CovaEdge-Net: A Covariance-Driven Adaptive Graph Network for Precise Industrial Point Cloud Edge Detection

Abstract

Reliable edge detection in industrial point clouds is difficult when Euclidean neighborhoods cross surface discontinuities and local descriptors vary with object pose. CovaEdge-Net addresses these effects through geometry-conditioned neighborhood construction, rotation-aware pairwise encoding, and temperature-controlled aggregation. Adaptive Covariance Neighbor Selection (ACoNS) ranks candidate neighbors with a covariance-weighted dissimilarity that combines spatial, normal, and curvature terms. The Tangent Orientation Descriptor (TOD) separates scalar rotation-invariant relations from learned feature differences, while Adaptive Temperature-controlled Aggregation (ATA) uses a positive, lower-bounded temperature to control neighborhood selectivity. On the ABC benchmark, CovaEdge-Net reports an F1 score of 0.973 and is evaluated under density, rotation, noise, ablation, and one model-level efficiency setting. The evaluation covers one dataset and reports aggregate results across three runs; cross-dataset generalization, formal convergence, and adversarial robustness remain outside the evaluated scope. The ablation results are used to interpret the roles of ACoNS, TOD, and ATA without extending the conclusions to unmeasured intermediate quantities.