Driven by several key advances such as novel image descriptors and efficient
reconstruction techniques, 3D vision made significant progress in the last
decade. Modern techniques are able to reconstruct detailed scene models from
very general image data. In this talk, I will first give an overview of the
robust 3D reconstruction pipeline developed in my lab. Since real-world input
images typically depict objects at various scales, geometry reconstructed using
multi-view stereo techniques is inherently multi-scale. I.e., individual depth
samples represent a specific, finite surface area. I will therefore next discuss
how to incorporate this property into surface merging techniques and describe
our scale-aware floating scale surface reconstruction technique (FSSR) in
detail. Photorealistic rendering of reconstructed geometry requires also high
resolution information about the surface color and texture. Unfortunately,
existing work is often restricted to small and/or laboratory-type datasets. I
will therefore finally describe our recently proposed texturing method that
allows us to texture real-world models with millions of triangles using hundrets
of photographs.