Are superquadrics learned only by neural networks?

No. They are mathematical shape models that can be created directly or fitted using various optimization or learned methods.

Can they represent rings?

Supertoroid members of the broader family can represent ring-like forms; a basic superellipsoid cannot represent that topology by itself.

Do their parameters describe object material?

No. Geometry parameters do not automatically specify stiffness, texture, weight, or other physical properties.

AI Glossary / Computer vision and multimodal AI / Intermediate

Superquadrics

Definition

Superquadrics are a family of parameterized three-dimensional shapes that generalize familiar forms such as ellipsoids. By changing a small number of size and shape parameters, they can describe rounded, box-like, pointed, and other related surfaces. They are used as compact geometric models in computer graphics, shape approximation, and computer vision.

What it means

A detailed mesh may contain thousands of vertices. A superquadric instead describes a shape through a mathematical formula and a relatively small parameter set. This makes it convenient to adjust broad proportions and roundness or to fit a simplified shape to measured data.

The family includes different forms, and authors' terminology can vary. Superellipsoids are a common example, while supertoroids describe related ring-like forms. A single superquadric does not represent every possible object topology or fine surface detail.

Shape parameters describe geometry, while pose parameters place the shape in the scene. Keeping those roles separate helps distinguish a wider object from an unchanged object viewed or positioned differently.

How it works

Choose a superquadric family suitable for the object or object part. Set size parameters along the coordinate axes and shape exponents that control the surface's broad form. Add rotation and translation to place it in the required coordinate frame.

For graphics, evaluate a parametric or implicit representation to generate a surface or test spatial relationships. For vision, estimate parameters from observations such as a point cloud or segmented depth image.

Fitting typically requires an objective, an initialization, and constraints on plausible parameters. Partial visibility, outliers, and multiple possible decompositions can make the estimate ambiguous. An algebraic residual from an implicit equation is not automatically an exact Euclidean distance to the surface.

Evaluate whether the simplified model preserves the properties needed by the application. Several superquadrics can be combined to approximate a more complex object, but selecting the number of parts and their relationships introduces additional modeling choices.

Three practical examples

1. Graphics authoring: A designer adjusts a few parameters to create rounded blocks or ellipsoid-like forms for a stylized scene.

2. Robot perception research: A system approximates object parts with compact shapes to support reasoning about broad geometry and candidate interactions.

3. Shape analysis: A depth scan is summarized by fitted size and roundness parameters when those properties are more useful than every individual measured point.

Think of it this way

Think of a family of adjustable molds. One control changes length, another changes width, and others change whether the mold looks rounded or more squared off.

The analogy captures compact parameter control, but real superquadrics are mathematical surfaces with particular limits. Adjusting a mold cannot automatically create arbitrary handles, holes, or intricate detail unless the selected family or a composition of shapes supports them.

Worked example

Consider a simple equal-exponent superellipsoid-style family defined by |x/a|^p + |y/b|^p + |z/c|^p = 1, with positive sizes a, b, c and exponent p. This is a restricted illustrative member, not the most general superquadric formula.

When a = b = c = 1 and p = 2, the equation is the unit sphere. At points with x = y = z = q, the surface condition becomes 3q^p = 1, so q = 3^(-1/p).

For p = 2, q is approximately 0.577. For p = 4, q is approximately 0.760. The surface extends farther toward the cube's corners while its axis intercepts remain at one, making the p = 4 form more box-like.

Some literature uses reciprocal exponent conventions, so smaller values of its named shape parameter may correspond to greater squareness. Always compare formulas, not parameter names alone.

Illustrative case study

Fictional case: A warehouse robotics team approximates boxed and rounded packages with fitted superellipsoids. The compact models are easy to store and visualize, but a package with a protruding handle is poorly represented by a single shape.

The team flags the residual structure and tests a multi-part model rather than pretending the handle is absent. It compares the approximation with the original depth points before using it in downstream planning research.

The project treats the fitted shape as a useful abstraction, not a complete physical description. Material properties, contact behavior, and unseen surfaces remain separate questions that the shape parameters alone do not answer.

Common misconceptions and mistakes

Do not confuse superquadrics with ordinary quadrics. The generalized exponents allow a broader family than conventional second-degree surfaces.

Do not assume a single compact shape can represent any object faithfully. Handles, holes, concavities, and multiple parts may require another family or decomposition.

Do not compare exponent values across implementations without checking their conventions. Equivalent geometry may use differently defined parameters.

Do not treat a low fitting loss as proof of correct hidden geometry. Partial observations can support multiple plausible shapes, especially when the object is heavily occluded.

Business use

Superquadrics offer a tradeoff between simplicity and geometric detail. They can reduce storage and make broad shape properties interpretable, which is useful when a downstream task needs approximate dimensions and structure rather than a photorealistic surface.

Evaluate approximation errors where they matter to the task. A small protrusion may contribute little to an average residual while being operationally important. Retain access to source measurements and document what the model omits, particularly when fitted geometry informs physical interaction or dimensional decisions.

Check your understanding

In the worked equal-exponent family, the axis intercepts stay fixed while p increases from two to four. Does the shape simply become a larger sphere?

Answer: No. Its proportions relative to diagonal directions change, making it more box-like while preserving those axis intercepts.

Related terms

Implicit surface
Superellipsoid
supertoroid
parametric surface
shape fitting
point cloud
geometric primitive
object decomposition

Learn next

Plot two-dimensional superellipse cross-sections with several exponent values before exploring three-dimensional forms. Then fit a simple shape to noisy points and inspect local residuals, especially around protrusions and missing regions.

Sources

Barr, Superquadrics and Angle-Preserving Transformations (1981), primary Caltech publication record and abstract: https://authors.library.caltech.edu/records/rtr62-f2882 . The record was read; its PDF download was unavailable in this session. The explicit equal-exponent example and warehouse case are original illustrations.

Previous: Superpixels Next: Supervised Biclustering