STAR-Edge: Structure-aware Local Spherical Curve Representation for Thin-walled Edge Extraction from Unstructured Point Clouds

TL;DR

STAR-Edge addresses thin-walled edge extraction from unstructured point clouds by introducing a local spherical curve representation that builds structure-aware neighborhoods, then encodes these curves with a rotation-invariant descriptor via spherical harmonics and classifies edge points with a lightweight MLP. It also uses the representation to estimate more accurate normals and to refine edge points by projecting them onto true edge locations. The approach yields superior performance on thin-walled datasets and ABC CAD data, with strong robustness to noise and sparse sampling, and provides thorough ablations and runtime analysis. The work advances edge extraction in challenging geometries and offers a practical, geometry-aware solution for downstream tasks in manufacturing and 3D modeling.

Abstract

Extracting geometric edges from unstructured point clouds remains a significant challenge, particularly in thin-walled structures that are commonly found in everyday objects. Traditional geometric methods and recent learning-based approaches frequently struggle with these structures, as both rely heavily on sufficient contextual information from local point neighborhoods. However, 3D measurement data of thin-walled structures often lack the accurate, dense, and regular neighborhood sampling required for reliable edge extraction, resulting in degraded performance. In this work, we introduce STAR-Edge, a novel approach designed for detecting and refining edge points in thin-walled structures. Our method leverages a unique representation-the local spherical curve-to create structure-aware neighborhoods that emphasize co-planar points while reducing interference from close-by, non-co-planar surfaces. This representation is transformed into a rotation-invariant descriptor, which, combined with a lightweight multi-layer perceptron, enables robust edge point classification even in the presence of noise and sparse or irregular sampling. Besides, we also use the local spherical curve representation to estimate more precise normals and introduce an optimization function to project initially identified edge points exactly on the true edges. Experiments conducted on the ABC dataset and thin-walled structure-specific datasets demonstrate that STAR-Edge outperforms existing edge detection methods, showcasing better robustness under various challenging conditions.

Figures (13)

• Figure 1: The characteristics of our method and visual comparison of edge extraction results in thin-walled structures. (a) defines the concept of a thin-walled edge. (b) highlights that the primary challenge in thin-walled edge extraction is the sensitivity to local neighborhood selection. (c) illustrates how a larger neighborhood may include points from both upper and lower surfaces as well as side-end faces, blurring boundary information. (d) shows that a smaller neighborhood may suffer from noise and insufficient sampling on side faces, lacking the contextual information needed for accurate edge point recognition. (e) depicts the spherical projection of the local neighborhood of four points, where points of the same color lie on the same underlying surface. This illustrates our key observation that points co-planar with the neighborhood center tend to align along a great circle arc after spherical projection. (f) provides a qualitative comparison of our STAR-Edge with RFEPS xu2022rfeps, EC-Net yu2018ec, and MFLE chen2021multiscale.
• Figure 2: Overview of the proposed method. STAR-edge comprises three main steps: (a) constructing a local spherical curve; (b) identifying edge points by integrating rotation-invariant local descriptors with a light-weight MLP layer; and (c) optimizing edge points by adjusting their positions based on the estimated normal vector from the local spherical curve.
• Figure 3: Neighborhood distribution of five surface points on the unit sphere. Local spherical points are shown in grey, sampled key points on the fitted spherical Curve in blue, and local spherical curves in red. We observe that the neighborhoods of edge points resemble two semicircles (e.g. ④), while those of non-edge points maintain a complete great circle (e.g. ①).
• Figure 4: Normal estimation results for points ③, ④, and ⑤ in \ref{['fig:LSC']} (a). In each sub-figure, the red curve is the great circle arc extracted by RANSAC, the blue curve is the non-great circle segment, and the yellow arrow is the estimated normal vector.
• Figure 5: Visual comparison of state-of-the-art methods on the thin-walled structure dataset. In these close-up views, our method demonstrates superior accuracy in edge point extraction. Edge points are in red color.