Blue noise sampling of surfaces from stereoscopic images

We propose an original sampling technique for surfaces generated by stereoscopic acquisition systems. The idea is to make the sampling of these surfaces directly on the pair of stereoscopic images, instead of doing it on the meshes created by triangulation of the point clouds given by the acquisition system. Point clouds are generally dense, and consequently the resulting meshes are oversampled (this is why a re-sampling of the meshes is often done). Moving the sampling stage in the 2D image domain greatly simplifies the classical sampling pipeline, allows to control the number of points from the beginning of the sampling/reconstruction process, and optimizes the size of the generated data. More precisely, we developed a feature-preserving Poisson-disk sampling technique applied to the 2D image domain - which can be seen as a parameterization domain - with inter-sample distances still computed in the 3D space, to reduce the distortion due the embedding in R^3. Experimental results show that our method generates 3D sampling patterns with nice blue noise properties in R^3 (comparable to direct 3D sampling methods), while keeping the geometrical features of the scanned surface. © 2016, Society for Imaging Science and Technology (IS&T).