Graph Constrained Data Representation Learning for Human Motion Segmentation

Recently, transfer subspace learning based approaches have shown to be a\nvalid alternative to unsupervised subspace clustering and temporal data\nclustering for human motion segmentation (HMS). These approaches leverage prior\nknowledge from a source domain to improve clustering performance on a target\ndomain, and currently they represent the state of the art in HMS. Bucking this\ntrend, in this paper, we propose a novel unsupervised model that learns a\nrepresentation of the data and digs clustering information from the data\nitself. Our model is reminiscent of temporal subspace clustering, but presents\ntwo critical differences. First, we learn an auxiliary data matrix that can\ndeviate from the initial data, hence confer more degrees of freedom to the\ncoding matrix. Second, we introduce a regularization term for this auxiliary\ndata matrix that preserves the local geometrical structure present in the\nhigh-dimensional space. The proposed model is efficiently optimized by using an\noriginal Alternating Direction Method of Multipliers (ADMM) formulation\nallowing to learn jointly the auxiliary data representation, a nonnegative\ndictionary and a coding matrix. Experimental results on four benchmark datasets\nfor HMS demonstrate that our approach achieves significantly better clustering\nperformance then state-of-the-art methods, including both unsupervised and more\nrecent semi-supervised transfer learning approaches.\n