Point Cloud Denoising using Deep Learning

The emergence of uncontrollable noise from diverse sources possess several difficulties in scanning 3D objects. In the case of animals in the wild this is especially hard to manage since their movements are unavoidable during the acquisition process. This causes distortions that compromise the reconstruction process significantly, rendering the whole acquisition procedure useless, or in the best case requiring strenuous assisted editing tasks in order to obtain viable results. In this work we propose a method for detecting and filtering noisy zones in meshes generated through point clouds acquired from in-situ scanning southern elephant seals in their natural habitat. We trained a CNN model with meshes resulting from noisy and clean acquisitions. The trained neural network is able to filter, in subsequent acquisitions, those parts in the mesh that do not belong to the original objects. This greatly reduces or eliminates the amount of manual editing work that is required in order to obtain a useful acquisition.