Fabio Cuzzolin, Diana Mateus, David Knossow, and Radu Horaud To submit to the International Journal of Computer Vision, 2011 Abstract The contribution of the paper is twofold: an analysis of a number of desirable geometric features of locally linear embedding (in the context of clustering) due to the related minimization problem is developed. As LLE conserves the affine coordinates of the points inside local neighborhoods, shape protrusions as high-curvature regions of the surface are also preserved. The covariance constraint acts instead like a force that stretches those protrusions, making them wider separated and lower dimensional. Based on these features a novel scheme for unsupervised bodypart segmentation along time sequences is proposed in which 3D shapes are clustered in the embedding space, clusters are propagated along time, and merge or split in an unsupervised fashion to accommodate changes of the body topology. Experiments on both synthetic and real sequences of dense voxelset data are shown which support the ability of the method to cluster bodyparts consistently in time, in a totally unsupervised fashion. Performance is measured against ground-truth labels obtained for real sequences from the available estimated kinematic model after model-based motion capture. sp  Download PDF  Zipped Postscript BibTeX entry @article{cuzzolin11ijcv,    AUTHOR = "Fabio Cuzzolin and Diana Mateus and David Knossow and Radu Horaud ",    TITLE = "Robust coherent Laplacian protrusion segmentation along 3D sequences",   JOURNAL = "to submit to the International Journal of Computer Vision",    YEAR = "2011"  }