Surface Reconstruction from Silhouette and Laser Scanners as a Positive-Unlabeled Learning Problem

Typical 3D reconstruction pipelines employ a combination of line-laser scanners and robotic actuators to produce a point cloud and then proceed with surface reconstruction. In this work we propose a new technique to learn an Implicit Neural Representation (INR) of a 3D shape S without directly observing points on its surface. We just assume being able to determine whether a 3D point is exterior to S (e.g. observing if the projection falls outside the silhouette or detecting on which side of the laser line the point is). In this setting, we cast the reconstruction process as a Positive-Unlabelled learning problem where sparse 3D points, sampled according to a distribution depending on the INR's local gradient, have to be classified as being interior or exterior to S. These points, are used to train the INR in an iterative way so that its zero-crossing converges to the boundary of the shape. Preliminary experiments performed on a synthetic dataset demonstrates the advantages of the approach.