We propose an algorithm for clustering large sets of images of a scene into smaller subsets covering different parts of the scene suitable for 3D reconstruction. Unlike the canonical view selection of simon et al., we do not focus only on the visibility information, but introduce an alternative similarity measure which takes into account the relative camera orientations and their distance from the scene. This allows us to formalize the clustering problem as a graph partitioning and solve it using spectral clustering. The obtained image clusters bring down the amount of data that has to be considered by the reconstruction algorithms simultaneously, thereby allowing traditional algorithms to take advantage of large multi-view data sets processing them significantly faster and at smaller memory costs compared to using the full image datasets. We tested our approach on a number of multi-view data sets and demonstrated that the clustering we obtain is suitable for 3D reconstruction and coincides with what a human observer would consider as a good clustering.     «

We propose an algorithm for clustering large sets of images of a scene into smaller subsets covering different parts of the scene suitable for 3D reconstruction. Unlike the canonical view selection of simon et al., we do not focus only on the visibility information, but introduce an alternative similarity measure which takes into account the relative camera orientations and their distance from the scene. This allows us to formalize the clustering problem as a graph partitioning and solve it usin...     »