LiteGE: Lightweight Geodesic Embedding for Efficient Geodesics Computation and Non-Isometric Shape Correspondence

AI summary

Overview

• Research area: 3D computer vision and geometry processing, specifically geodesic distance computation and non-isometric shape correspondence.
• Technical level: Intermediate. Familiarity with 3D shape representations, PCA, and basic neural network architectures helps, but the core ideas are explained without heavy differential geometry.
• Scope: The paper introduces a lightweight, PCA-based alternative to heavy neural backbones for predicting geodesic distances on 3D shapes, and applies it to fast non-isometric shape matching.

What This Paper Is About

Computing geodesic (surface) distances on 3D shapes is a core primitive for shape matching, segmentation, and reconstruction, but classical algorithms are slow and recent learning-based methods (like NeuroGF and GeGNN) demand large 3D backbones with gigabytes of GPU memory. LiteGE asks whether the heavy backbone is actually necessary: it replaces it with a compact shape descriptor built from PCA applied to an unsigned distance field (UDF) around the shape. The goal is to match or beat prior accuracy while reducing memory by up to 300× and inference time by up to 100×, and to extend this efficiency to non-isometric shape matching between different shapes.

Key Contributions

1. UDF-PCA shape descriptor: A lightweight shape embedding built by applying Principal Component Analysis to unsigned distance field samples taken at informative voxels around a canonically aligned shape, retaining only 50–400 dimensions instead of high-dimensional neural features.
2. Backbone-free geodesic prediction: A small MLP-based network (~600K parameters, roughly 10% of GeGNN) that predicts geodesic distances without any large 3D convolutional or graph backbone, enabling interactive and memory-constrained use.
3. Robustness on sparse and degraded inputs: The method works on point clouds with as few as 300 samples, on noisy inputs, on meshes with missing regions, and on inputs with up to 40% of faces removed, where prior methods fail outright.
4. Geodesic-supervised non-isometric matching: A coarse-to-fine matching scheme that uses predicted geodesic distance as the primary supervisory signal between two different shapes, achieving accuracy comparable to the mesh-based state-of-the-art SMS while being up to 1000× faster.

Main Findings

• Dramatic memory and speed reductions: LiteGE uses up to 300× less memory and is over 100× faster than NeuroGF for geodesic inference. At 96 simultaneous queries, NeuroGF fails with out-of-memory while LiteGE completes using 51.8 MB.
• Strong accuracy on sparse point clouds: On 300-point models, LiteGE achieves a mean L1 error of 3.4 (ground-truth normalized to mean 100), while NeuroGF produces an error of 75. At 2,000 points, LiteGE's error is 2.3 versus NeuroGF's 13.
• Compact descriptors are sufficient: The first 50 principal components of the UDF representation already capture over 95% of the UDF variance, confirming strong redundancy in the full 20K–30K dimensional representation.
• Competitive non-isometric matching: On template meshes, LiteGE achieves comparable geodesic error (2.5 vs. 2.2) and AUC (79.3 vs. 79.4) to SMS while running in 185 ms versus SMS's 21,300 ms.
• Generalization to remeshed and broken shapes: On remeshed models, LiteGE slightly outperforms SMS (7.2 vs. 7.44 error), and on broken meshes with 40% of faces removed, SMS collapses (AUC 33.2, error 28) while LiteGE holds up (AUC 69.5, error 8).
• Generalization across datasets: When trained on SMAL and SURREAL, LiteGE generalizes to the unseen FAUST human dataset with a median L1 error of 2.6, using only 700 test points. It also generalizes across diverse categories when trained on Objaverse-XL.
• Robustness to noise and occlusion: Adding 0.01-scale Gaussian noise yields a mean L1 error of 4.3; removing 10% and 15% of points yields errors of 3.7 and 4.5 respectively.
• Efficient path tracing: Gradient-based backtracking reconstructs geodesic paths on sparse point clouds of under 1,000 points with a memory footprint below 5 MB.
• Predicting residuals helps: Training the network to predict the difference between Euclidean straight-line distance and geodesic distance improves accuracy by 5–7% over direct geodesic prediction.

Methodology in Plain English

The method rests on the observation that if you constrain yourself to a known set of shape categories (animals, humans, faces), shapes share a lot of structure—so a heuristic, task-specific descriptor can replace a heavy learned backbone. The pipeline has three stages:

1. Canonicalize the shape. Every input (mesh or point cloud) is centered at the origin, scaled to a consistent size (fixed bounding-box surface area for regression, unit surface area for matching), and oriented. For orientation, geodesic regression uses PCA-aligned principal axes, while shape matching uses a trained T-Net that predicts a rotation matrix, supervised by a geodesic angular loss plus an L2 coordinate loss.

