Attention-Guided Multi-Scale Point Cloud Normal Estimation with Noise-Aware Training

This study presents an attention-guided multi-scale framework for surface normal estimation on point clouds, extending an attention-based dynamic graph baseline with four targeted modifications. Multi-scale neighborhood features are adaptively fused via attention, and normals are predicted directly rather than reconstructed through surface fitting. To reduce the impact of measurement noise, training targets for corrupted patches are replaced with the nearest noise-free ground-truth normals identified by a Chamfer-distance association. Two lightweight geometric regularizers — a coplanarity-guided weighting term and a Z-axis alignment constraint — yield a balanced training objective without added model complexity. On synthetic benchmarks, the method is competitive with the baseline, with the clearest gains on primitive geometries such as planes, cylinders, and spheres; multi-seed experiments confirm statistically comparable overall accuracy alongside category-level gains under noise-free and varying-density conditions. On large-scale indoor scans, the method remains stable and robust to sensor noise and sampling-density variation. However, the coplanarity supervision introduces an implicit planarity bias, yielding slightly higher error on complex real-world scenes than methods without such regularization. The framework keeps a moderate parameter count and runtime suitable for commodity hardware and requires no auxiliary inputs such as edge detectors or resampled patches. The code is available at https://github.com/inyoungoh-cde/DeepLearningNormalEstimation.