Integration of motion fields through shape

TitleIntegration of motion fields through shape
Publication TypeConference Papers
Year of Publication2005
AuthorsJi H, Fermüller C
Conference NameIEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2005. CVPR 2005
Date Published2005/06/20/25
ISBN Number0-7695-2372-2
Keywords3D motion estimation, Automation, CAMERAS, computational geometry, Computer vision, constrained minimization problem, decoupling translation from rotation, Educational institutions, image colour analysis, image gradients, image resolution, Image segmentation, image sequence, Image sequences, integration of motion fields, Layout, minimisation, Motion estimation, motion field integration, motion segmentation, parameter estimation, planar patches, rank-3 constraint, scene patches, SHAPE, shape and rotation, shape estimation, structure estimation

Structure from motion from single flow fields has been studied intensively, but the integration of information from multiple flow fields has not received much attention. Here we address this problem by enforcing constraints on the shape (surface normals) of the scene in view, as opposed to constraints on the structure (depth). The advantage of integrating shape is two-fold. First, we do not need to estimate feature correspondences over multiple frames, but we only need to match patches. Second, the shape vectors in the different views are related only by rotation. This constraint on shape can be combined easily with motion estimation, thus formulating motion and structure estimation from multiple views as a practical constrained minimization problem using a rank-3 constraint. Based on this constraint, we develop a 3D motion technique, which locates through color and motion segmentation, planar patches in the scene, matches patches over multiple frames, and estimates the motion between multiple frames and the shape of the selected scene patches using the image gradients. Experiments evaluate the accuracy of the 3D motion estimation and demonstrate the motion and shape estimation of the technique by super-resolving an image sequence.