2. Build the UDF-PCA descriptor. The canonicalized shape is voxelized at 128³. Voxels whose occupancy variance across the training set exceeds a threshold are kept as "informative voxels"—these cluster near the surface. For each informative voxel, the unsigned distance to the shape surface is computed, producing a 20K–30K dimensional vector. PCA reduces this to a 50–400 dimensional descriptor, and 50 components already retain 95% of the variance.

3. Predict with small MLPs. The descriptor passes through a shape MLP to produce a global shape embedding. Query point coordinates pass through a shared coordinate MLP. Each point embedding is concatenated with the shape embedding and fed through another shared MLP to produce shape-specific point embeddings. The difference between the source and destination point embeddings is fed to a final MLP that predicts the geodesic distance (or the residual over the straight-line distance).

For shape matching, both shapes' descriptors are fed into the same shared shape MLP, and the network predicts the average geodesic distance between a point on shape X and a point on shape Y as if both lay on a common surface. Matching proceeds coarse-to-fine using a precomputed multi-tier nearest-neighbor cache: start from a sparse candidate set on the target shape, find the candidate minimizing predicted geodesic distance, then expand to its nearest neighbors in the next tier, repeating until the finest tier. Only a small number of network queries are needed per match.

For path tracing, the destination point is iteratively updated along the negative gradient of the predicted geodesic distance and projected back onto the surface until convergence.

Why This Matters

This work challenges the assumption that accurate geodesic prediction requires large 3D backbones, showing that for category-constrained shape collections, a fixed PCA-based descriptor plus a tiny MLP can match or beat much heavier models. This has direct implications for how the field designs shape-aware networks—particularly for deployment settings where GPU memory and latency are real constraints.

Real-world applications:

• Interactive 3D editing and modeling tools, where users need near-instant geodesic brushes, surface distances, and region selections without a GPU cluster.
• Augmented and virtual reality, where markerless shape correspondence and distance queries must run at interactive framerates on headset hardware.
• Robotics and motion planning, where surface distance queries on scanned objects inform grasping, manipulation, and collision reasoning in memory-constrained onboard compute.
• Medical and anatomical shape analysis, where matching deformed or incomplete organ scans benefits from a method that tolerates broken meshes, noise, and sparse point clouds.
• Animation and gaming pipelines, where shape matching between characters (retargeting, morphing) can be done at a fraction of the previous cost.

Industry relevance centers on the deployment gap: prior neural geodesic methods are impractical on edge devices, mobile hardware, or in multi-shape batch inference because of memory blowup. LiteGE's fixed descriptor size means memory scales with the number of query points, not with a large model, which is exactly the property needed for real-time interactive software and embedded 3D vision.

Future Directions

• Extending the descriptor beyond geodesics: The authors propose reusing geodesic-aware shape descriptors for segmentation, shape analysis, and retrieval, which would test whether the UDF-PCA representation transfers to other geometric tasks.
• Temporal consistency for dynamic shapes: Applying LiteGE to time-varying 4D shapes (animation, motion capture, dynamic scans) would require enforcing consistency across frames, which the current formulation does not address.
• Reducing label dependence: Self-supervised or few-shot variants could lower reliance on expensive ground-truth geodesic data, which currently must be precomputed with FastDGG.
• Pushing compactness further: The paper notes that smaller PCA dimensions would give further memory reductions but leaves systematic exploration of that trade-off to future work. Extending the approach to less constrained shape collections (outside a limited category set) is also an open question.

Target Audience

This paper is most useful for 3D vision and geometry processing researchers working on geodesic computation, shape correspondence, or efficient neural shape representations; graphics engineers building interactive tools that need fast surface distance queries; and practitioners deploying 3D vision models on memory-constrained hardware. Readers interested in the practical trade-off between hand-crafted descriptors and learned backbones, or in non-isometric shape matching that tolerates degraded inputs, will find the results directly applicable. The paper assumes comfort with geodesic distance concepts and PCA but does not require deep differential geometry or 3D deep learning expertise.

Authors’ abstract

Computing geodesic distances on 3D surfaces is fundamental to many tasks in 3D vision and geometry processing, with deep connections to tasks such as shape correspondence. Recent learning-based methods achieve strong performance but rely on large 3D backbones, leading to high memory usage and latency, which limit their use in interactive or resource-constrained settings. We introduce LiteGE, a lightweight approach that constructs compact, category-aware shape descriptors by applying Principal Component Analysis (PCA) to unsigned distance field (UDFs) samples at informative voxels. This descriptor is efficient to compute and removes the need for high-capacity networks. LiteGE remains robust on sparse point clouds, supporting inputs with as few as 300 points, where prior methods fail. Extensive experiments show that LiteGE reduces memory usage and inference time by up to 300$\times$ compared to existing neural approaches. In addition, by exploiting the intrinsic relationship between geodesic distance and shape correspondence, LiteGE enables fast and accurate shape matching. Our method achieves up to 1000$\times$ speedup over state-of-the-art mesh-based approaches while maintaining comparable accuracy on non-isometric shape pairs, including evaluations on point-cloud inputs.