Puy, GillesVandergheynst, Pierre2013-01-142013-01-142013-01-142013https://infoscience.epfl.ch/handle/20.500.14299/87819We extend the domain of application of a method developed recently for joint reconstruction of l images, representing the same scene, from few multi-view measurements. While this method was initially designed for planar scenes, we show here that parametric smooth transformations can be used to handle more challenging scene structures. Our algorithm estimates one reference image common to all viewpoints, l complementary images modeling details in the scene that are not always visible, and few transformation parameters modeling the inter-correlation between the observations. The algorithm is an alternating descent method built to minimize a non-convex objective function and which produces a sequence converging to one of the critical points of this function.Compressed sensingNon-convex optimizationMulti-view imagingtext::conference output::conference paper not in proceedings