A subspace approach to layer extraction, patch-based SFM, and video compression

摘要: Abstract: "Representing videos with layers has important applications such as video compression, motion analysis, 3D modeling and rendering. This thesis proposes a subspace approach to extracting from by taking advantages of the fact that homographies induced planar patches in scene form low dimensional linear subspace. In subspace, input images are mapped onto well-defined clusters, can be reliably identified standard clustering algorithm (e.g., mean-shift). Global optimality is achieved since both spatial temporal redundancy simultaneously taken into account, noise effectively reduced enforcing constraint. The existence also enables outlier detection, making computation robust. Based on constraint, we propose patch-based scheme for affine structure (SFM), which recovers plane equation each patch scene, well camera epipolar geometry. We two approaches SFM: (1) factorization approach; (2) layer based approach. Patch-based SFM provides compact representation used construct high quality texture map layer. plan apply our generating Video Object Planes (VOPs) defined MPEG-4 standard. VOP generation critical but unspecified step Our model consists global localized deformations, closed-form solution. goals are: combining different level cues VOPs; VOPs undergo more complicated (non-planar or non-rigid)."