The Euclidean Geodesic Discrepancy in 3D Vision

40 concepts40 papers1 proposal

Imagine a 3D model of a hand with fingers curled tightly into a fist, or a long-sleeved coat pressing against a torso. To a human observer, the palm and fingers, or the sleeve and torso, are distinct surfaces separated by empty space. However, to a standard 3D machine learning model, they present a fundamental trap: spatial proximity ambiguity.

Because these distinct surfaces are physically close in 3D space, algorithms that rely on straight-line Euclidean distance often treat them as a single, continuous entity. This phenomenon—where extrinsic spatial closeness overrides intrinsic topological separation—is known as the Euclidean-geodesic discrepancy, or topological proximity error. It represents a pervasive failure mode across 3D reconstruction, segmentation, and shape analysis.

The Core Conflict: Extrinsic vs. Intrinsic Space

To understand why this error occurs, we must look at how computers represent and navigate 3D shapes. Most models operate in extrinsic Euclidean space ($\mathbb{R}^3$), where the distance between two points is a straight line. However, the data itself often lives on an intrinsic 2D manifold (the surface of the object), where distance should be measured along the surface (geodesic distance).

        Euclidean Distance (Short straight line)
     [Surface A] ------------------------> [Surface B]
          |                                    ^
          |                                    |
          +====== Geodesic Path ========>======+
                 (Long path along the manifold)

This tension was recognized early in computer vision history. In their foundational work, Gerald J. Agin et al. (Computer Description of Curved Objects) noted that geometric segmentation algorithms routinely struggle when unconnected parts of an object are in close physical proximity, leading to the erroneous merging of separate components.

Decades later, the problem remains acute. For instance, Herve Lombaert et al. demonstrated this failure mode vividly in cortical parcellation (Spectral Forests: Learning of Surface Data, Application to Cortical Parcellation): when mapping the highly convoluted folds of the human brain, extrinsic coordinates cause severe spatial ambiguity because points on opposite sides of a deep surface fold are physically adjacent but anatomically and functionally distinct. Similarly, Flora Ponjou Tasse et al. showed in Cluster-Based Point Set Saliency that clustering based on Euclidean distance erroneously groups surfaces like a palm and fingers, blending distinct geometric features and degrading local descriptors.

How the Error Manifests Across Domains

The discrepancy between Euclidean and geodesic metrics causes severe artifacts across three major areas of 3D deep learning.

1. Thin-Sheet and Thin-Walled Geometries

When a structure is exceptionally thin (such as sheet metal or a plastic shell), the distance between its opposing surfaces is often smaller than the local neighborhood radius used by processing algorithms. As Jiating Chen et al. observed in Bilateral blue noise sampling, points on opposite sides of a thin surface are spatially proximate but possess opposing surface normals. Standard spatial sampling treats them identically, leading to holes and jagged edges.

During optimization or denoising, this cross-surface interference can cause the geometry to collapse entirely. As Dong Xiao et al. explained in Winding Clearness for Differentiable Point Cloud Optimization, local neighborhood analysis incorrectly groups points from opposing sides of a thin structure during smoothing, causing the distinct boundaries to blend and collapse into a single non-existent structure.

2. Feature Bleeding and Segmentation Leaking

In 3D scene understanding, standard operations like voxel convolutions or self-attention mechanisms aggregate information from spatial neighborhoods. Xuyang BAI et al. (VMNet: Voxel-Mesh Network for Geodesic-Aware 3D Semantic Segmentation) showed that voxel networks blur boundaries between physically touching but semantically distinct objects (such as a chair resting against a wall), treating separate entities as a single unit.

When projecting 2D detections into 3D, this proximity causes "segmentation leaking," where distinct parts are erroneously merged because standard projection techniques do not respect the intrinsic topological structure of the mesh surface (Ahmed Abdelreheem et al. in SATR: Zero-Shot Semantic Segmentation of 3D Shapes). Even advanced architectures like Transformers suffer from this; Zhengyu Li et al. (Geodesic Self-Attention for 3D Point Clouds) noted that standard attention mechanisms treat points on opposite sides of a thin gap as highly related, ignoring the true surface flow.

3. Topological Self-Contact in Deformable Shapes

For articulated shapes like humans or animals, motion brings different body parts into contact. Hyomin Kim et al. (Deep Virtual Markers for Articulated 3D Shapes) and Dan Raviv et al. (DPC: Unsupervised Deep Point Correspondence via Cross and Self Construction) highlighted that when a hand touches a hip, Euclidean-based supervision fails, leading to severe tracking and correspondence errors. The model cannot distinguish between the surfaces in contact, causing visual "bleeding" of textures or geometry across distinct anatomical regions (Gusi Te et al. in Neural Capture of Animatable 3D Human from Monocular Video).

Emerging Solutions: Injecting Topological Awareness

To resolve this ambiguity, researchers are developing methods that explicitly bridge or replace Euclidean metrics with topology-aware mechanisms.

The Intrinsic-Extrinsic Ratio (IER)

One elegant way to detect topological boundaries is to compare the straight-line distance between two points to their surface distance. Minghua Liu et al. (Meshing Point Clouds with Predicted Intrinsic-Extrinsic Ratio Guidance) trained neural networks to predict this Intrinsic-Extrinsic Ratio (IER).

$$\text{IER} = \frac{\text{Geodesic Distance}}{\text{Euclidean Distance}}$$

For points on a flat, continuous surface, the IER is close to $1$. For points on opposite sides of a fold or thin gap, the Euclidean distance is tiny, but the geodesic distance is massive, causing the IER to spike. By predicting this ratio, models can accurately determine which points should form triangular faces during mesh generation without performing expensive global geodesic calculations.

Intrinsic Coordinate Embeddings

Instead of feeding raw $(x,y,z)$ coordinates into a network, models can map points into a geometry-aware coordinate system. Lukas Koestler et al. (Intrinsic Neural Fields: Learning Functions on Manifolds) utilized embeddings based on the Laplace-Beltrami Operator (LBO). Because the LBO depends strictly on the metric tensor of the surface, functions learned within this embedding space are naturally aware of the shape's geometry and invariant to bending, completely eliminating cross-surface bleeding artifacts.

Alternative Geometric Primitives

Moving away from point-based representations can also bypass the issue entirely. Jianwei Hu 0003 et al. (MAT-Net: Medial Axis Transform Network for 3D Object Recognition) used the Medial Axis Transform (MAT)—a "skeleton-and-skin" representation—to capture topological connections explicitly via medial spheres.

Meanwhile, Siyu Ren et al. (Shape as Line Segments: Accurate and Flexible Implicit Surface Representation) replaced traditional point-based distance fields with a Line Segment Field. Instead of prioritizing the single nearest surface point (which fails near complex or thin structures), their method evaluates if and where a line segment intersects a surface, providing a mathematically elegant way to preserve connectivity information and open boundaries.

Key Foundational Work

Computer Description of Curved Objects (1973) — contributed 1 concept elements to this direction.
Landmark Correspondence Optimization for Coupled Surfaces (2007) — 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_topological_proximity_error_05158549,
  title = {topological proximity error},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-proximity-error/rdr_051585499a994d8bbc72a713bffb9f61},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_topological_proximity_error_05158549_v1,
  title = {topological proximity error},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-proximity-error/rdr_051585499a994d8bbc72a713bffb9f61/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake