Manifold-Regularized Dual Dictionary Learning

Manifold-Regularized Dual Dictionary Learning for Human Pose Estimation

Human pose estimation is commonly defined as estimating the configuration of the human body from an image or video clip. The input image is typically captured using a camera and the pose is modeled with angles that describe the configuration of each major body joint.

In this project, we wish to present a framework for dual dictionary learning that takes advantage of two assumptions, collectively known as the manifold assumption. First, that high-dimensional input data often lie on a lower-dimensional manifold and second, that the output changes smoothly over the input manifold. To this end, we attempt to approximate input data on affine subspaces spanned by a few dictionary atoms while minimizing a measure of smoothness for the output data. Optimization programs are derived for both supervised and semi-supervised settings. The framework is applied to human pose estimation from monocular single depth images.

Related Publication:

  • A. Soltani-Farani, H. R. Rabiee, S. A. Hosseini, P. Mianjy “Manifold-Regularized Dual Dictionary Learning for Human Pose Estimation”, Manuscript in preparation, Dec. 2014

Project code:

  • The project code will be uploaded soon.

People involved:

  • Ali Soltani-Farani, Abbas Hosseini, Poorya Mianjy,  Hamid R. Rabiee

Sparse Signal Processing Group