Non-rigid structure from motion

People: Rohith MV, Xiaolong Wang and Chandra Kambhamettu

Estimation of non-rigid structure from motion (NRSFM) has often been performed as a linear combination of basis shapes. However, when dealing with scenes containing human articulated motion (especially in presence of clothing), the number of basis shapes precludes accurate results. We model deformation as a combination of articulated and non-rigid surface deformation. We propose a novel algorithm for segmenting motion to find articulated components and a hierarchical NRSFM algorithm which estimates articulated and non-rigid structure separately. Our method attempts to remove the articulated motion from the observation matrix leaving behind only the non-rigid component which can be represented with fewer bases. Results show that our method successfully segments motion in a variety of cases and structure reconstruction is improved using the hierarchical framework.

Illustration of one reconstructed face based on 2D points

Related publications