Iteratively weighted principal component analysis and orientation consistency for normal estimation in point cloud

Estimating accurate surface normals is a prerequisite for point cloud surface reconstruction, rendering, and geometric analysis. The standard PCA approach fits a plane to each point’s k-nearest neighbours and uses the smallest eigenvector as the normal estimate. This works well on clean, uniformly sampled point clouds but degrades in two common practical situations: in the presence of noise and outliers, which corrupt the covariance matrix, and near sharp features, where neighbours from geometrically different surface patches mix and produce averaged, incorrect normals. This paper addresses both problems. The iteratively re-weighted PCA (IWPCA) assigns lower weights to neighbours that deviate from the current normal estimate and iterates until convergence, effectively ignoring outliers and cross-feature contamination without requiring explicit outlier detection. An orientation consistency propagation step then resolves the normal sign ambiguity inherent to PCA-based estimation: starting from seed points with reliably oriented normals, consistent orientation is propagated across the cloud via a minimum spanning tree, without user interaction. Published in the International Journal of Wireless and Mobile Computing (2020), IWPCA yields more accurate angular errors than standard PCA and achieves globally consistent normal orientation on datasets with sharp features and moderate noise.

Algorithm principle

The method estimates normals by iteratively re-weighting each point’s local neighborhood so that outliers and cross-feature neighbours contribute less to the covariance estimate. A follow-up orientation propagation step resolves the normal sign ambiguity across the point cloud.

Visual material

Additional method and result figures will be added when the original assets are available.

Results and impact

The reported evaluation compares normal-estimation accuracy and orientation consistency against standard PCA-style baselines under noise, outliers, and sharp-feature contamination.