Shape Abstraction Using Parameterized Geometric Primitives

58 concepts48 papers1 proposal

When humans look at a chair, we do not perceive a collection of millions of independent points or a dense triangle mesh. Instead, our minds naturally decompose the object into a meaningful arrangement of parts: a flat rectangular board for the seat, another for the backrest, and four elongated cylinders for the legs.

Traditional computer vision and deep learning models often force algorithms to operate on unstructured point clouds, dense voxel grids, or complex meshes. While highly detailed, these representations suffer from an informational bottleneck: they are high-dimensional, computationally expensive, and lack inherent semantic structure.

As Pentland argued in the seminal paper Parts: Structured Descriptions of Shape, moving away from "pixels and edges" toward parameterized geometric primitives drastically reduces the informational cost of a scene. This mathematical compression transforms complex spatial reasoning tasks—which are often NP-complete when using raw data—into computationally tractable problems.

The Mathematics of Morphable Primitives

If we restrict our geometric vocabulary to rigid, basic shapes like standard cubes or spheres, we quickly run into a representation problem. Real-world objects are full of rounded corners, tapered edges, and irregular bulges. Forcing a model to fit these nuances using rigid shapes requires packing an excessive number of primitives together, destroying the simplicity we set out to achieve.

The core breakthrough that makes parametric shape abstraction work is the use of morphable mathematical families, most notably superquadrics.

A superquadric is an implicit surface defined by an algebraic equation where the exponents are not fixed integers, but continuous variables known as shape exponents or squareness parameters. By adjusting these parameters, a single mathematical function can smoothly transition across a continuous spectrum of geometries:

$$\left( \left| x \right|^{\frac{2}{\epsilon_2}} + \left| y \right|^{\frac{2}{\epsilon_2}} \right)^{\frac{\epsilon_2}{\epsilon_1}} + \left| z \right|^{\frac{2}{\epsilon_1}} = 1$$

By varying $\epsilon_1$ and $\epsilon_2$, the shape morphs seamlessly from a perfect sphere to a cylinder, and all the way to a sharp-edged box. Early work by Solina et al. in Range Image Interpretation of Mail Pieces with Superquadrics demonstrated that these parameters allow a single model to absorb the "messiness" of real-world objects, such as the bulging sides of a parcel or the rounded edges of a tube.

  [ ε₁=1, ε₂=1 ]  ------>  [ ε₁=0.1, ε₂=1 ]  ------>  [ ε₁=0.1, ε₂=0.1 ]
     Sphere                    Cylinder                     Cuboid

This continuous flexibility is highly efficient. Zuo et al. demonstrated in QuadricFormer: Scene as Superquadrics for 3D Semantic Occupancy Prediction that morphable primitives eliminate the need to pack multiple rigid ellipsoids to represent sharp corners or elongated structures, significantly lowering the computational latency and memory footprint required to model complex driving environments.

Other variations expand this parametric vocabulary beyond superquadrics:

Learning to Decompose Scenes Unsupervised

How do we automatically break an object down into these primitives without requiring human annotations for every part? Modern frameworks achieve this by framing primitive assembly as an unsupervised reconstruction or rendering problem.

The foundational deep learning approach, introduced by Tulsiani et al. in Learning Shape Abstractions by Assembling Volumetric Primitives, uses a neural network to predict the shape parameters (dimensions, position, orientation, and an existence probability) for a fixed budget of primitives. By enforcing a loss function that penalizes the discrepancy between the assembled primitives and the target 3D shape, the network naturally learns to discover consistent, interpretable part decompositions across entire object categories.

To move beyond simple bounding boxes, Paschalidou et al. (Superquadrics Revisited: Learning 3D Shape Parsing beyond Cuboids) derived a novel analytical loss function based on the radial distance function of superquadrics. This mathematical formulation allows neural networks to gradient-optimize shape exponents directly. The authors later extended this in Learning Unsupervised Hierarchical Part Decomposition of 3D Objects from a Single RGB Image to predict a tree structure of parts, allowing a model to understand how fine-grained components structurally compose a larger object.

With the rise of differentiable rendering, this decomposition can even be learned directly from 2D images without 3D supervision. For example, Monnier et al. (Differentiable Blocks World: Qualitative 3D Decomposition by Rendering Primitives) leverage a transparency-based optimization framework to render superquadric primitives. By minimizing the difference between the rendered primitives and input images, they bypass the unstructured "soup of particles" typical of radiance fields, yielding compact and editable 3D models.

Why Primitives are Actionable

The ultimate value of a parametric primitive representation is that its parameters are directly "actionable." A raw point cloud tells a system where matter exists in space, but it provides no structural insight. In contrast, a primitive representation explicitly exposes size, orientation, and shape type.

Robotics and Manipulation

As Wu et al. (Primitive-Based Shape Abstraction via Nonparametric Bayesian Inference) note, decomposing objects into parametric primitives yields low-dimensional parameters that provide immediate semantic data for object manipulation and spatial reasoning.

This low-dimensional nature enables real-time performance where dense data would fail:

Furthermore, the implicit mathematical functions of these primitives make calculating physical constraints trivial. Liu et al. (Primitive-Based 3D Human-Object Interaction Modelling and Programming) demonstrate that evaluating surface distances via implicit equations is significantly faster than computing mesh-based signed distance fields, vastly simplifying the modeling of contact constraints in human-object interactions.

Object-Centric SLAM and Mapping

This compact abstraction is equally powerful for spatial mapping and localization. Traditional Simultaneous Localization and Mapping (SLAM) tracks thousands of individual points or planes. Object-centric SLAM instead treats semantic objects as first-class geometric landmarks.

By representing objects as dual quadrics (ellipsoids), frameworks like those proposed by Hosseinzadeh et al. (Real-Time Monocular Object-Model Aware Sparse SLAM) and Liao et al. (RGB-D Object SLAM Using Quadrics for Indoor Environments) can encapsulate an object's position, orientation, and spatial extent into a single, mathematically elegant landmark. These ellipsoidal abstractions plug directly into geometric pose-graph optimization, mitigating scale drift without the computational overhead of dense tracking.

Raw Point Cloud Map                  Object-Centric Quadric Map
   . . : : . . .                        _______
 . : : :: : : . .                      /       \
  . . : : . . .       ======>         |  Dual   |  --> Single landmark in
 . : : :: : : . .                      \_______/      pose-graph optimization
(Thousands of points)                (1 Ellipsoid)

On a massive scale, such as autonomous driving, Xia et al. (Quadric Representations for LiDAR Odometry, Mapping and Localization) showed that decomposing heavy LiDAR point clouds into sparse patches fitted to quadric implicit functions (like cylinders and hyperboloids) slashes map storage requirements by over 1,500x while maintaining high-fidelity localization accuracy.

By replacing raw data with the elegant language of geometry, parametric shape abstraction bridges the gap between low-level perception and high-level, actionable reasoning.

Key Foundational Work

Parts: Structured Descriptions of Shape (1986) — contributed 2 concept elements to this direction.
Description and Recognition of Curved Objects (1977) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_superquadric_representation_94496606,
  title = {superquadric representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/superquadric-representation/rdr_94496606dd3f43038337ad347afb50ab},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_superquadric_representation_94496606_v1,
  title = {superquadric representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/superquadric-representation/rdr_94496606dd3f43038337ad347afb50ab/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake