Point cloud occlusion recovery with shallow feedforward neural networks

3D scenes reconstructed from point clouds, acquired by either laser scanning or photogrammetry, are subject to data voids generated by occluding objects. Modeling from incomplete data is usually a manual process in which human interpretation plays an essential role. This paper presents a machine learning algorithm based on neural networks capable of recovering point cloud occlusions for surfaces that can be approximated with injective functions. Starting from the point clouds acquired around the occlusion, a set of single-layer feedforward networks with a variable number of neurons is trained and validated with a subset of the original cloud, which is preliminarily decimated using local curvature to reduce CPU cost. The averaged result of the best neural networks is evaluated on a spatial domain that contains the 2D projection of the void, obtaining a complete 3D point cloud for the occluded volume. Criteria for choosing the number of neurons and the activation function for hidden and output layers are illustrated and discussed. Results are presented for both simulated and real occlusions, describing the pros and cons of the proposed method.