WaveNormal: Wavelet-based Geometry-Aware Disentanglement for Point Cloud Normal Estimation

Normal estimation from point clouds is fundamental to 3D geometric processing and applications. Existing methods typically infer normals from local patches and therefore focus on designing feature extraction architectures and selecting appropriate neighborhood scales to better capture geometric context. Although integrating graph convolutions and multi-scale feature fusion improves performance, accurate normal estimation remains challenging in regions with complex geometry. This limitation arises from the indiscriminate coupling of high- and low-frequency information within feature representations, which obscures fine-grained geometric details and consequently leads to suboptimal normal estimation. To address these issues, we propose WaveNormal, a wavelet-based geometry-aware disentanglement framework that models the underlying geometry of local patches through a multi-scale geometric perception (MGP) module and a feature aggregation-disentanglement (FAD) module. Specifically, the MGP module extracts discriminative features at multiple scales and guides cross-scale feature fusion, thereby capturing both large-scale structural information and fine-grained geometric details. The FAD module employs a learnable wavelet lifting scheme to decompose coupled representations into high- and low-frequency components. The high-frequency components are selectively modulated to emphasize informative geometric details and subsequently fused with the low-frequency components through an adaptive fusion layer. In addition, a feature compensation layer is incorporated to mitigate potential information loss caused by successive wavelet decompositions, thereby preserving the underlying geometry of local patches. Extensive evaluations on both synthetic and real-world datasets demonstrate that WaveNormal improves normal estimation performance, particularly around sharp features and intricate surface details.