Contact: Prof. Dr. Daniel Cremers

For a human, it is usually an easy task to get an idea of the 3D structure shown in an image. Due to the loss of one dimension in the projection process, the estimation of the true 3D geometry is difficult and a so called ill-posed problem, because usually infinitely many different 3D surfaces may produce the same set of images.

The goal of multiview 3D reconstruction is to infer geometrical structure of a scene captured by a collection of images.

Considering a dynamic scene that changes over time, 3D reconstruction can be applied to every time step independently. However, one can achieve temporally more consistent results by using the information from several time frames together, thus computing a spatio-temporal hyper-surface in 4D space. See this link for details.

The estimation of 3D geometry from a single image is a special case of image-based 3D reconstruction from several images, but is considerably more difficult since depth cannot be estimated from pixel correspondences. Thus, further prior knowledge or user input is needed in order to recover or infer any depth information.