Geodesic Object Representation and Recognition

A. Ben Hamza, Hamid KrimLecture Notes in Computer Science 2003 · DOI · OpenAlex

Abstract

Comparing 3D objects is a computationally intensive task, often hindered by the limitations of Euclidean distance metrics when dealing with non-rigid or curved surfaces. This paper presents a novel framework for 3D object recognition that simplifies complex 3D meshes into one-dimensional "geodesic shape distributions." By leveraging intrinsic geodesic distances, the proposed method creates shape signatures that are invariant to rotation and bending. These signatures are then compared using information-theoretic measures, enabling efficient and accurate object matching.

Introduction: The Challenge of 3D Matching

Traditional 3D object comparison methods frequently rely on Euclidean distance—the straight-line distance between two points in space. While intuitive, this "as the crow flies" approach fails to respect the actual topology of nonlinear manifolds, such as a human face or a flexible hose. For instance, bending a 3D model of a hose significantly alters the Euclidean distance between its ends, even though the intrinsic shape remains unchanged.

To address this, the authors propose a descriptor that is compact, robust to noise, and, most importantly, intrinsic. By measuring distances "as the ant crawls" along the surface, the representation becomes isometrically invariant, capturing the true geometry regardless of the object's pose or orientation.

Comparison of Euclidean and geodesic distances on a curved manifold.

Methodology: From 3D Mesh to 1D Signature

The proposed pipeline transforms a complex 3D triangular mesh into a probabilistic 1D signature through several key steps:

  1. Feature Extraction: The object is represented by the centroids of its constituent triangles, providing a uniform set of representative points across the surface.
  2. Geodesic Distance Calculation: A neighborhood graph is constructed from these centroids. Dijkstra’s algorithm is then employed to calculate the shortest path along the surface between all pairs of points, yielding the geodesic distance.
  3. Global Geodesic Shape Function: For each point on the surface, the method calculates the average squared geodesic distance to all other points. This function effectively maps how "central" or "peripheral" a region is relative to the entire object.
  4. Kernel Density Estimation (KDE): These values are treated as a random sample to produce a smooth 1D probability density function (PDF), termed the Geodesic Shape Distribution.
  5. Dissimilarity Measure: Comparison between two objects is performed by calculating the Jensen-Shannon (JS) divergence between their respective PDFs. JS divergence offers a symmetric, bounded score between 0 (identical) and 1 (completely different), making it ideal for classification.

Experimental Results and Insights

The researchers evaluated their method using a diverse database of 3D models, including Genus-0 objects (sphere-equivalent, like animals) and Genus-1 objects (with one hole, like teapots). All models were normalized to 1,000 centroids to ensure consistency.

The results demonstrate that a 1D histogram of surface distances captures sufficient "topological DNA" to distinguish between different classes of objects effectively. The system achieved a matching rate of approximately 91% for animal models and improved to 98% for topological classes like pots and rings.

Examples of 3D models (T-rex, Triceratops, Lion, Horse) and their corresponding 1D geodesic shape distributions.

The move from extrinsic (Euclidean) to intrinsic (geodesic) metrics is the core innovation. Because the signature follows the surface, a model of a person standing or sitting produces nearly identical signatures, a property known as isometric invariance. Furthermore, treating shape as a distribution provides inherent robustness against minor mesh errors or scanning noise.

Limitations and Future Directions

While the comparison of 1D distributions is extremely fast, the initial computation of all-pairs shortest paths on dense meshes is computationally demanding. Additionally, as a global descriptor, the method is optimized for whole-object recognition rather than partial matching (e.g., identifying an object from a fragment).

Future work may involve parallelizing the geodesic calculations to enhance performance. The authors also suggest exploring more sophisticated information-theoretic measures to further refine the discriminative power of the geodesic signatures.

Summary updated

BibTeX
@inproceedings{hamza2003geodesic,
  author = {A. Ben Hamza and Hamid Krim},
  title = {{Geodesic Object Representation and Recognition}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2003},
  doi = {10.1007/978-3-540-39966-7\_36},
}
Figure