Decomposing Complex Shapes into Geometric Primitives

51 concepts45 papers1 proposal

The physical world is composed of complex, continuous, and often highly irregular geometries. When a robot or computer vision system observes this world via sensors—whether through laser range finders, RGB-D streams, or LiDAR—it receives an unstructured, noisy torrent of data: millions of individual points or pixels. Processing this raw data directly is computationally punishing and semantically shallow.

To bridge this gap, a powerful paradigm has emerged: geometric primitive decomposition. Instead of treating a scene as an arbitrary collection of points, this approach abstracts the environment into a structured vocabulary of simple, mathematically well-defined shapes—such as planes, cylinders, spheres, quadrics, and convex hulls. By breaking down complex structures into these fundamental building blocks, vision and robotic systems drastically reduce computational complexity, filter out sensor noise, and unlock high-level reasoning.

Suppressing Noise Through Structural Aggregation

Raw sensor data is inherently noisy. If a system calculates geometric properties, such as surface normals, point by point, minor sensor fluctuations can lead to massive errors. Geometric primitive decomposition solves this by acting as a spatial aggregator.

When a system fits a mathematical primitive (like a cylinder or a plane) to a cluster of points, it effectively averages out local noise across the entire surface. For instance, Porrill et al. demonstrated that representing 3D scenes via lines and circular arcs provides robust object recognition even in cluttered, noisy environments.

Similarly, Somani et al. showed that deriving surface normals from fitted geometric primitives rather than individual points yields a nearly fifty-fold reduction in the number of hypotheses required for 3D object matching, accelerating processing speeds significantly. By anchoring perception to stable mathematical abstractions, systems become resilient to partial occlusions and sensor imperfections, as further explored by Somani et al. and Wang et al..

Taming Complexity with Discrete Vocabularies

Matching an observed 3D point cloud against a reference model is an optimization nightmare if done at the individual point level. Decomposing the scene into primitives transforms this continuous, high-dimensional search into a discrete matching problem.

By segmenting raw data into planes or quadric surfaces, early foundational work by Faugeras et al. turned complex 3D pose estimation into manageable linear least-squares and quaternion optimizations. Rather than comparing millions of points, the algorithm matches a handful of primitive surfaces.

This concept of a "structural vocabulary" extends across domains. Shrivastava et al. clustered surface normals into a dictionary of geometric primitives to define object parts, while Kang et al. utilized learned complex filters to identify elementary curves and planes within complex interferometric radar data. In modern SLAM systems, Davison et al. utilized quadric surfaces as high-level landmarks, allowing back-end optimizers to map structured environments with a fraction of the data required by dense methods.

Constraining Ill-Posed Problems via Inductive Biases

Inferring 3D structure from limited data—such as a single 2D image or a line drawing—is fundamentally ill-posed; infinite 3D shapes could project to the same 2D representation. Geometric primitives introduce a powerful inductive bias that constrains the solution space to physically plausible shapes.

For example, Zhang et al. enabled 3D insect modeling from a single 2D viewpoint by fitting generalized cylinders and ellipsoids to segmented body parts, making the reconstruction mathematically tractable. Similarly, Zou et al. decomposed complex 2D line drawings into simpler geometric primitives, allowing optimization algorithms to solve for small, manageable parts rather than the entire intricate structure at once.

This principle remains vital in modern neural rendering. Held et al. replaced unstructured or ellipsoidal elements in radiance fields with smooth 3D convexes, showing that primitives aligning with physical boundaries yield sharper edges and higher visual quality with fewer parameters.

Bridging Perception to High-Level Reasoning and Labeling

Ultimately, primitive decomposition serves as a translation layer between raw perception and semantic understanding. It converts spatial coordinates into a symbolic language that humans, downstream models, and even large language models (LLMs) can easily interpret.

In data annotation, encapsulating point clouds within primitives like cuboids, cylinders, or lines drastically reduces human effort. Ibrahim et al. and Unal et al. demonstrated that primitive-based labeling cuts manual annotation time by up to 90% compared to point-by-point selection. For downstream semantic tasks, Jones et al. and Hu et al. showed that partitioning scenes into geometrically consistent clusters or primitives provides excellent spatial priors for few-shot and open-vocabulary segmentation.

Perhaps the clearest distillation of this idea's power is found in robotic manipulation. The authors of ShapeGrasp: Zero-Shot Task-Oriented Grasping with Large Language Models through Geometric Decomposition decomposed complex, unfamiliar objects into basic primitives like rectangles and ellipses. By converting raw sensor data into this standardized vocabulary of shapes, they enabled LLMs to apply physical commonsense reasoning to zero-shot grasping tasks. Long before LLMs, Aida et al. recognized this exact potential, decomposing geometric features into logical propositions and decision trees to achieve object recognition through binary spatial truths.

Key Foundational Work

Recognition Of An Object in A Stack Of Industrial Parts (1975) — contributed 1 concept elements to this direction.
Visual-Tactile Symbiotic System for Stereometric Pattern Recognition (1971) — contributed 1 concept elements to this direction.
Forming Models Of Plane-And-Cylinder Faceled Bodies From Light Stripes (1975) — 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_geometric_primitive_decomposition_9435d884,
  title = {geometric primitive decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-primitive-decomposition/rdr_9435d884768345a49e7bc5702ba76501},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_geometric_primitive_decomposition_9435d884_v1,
  title = {geometric primitive decomposition},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometric-primitive-decomposition/rdr_9435d884768345a49e7bc5702ba76501/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake