GeoRefine-Net: A Geometry-Consistent Boundary Refinement Network for Industrial-Style Point Cloud Edge Detection

Industrial-style 3D point clouds frequently exhibit intricate geometric structures such as abrupt curvature changes, multi-surface junctions, and fine surface seams. Extracting fine boundary cues from these data remains challenging due to three intertwined factors: cross-boundary neighborhood contamination caused by purely spatial neighbor selection, orientation-sensitive boundary contrast in arbitrary poses, and over-smoothing from uniform feature aggregation. To address these issues, we present GeoRefine-Net, a geometry-consistent boundary refinement framework that progressively refines boundary predictions through three dedicated mechanisms. The geometry-consistent graph construction (GCG) module incorporates multi-order geometric statistics and a learnable boundary prior to actively discourage erroneous cross-surface neighbor associations during graph construction. The Rotation-aware Boundary Contrast Encoding (RBC) module captures local directional disparities using a composite descriptor incorporating normal cross-product interaction, angular divergence, curvature discrepancy, and tangent-plane projection, thereby providing rotation-consistency-enhanced boundary cues across varying object orientations. The Confidence-Guided Boundary Refinement (CBR) module introduces a gating mechanism that adaptively amplifies or suppresses neighbor contributions based on learned consistency scores, producing sharper boundary activation while suppressing noisy responses. The main novelty of GeoRefine-Net lies in the coupled design of geometry-consistent graph construction, rotation-consistency-enhanced boundary contrast encoding, and confidence-guided aggregation, rather than in any single handcrafted descriptor or isolated network component. Additional evaluations using SymBLE, PR-AUC, directed-versus-symmetric GCG, normal and curvature perturbations, normal-free variants, coefficient sensitivity analysis, and inference-time profiling further characterize the robustness and limitations of the proposed framework. Comprehensive evaluations on the ABC CAD dataset demonstrate that GeoRefine-Net outperforms the strongest directly compared baseline by 2.8 percentage points in the F1 score.