What does the scene look like from a scene point?

In this paper we examine the problem of synthesizing virtual views from scene points within the scene, i.e., from scene points which are imaged by the real cameras. On one hand this provides a simple way of defining the position of the virtual camera in an uncalibrated setting. On the other hand, it implies extreme changes in viewpoint between the virtual and real cameras. Such extreme changes in viewpoint are not typical of most New-View-Synthesis (NVS) problems. In our algorithm the virtual view is obtained by aligning and comparing all the projections of each line-of-sight emerging from the “virtual camera” center in the input views. In contrast to most previous NVS algorithms, our approach does not require prior correspondence estimation nor any explicit 3D reconstruction. It can handle any number of input images while simultaneously using the information from all of them. However, very few images are usually enough to provide reasonable synthesis quality. We show results on real images as well as synthetic images with ground-truth.