We address the problem of large-scale 3D reconstruction from calibrated images relying on a viewpoint-based approach. The representation is in the form of a collection of depth maps, which are fused to blend consistent depth estimates and minimize violations of visibility constraints. We adopt a least commitment strategy by allowing multiple candidate depth values per pixel in the fusion process and deferring hard decisions as much as possible. To address the inevitable noise in the depth maps, we explicitly model its sources, namely mismatches and inaccurate 3D coordinate estimation via triangulation, by measuring two types of uncertainty and using the uncertainty estimates to guide the fusion process. To the best of our knowledge, this is the first attempt to model both geometric and correspondence uncertainty in the context of dense 3D reconstruction. We show quantitative results on datasets with ground truth that are competitive with the state of the art.