landmarks, curves, surfaces, or other structured descriptions, depending on the problem.

What it means

Two copies of a triangle can have the same shape even when one is moved, rotated, or uniformly enlarged. Shape analysis aims to distinguish those changes from changes in the triangle's proportions or angles.

The transformations removed depend on the scientific or business question. If absolute size matters, removing it would discard useful information; an analysis of form may therefore retain size alongside shape.

Statistical shape analysis is not simply image classification. It asks how geometry varies across examples, how to summarize that variation, and how strongly observed differences are supported after accounting for measurement and sampling uncertainty.

How it works

Choose a representation and establish correspondence. In a landmark approach, the same named location must appear in the same position in each object's coordinate list. Inconsistent correspondence can masquerade as real shape variation.

Align configurations according to the transformations considered irrelevant. Procrustes methods commonly remove translation and rotation, and may normalize scale. A collection can be aligned jointly rather than choosing an arbitrary single specimen as the final reference.

Compute summaries such as mean configurations, distances, or dominant variation patterns using methods appropriate to the resulting shape space. Ordinary Euclidean statistics are not automatically valid for every geometric representation, especially with large variation.

Assess uncertainty, group structure, and measurement repeatability. Check whether apparent differences reflect imaging pose, inconsistent landmark placement, sampling bias, or genuine shape variation. Visualization should show what a statistical change means geometrically rather than relying only on abstract scores.

Three practical examples

1. Manufacturing research: Corresponding points on repeated parts are analyzed to understand variation in proportions after separating placement differences from geometry.

2. Biological research: Landmark configurations describe how specimen shapes vary across a sampled population under an explicit correspondence protocol.

3. Computer vision: A statistical model of plausible shapes helps constrain object localization or segmentation when individual image observations are noisy.

Think of it this way

Think of tracing several leaves on transparent sheets, then moving, rotating, and perhaps resizing the tracings before comparing their outlines. Once irrelevant placement differences are removed, their proportions become easier to compare.

The analogy omits statistical uncertainty and the difficulty of matching corresponding points. It also shows why scale removal is a choice: resizing the tracings would make a study of absolute leaf size impossible from those normalized shapes alone.

Worked example

Consider landmark triangle A with coordinates (0,0), (2,0), and (0,1). Triangle B has coordinates (5,7), (9,7), and (5,9).

Subtracting (5,7) from B and dividing its coordinates by two produces exactly A. Thus B differs from A only by translation and uniform scale in this example, so they have the same shape when those transformations are excluded.

Triangle C has coordinates (0,0), (2,0), and (0,1.5). Its vertical-to-horizontal leg ratio is 1.5/2 = 0.75, compared with 1/2 = 0.5 for A. No uniform scaling and rigid motion can remove that proportional difference.

If the task is checking physical part dimensions, B may still be unacceptable because it is twice as large. A shape-only conclusion must therefore be reported with its transformation convention rather than treated as a complete dimensional assessment.

Illustrative case study

Fictional case: A design team studies variation in hand-cut packaging inserts using six landmarks. Initial coordinate differences are dominated by how the inserts were placed under the camera.

The team aligns their configurations and separately retains an overall size measure. It repeats annotations on a sample to estimate measurement variability before interpreting the remaining shape differences.

One landmark proves difficult to define on rounded corners, so its apparent variation is partly annotation noise. The team improves the protocol and presents mean shape and variation plots with the size results. This prevents a tidy statistical summary from hiding a poorly measured feature.

Common misconceptions and mistakes

Do not assume all geometric variation is shape variation. Translation, orientation, scale, and perspective can contribute different effects.

Do not normalize away size when size is part of the question. State whether the analysis concerns shape alone or a broader notion of form.

Do not treat landmark correspondence as a minor formatting detail. Comparing different physical locations can invalidate an otherwise careful analysis.

Do not assume the largest statistical variation direction is automatically meaningful or causal. It may reflect acquisition artifacts, annotation differences, or the composition of the sample.

Business use

Statistical shape analysis can help teams understand geometric consistency, compare designs, and build interpretable models of allowable variation. It is useful when the question concerns structured geometry rather than only whether an image belongs to a category.

The practical value depends on trustworthy capture and correspondence. Establish repeatability, preserve physical units and size measures where needed, and evaluate whether statistical differences matter operationally. A small statistically detectable change may be irrelevant to assembly, while a localized geometric deviation may matter even if it contributes little to total variance.

Check your understanding

Two triangular parts become identical after one is uniformly enlarged. Can a shape-only analysis establish that their physical dimensions meet the same specification?

Answer: No. Removing scale discards absolute size. Dimensional conformity requires retaining and checking the relevant physical measurements.

Related terms

Landmark point
Principal component analysis
Procrustes analysis
shape space
point distribution model
geometric morphometrics
centroid size

Learn next

Compare several landmark triangles before and after translation, rotation, and scale normalization. Keep a separate size measurement, then repeat landmark placement to see how annotation uncertainty contributes to the apparent shape variation.

Sources

Dryden and Mardia, Statistical Shape Analysis, author-hosted book overview and resources: https://people.stat.sc.edu/dryden/book/ . The primary author page was read for the field's geometric scope and treatment of location, rotation, and scale. The triangles and packaging case are original teaching examples.

Previous: Statistical semantics Next: Statistical significance