Point Cloud Normal Estimation via Representation Learning on Height Maps

Point cloud is a natural representation of 3D data. A fundamental research problem in point cloud analysis involves estimation of point normals. Both traditional optimization methods and more recent deep learning approaches show sensitivity to noise and difficulty in inferring normals at points with sharp geometric features. To address this, we propose a novel normal estimation method that first distinguishes between sharp and non-sharp feature points. This classification forms the basis for two designed regression networks, creating a dual-branch normal estimation architecture. The two networks independently infer normals at sharp and non-sharp feature points. Moreover, the feature encoders of the regression networks are pretrained using a contrastive learning mechanism. This mechanism exploits noise-corrupted augmentations of 2D input height maps derived from a specific underlying surface. Thereby, latent representations of height maps from the same clean surface remain similar despite different noise levels, yielding better regression results. Experiments demonstrate that our method achieves state-of-the-art results.