NeuroGF: A Neural Representation for Fast Geodesic Distance and Path Queries

Yohanes Yudhi Adikusuma, Ying He, Junhui Hou, Wenping Wang, Qijian Zhang

TL;DR

This paper introduces NeuroGF, a novel neural implicit representation for efficiently querying geodesic distances and shortest paths on 3D meshes. By learning a unified representation of both geometry and geodesics, NeuroGF significantly accelerates these computations, achieving near real-time performance for complex queries.

What Problem Does This Paper Address?

Computing geodesic distances and paths on 3D meshes is crucial for many geometry processing tasks, but traditional methods are often computationally expensive and slow, especially for applications requiring frequent, arbitrary point-to-point queries. While neural implicit representations have shown promise for 3D shape modeling, their application to geodesic computations has been largely unexplored. This paper aims to bridge this gap by developing a neural network that can directly predict geodesic properties, overcoming the limitations of existing methods in terms of speed and efficiency.

Methods

The core of NeuroGF is a neural network trained to represent geodesic information for a given 3D mesh. The authors propose an "overfitted" paradigm where a single neural network is trained for each specific mesh, allowing for high-fidelity representation. The network takes pairs of 3D points (source and target) as input and outputs:

  • Geodesic Distance: Predicted using a dedicated regression branch that takes the "feature-space difference" between source and target point embeddings.
  • Shortest Path: Modeled as a sequence of ordered 3D points, generated through a feature-guided curve deformation process. This involves deforming a straight line segment between the query points into a geodesic curve, guided by learned features.
  • Signed Distance Field (SDF): An auxiliary branch is included to predict the SDF, enabling a unified representation of both geometry and geodesics.

The learning objective combines supervision for geodesic distances, shortest paths, and signed distances, along with consistency constraints to ensure the predicted path length matches the predicted distance and that path points lie on the surface. The authors also explore generalizable learning frameworks by incorporating different feature encoders.

Datasets

The paper uses common 3D shape models for testing, including armadillo, bimba, bucket, bunny, cow, dragon, fandisk, heptoroid, maxplanck, and nail. For training, they generate ground-truth geodesic distances and shortest paths using DGG-VTP and fast discrete geodesic graphs (fDGG). For generalizable frameworks, they use the ShapeNet dataset.

Key Insights and Ideas

  • Neural Implicit Geodesics: This is the first work to represent geodesic distances and paths using neural implicit functions, opening a new avenue for efficient geodesic computation.
  • Unified Representation: Encoding both 3D geometry (via SDF) and geodesics within a single neural network offers a powerful and compact representation.
  • Speed and Accuracy Trade-off: NeuroGF achieves a compelling balance between high accuracy (mean relative error < 0.5%) and remarkable speed, outperforming traditional methods by orders of magnitude.
  • Generalizability: The framework can be extended to learn generalizable geodesic representations for unseen shapes and categories, demonstrating its versatility.
  • Overfitting for Fidelity: The initial "overfitted" approach, where a network is trained per mesh, proves highly effective for achieving accurate geodesic predictions.

Key Results

  • Speed: NeuroGF achieves query times of less than 1 ms for SSAD (single-source all-destination) on meshes with 400K vertices, significantly faster than traditional methods.
  • Accuracy: Mean relative errors for geodesic distances are consistently low, often below 0.5%, and comparable to or better than state-of-the-art methods like HM and fDGG.
  • Shortest Paths: Predicted shortest paths show high visual similarity to ground truth, with low Chamfer-L₁ errors.
  • Memory Efficiency: NeuroGF has a significantly smaller memory footprint for online queries compared to HM and fDGG, especially for larger meshes.
  • Robustness: The approach demonstrates robustness to anisotropic meshes, where other methods degrade.

Limitations and Future Work

  • Path Surface Adherence: There's no strict guarantee that predicted geodesic paths lie perfectly on the underlying surface. Post-projection/refinement techniques using SDFs are suggested to improve this.
  • Maximal Error: While mean errors are low, maximal errors can sometimes reach higher values (up to ~5%), suggesting potential for optimization strategies like online hard example mining.
  • Generalizable Frameworks: While promising, the generalizable frameworks still show some performance degradation compared to overfitted models, particularly for unseen categories. Further improvements in feature encoders are suggested.

Summary updated

BibTeX
@inproceedings{adikusuma2023neurogf,
  author = {Yohanes Yudhi Adikusuma and Ying He and Junhui Hou and Wenping Wang and Qijian Zhang},
  title = {{NeuroGF: A Neural Representation for Fast Geodesic Distance and Path Queries}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=cGdGh3Mp2W},
}
Figure