Local Polynomial Fitting for 3D Geometric Representation

20 concepts16 papers1 proposal

3D geometric data is inherently discrete. Whether captured as a sparse point cloud from a LiDAR sensor, structured as a polygon mesh, or tracked as a set of moving particles, it lacks the continuous, differentiable nature of the physical surfaces it represents. Traditional deep learning architectures often struggle with this irregularity, either relying on computationally heavy voxel grids or oversimplifying local neighborhoods by treating them as flat planes.

Assuming a local neighborhood is a flat plane fails catastrophically at edges, corners, and complex curvatures. To resolve this, a powerful research direction has emerged: treating discrete neighborhoods as continuous mathematical functions by fitting local polynomial surfaces directly to the data.

The Intuition: Taylor Expansions as Geometric Inductive Biases

Why do polynomials bridge the gap between discrete points and continuous geometry so effectively? The core intuition relies on the principle of a Taylor series expansion: any smooth, continuous surface can be approximated locally by a polynomial function. By framing local geometric feature extraction as a polynomial fitting problem, we introduce a rigorous mathematical inductive bias into geometric learning models.

In a Taylor expansion, different polynomial terms correspond to different geometric properties:

  • Low-order terms (0th and 1st order) capture the baseline topology, such as local position and spatial orientation (the tangent plane).
  • High-order terms (2nd, 3rd, and above) capture intricate structural variations—the curvatures, ridges, valleys, and sharp edges.

This mathematical decomposition allows neural networks to process irregular data far more efficiently. For example, Changshuo Wang et al. demonstrated in Taylor Series-Inspired Local Structure Fitting Network for Few-shot Point Cloud Semantic Segmentation that explicitly separating low-order terms for coarse shapes and high-order terms for fine-grained details allows models to learn highly descriptive features without requiring massive pre-training datasets.

This capability is further expanded by Changshuo Wang et al. in Dynamic Taylor Convolutional Neural Network for Few-Shot Point Cloud Semantic Segmentation, and by Changshuo Wang in the single-author work DyPolySeg: Taylor Series-Inspired Dynamic Polynomial Fitting Network for Few-shot Point Cloud Semantic Segmentation, which utilizes dynamic high-order kernels with learned spatial priors to capture complex variations on the fly.

Bridging Discrete Points and Continuous Calculus

One of the most compelling reasons to use polynomial fitting is that polynomials are analytically differentiable. On a raw point cloud, calculating curvature or normal vectors requires noisy, error-prone discrete approximations. By fitting a local polynomial—specifically an n-jet (a truncated Taylor expansion)—we can compute exact derivatives using standard calculus.

Discrete Points  ───(Weighted Least Squares)───►  Continuous Polynomial (n-jet)
     •   •                                              ┌─┐
   •   •   •     ─────────────────────────────►        /   \   (Analytically Differentiable)
     •   •                                            └────┘

The authors of DeepFit: 3D Surface Fitting via Neural Network Weighted Least Squares leveraged this by combining deep learning with classical geometry. They configured a neural network to learn point-wise weights for a Weighted Least Squares (WLS) polynomial fit, allowing the system to analytically derive principal curvatures and normals directly from the fitted coefficients.

For dynamic shapes, Ruochen Chen et al. showed in Patch-based Representation and Learning for Efficient Deformation Modeling that patch-based local n-jet functions provide a highly compact and computationally efficient way to model complex deformations, drastically reducing parameter counts while maintaining fast inference times.

This principle also extends to classical physics simulations; Hongkai Zhao demonstrated in A grid based particle method for moving interface problems that local interface reconstruction via least-squares polynomial fitting of neighboring particles enables accurate tracking of moving boundaries and fluid curvatures during complex topological changes.

Polynomial Convolutions and Transformers

To embed this mathematical intuition directly into deep learning pipelines, researchers have redesigned standard network operators to natively perform polynomial fitting.

Traditional convolutional neural networks (CNNs) rely on fixed grid kernels that cannot handle irregular point distributions. To translate convolutions to non-Euclidean domains, Yifan Xu et al. introduced SpiderCNN: Deep Learning on Point Sets with Parameterized Convolutional Filters, which defines convolutional filters as a product of step functions and Taylor polynomials. This allows the network to interpolate features smoothly across irregular neighborhoods.

In the context of attention-based architectures, Changshuo Wang et al. proposed GPSFormer: A Global Perception and Local Structure Fitting-Based Transformer for Point Cloud Understanding, which pairs global context with Taylor series-inspired convolutions to explicitly encode local surface structures into learned polynomial coefficients.

For mesh structures, Yang Zheng et al. demonstrated in 3D Mesh Model Classification with a Capsule Network that high-order polynomial parameters serve as highly descriptive feature descriptors for capsule networks, effectively capturing intricate local geometry despite irregular mesh topologies.

Applications Across Domains

Moving beyond flat or linear assumptions unlocks massive performance gains across diverse domains:

Upsampling and Reconstruction

Instead of directly predicting noisy point coordinates, the architecture proposed by Pingping Cai et al. in Parametric Surface Constrained Upsampler Network for Point Cloud estimates coefficients for local bicubic polynomial functions. This forces the network to generate points that lie on a smooth mathematical surface, eliminating outliers. For sparse reconstruction, Junhui Hou et al. utilize quadratic polynomial fitting in GeoUDF: Surface Reconstruction from 3D Point Clouds via Geometry-guided Distance Representation to stabilize unoriented normal vectors.

Mapping and Topography

In large-scale robotics, Jianping Li et al. show in Graph Optimality-Aware Stochastic LiDAR Bundle Adjustment with Progressive Spatial Smoothing that projecting point clouds onto second-order polynomial surfaces provides far more robust feature representations in non-structured environments than traditional flat-plane assumptions. Similarly, Jia Li et al. in A New Method to Estimate Changes in Glacier Surface Elevation Based on Polynomial Fitting of Sparse ICESat - GLAS Footprints replace planar assumptions with higher-order polynomials to accurately model complex mountainous glacier terrain from sparse altimetry data.

Topographic Feature Classification

Fitting local polynomials allows researchers to classify complex topologies into primitive geometric labels (such as ridges, valleys, and pits). Jun Wang et al. applied this to 3D Facial Expression Recognition Based on Primitive Surface Feature Distribution by fitting smooth cubic polynomials to mesh neighborhoods, achieving view-invariant tracking. In neuroimaging, Tuo Zhang et al. demonstrated in Parametric Representation of Cortical Surface Folding Based on Polynomials that a compact 3rd-order polynomial with just four coefficients can successfully parameterize and classify complex cortical surface folding patterns in the human brain.

Key Foundational Work

3D Facial Expression Recognition Based on Primitive Surface Feature Distribution (2006) — contributed 1 concept elements to this direction.
R-D analysis of adaptive edge representations (2002) — 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_polynomial_surface_approximation_0d759aa1,
  title = {polynomial surface approximation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/polynomial-surface-approximation/rdr_0d759aa11b7f44d1a80cac9bb9ac8586},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_polynomial_surface_approximation_0d759aa1_v1,
  title = {polynomial surface approximation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/polynomial-surface-approximation/rdr_0d759aa11b7f44d1a80cac9bb9ac8586/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake