F2S3: Robustified determination of 3D displacement vector fields using deep learning

Areal deformation monitoring based on point clouds can be a very valuable alternative to the established point-based monitoring techniques, especially for deformation monitoring of natural scenes. However, established deformation analysis approaches for point clouds do not necessarily expose the true 3D changes, because the correspondence between points is typically established naively. Recently, approaches to establish the correspondences in the feature space by using local feature descriptors that analyze the geometric peculiarities in the neighborhood of the interest points were proposed. However, the resulting correspondences are noisy and contain a large number of outliers. This impairs the direct applicability of these approaches for deformation monitoring. In this work, we propose Feature to Feature Supervoxel-based Spatial Smoothing (F2S3), a new deformation analysis method for point cloud data. In F2S3 we extend the recently proposed feature-based algorithms with a neural network based outlier detection, capable of classifying the putative pointwise correspondences into inliers and outliers based on the local context extracted from the supervoxels. We demonstrate the proposed method on two data sets, including a real case data set of a landslide located in the Swiss Alps. We show that while the traditional approaches, in this case, greatly underestimate the magnitude of the displacements, our method can correctly estimate the true 3D displacement vectors.