Sharpness fields in point clouds using deep learning

Prashant Raina, Sudhir P. Mudur, Tiberiu Popa

Abstract: Highlights • Our sharpness field localizes sharp features in point clouds. • It is computed with deep learning and an appropriate neighborhood representation. • It can be used for improving feature-aware smoothing. • It allows sampling points lying exactly on edges. • It allows segmenting point clouds into patches. Abstract Sharp features in 3D objects are often lost when the shapes are digitally acquired using a 3D scanning device. The resulting noisy point clouds do not usually contain points lying on sharp features, making it difficult to reconstruct these features or to reverse-engineer the original CAD model of the object. We presented a method in [1] to calculate a scalar field, defined on the underlying Moving Least-Squares surface of the point cloud, whose local maxima correspond to sharp edges. As this sharpness field is defined over a continuous domain, it allows us to determine the locations of sharp features on the object even at points not originally sampled from the scanned surface. The computation of the sharpness field is made possible by training deep neural networks to compute the value of the field for any given local neighborhood of a surface. We also presented in [1] a feature-aware smoothing algorithm which uses the computed sharpness field to smooth noisy point clouds while preserving sharp edges and yielding points that lie exactly on these edges. In addition, we described two approaches to segmentation of a point cloud using the sharpness field. In this paper, we extend our previous work in [1] with different neural network architectures and local neighborhood representations that improve sharpness field computation, resulting in significantly improved feature-aware smoothing. We also numerically evaluate the accuracy of different deep learning models for sharpness field computation.