Abstract
Skeletonization is a fundamental task in 3D vision, yet it lacks quantitative evaluation metrics. This paper introduces a framework to assess skeleton quality by analyzing the stability of the contraction process. By monitoring the statistical distributions of geometric curvature and surface normal changes during Laplacian-based contraction, the authors propose the "Stable Convergence Hypothesis." Results show that high-quality skeletons exhibit symmetric, unimodal distributions, providing a mathematical signature for evaluating 3D skeletonization algorithms.
Introduction
In 3D vision, a "skeleton" is a thin, centered representation of a shape—essentially a "stick figure" version of a complex point cloud. These skeletons are vital for robotics and computer graphics, enabling tasks like grasp planning and navigation. However, the field faces a major hurdle: there is no universally accepted "ground truth" for what an ideal skeleton looks like. Most researchers rely on subjective visual intuition, making it difficult to compare algorithms or guarantee topological accuracy. This work seeks to establish mathematical benchmarks based on the geometric evolution of the point cloud during the contraction process.
Methods
The study focuses on Laplacian-based contraction (LBC), a popular method that iteratively "shrinks" a point cloud inward until it forms a medial surface or curve. To evaluate this process, the researchers track two key geometric properties across iterations:
- Normal Vector Changes: Measured using cosine similarity and normalized angular distance to see how surface orientation shifts.
- Curvature Differences: Estimated via the eigenvalues of local covariance matrices to track changes in surface "pointiness" or flatness.
The Stable Convergence Hypothesis
The core of the methodology is the Stable Convergence Hypothesis. The authors propose that a high-quality skeletonization process results in a distribution of geometric changes that is symmetric and unimodal (resembling a bell curve). Specifically, they look for normal differences to converge toward an average of 0.5 and curvature differences to converge toward 0.
Results and Discussion
Testing on the OmniObject3D dataset revealed a clear correlation between "clean" bell-curve distributions and skeletons that visually preserved the object's topology. Objects with clear structures, such as the hammer, skateboard, carrot, and horse, exhibited the highest quality and the most stable convergence patterns.
Conversely, "failed" or difficult cases like the cabinet or watermelon resulted in highly skewed or multi-modal histograms. This indicates that the algorithm struggled to find a stable medial axis, resulting in skeletons that lacked significant topological information. The metrics also proved sensitive to local errors; for instance, in the dumbbell model, an unexpected peak in the histogram correctly flagged an incomplete contraction on one side of the object.
Conclusion
This research provides a novel quantitative approach to skeleton evaluation by focusing on the stability of the contraction process rather than just the final visual output. While the stable convergence pattern is currently an empirical observation, it offers a promising path toward standardized benchmarks in 3D vision. Future work will focus on formalizing these definitions and extending the analysis to more complex, non-manifold geometries.
Summary updated