A major challenge in the unsupervised modeling of articulated objects observed from multiple-view videos is the capture of motion. This problem implies establishing correspondences between the objects across frames. We provide three approaches to solve the problem based on computer vision techniques and spectral graph theory. The first relies on modeling the scene as a sparse collection of 3-D points (surfels). We propose two multi-view extensions of the Lucas-Kanade algorithm to track the features in 3-D and efficiently recover the scene-flow. The second is based on spectral graph theory and searches to establish dense correspondences between pairs of articulated shapes represented by graphs. We revisit classical methods and propose an alternative solution for matching large and sparse graphs. Finally, we consider the consistent segmentation of the object in time based on the extension of spectral clustering methods to sequences.