Introduction
Skeletonization is a fundamental technique in shape analysis, reducing complex 3D point clouds to 1D curve skeletons or 2D medial surfaces. In robotics, these representations are critical for path planning, object manipulation, and environment mapping. Despite its utility, the evaluation of skeletonization algorithms has historically relied on subjective visual inspection. This paper introduces a formal numerical framework to quantitatively assess skeleton quality through four geometric metrics: topological similarity, boundedness, centeredness, and smoothness.
Methodology
The proposed framework evaluates skeletons based on their fidelity to the original point cloud across several dimensions.
Topological Similarity
To ensure the skeleton preserves the "connectedness" of the original shape, the authors employ persistent homology. By constructing a simplicial complex and tracking features like loops and voids across scales, they generate persistence barcodes. The similarity between the original shape and the skeleton is then quantified using bottleneck ($d_B$) and Wasserstein ($d_{W_p}$) distances. This metric is crucial for identifying when a skeleton fails to represent essential features like holes or loops.
Boundedness and Centeredness
Boundedness measures whether the skeleton remains contained within the object's volume. This is calculated by projecting direction vectors from skeletal points onto a unit sphere; a point is considered bounded if its projections cover the sphere. Centeredness evaluates how well the skeleton aligns with the geometric center of the shape, using an ellipse-fitting approach on local cross-sections to measure the offset. These metrics are vital for tasks like robotic grasping, where the center of mass or volume is a key target.
Smoothness
For robotic navigation, a skeleton must be smooth to avoid erratic motion commands. Smoothness is measured by analyzing the variation in tangent directions along the skeletal branches using cosine similarity. This ensures the resulting path is feasible for a physical robot to follow.
Experimental Evaluation and Results
The framework was tested on a diverse array of models, including inanimate objects, organic shapes, and complex environments like synthetic trees and cave scans. To assess robustness, the authors introduced Gaussian noise and varied point density.
A key finding is that Laplacian-based skeletonization is highly sensitive to point sparsity; reducing density from 4096 to 1024 points significantly increased topological distance scores. Computational analysis shows that topological similarity and boundedness are the most resource-intensive metrics, together accounting for 99% of the processing time. On average, evaluating a 4096-point cloud takes approximately 34 seconds on a standard CPU.
Conclusion
This research provides the first objective, open-source tool for evaluating point cloud skeletonization. By moving beyond visual judgment, the framework enables rigorous comparison of algorithms and ensures skeletal models are accurate enough for safety-critical robotic tasks. Future work will focus on optimizing computation times for real-time applications and refining metrics for non-convex local geometries.
Summary updated