Abstract : In 3D video, recorded object behaviors can be observed from any viewpoint, because the 3D video registers the object's 3D shape and color. However, the real-world views are limited to the views from a number of cameras, so only a coarse model of the object can be recovered in real-time. It becomes then necessary to judiciously texture the object with images recovered from the cameras. One of the problems in multi-texturing is to decide what portion of the 3D model is visible from what camera. We propose a texture-mapping algorithm that tries to bypass the problem of exactly deciding if a point is visible or not from a certain camera. Given more than two color values for each pixel, a statistical test allows to exclude outlying color data before blending.
