Reliable 3D reconstruction extending pixel-level certainty measures

A new method for obtaining a three-dimensional volumetric reconstruction from a set of views improving the classical Shape from Silhouette method (SFS) is presented. SFS approaches can be easily accelerated through hardware and software techniques but they are very sensible to errors arising during calibration and segmentation processes so they present difficulties when dealing with real images. This paper proposes a new algorithm which uses the information about pixel segmentation uncertainty contained in each view in order to get a reliable 3D reconstruction of the scene. Aggregation of the projected uncertainties permits to classify scene's voxels by means of a decision rule but also makes it possible to create a three-dimensional confidence map of the scene. As a consequence, the regions where more information is needed can be foreseen. Sample reconstructions from real image sets are presented and evaluated.