Dense Correspondence Finding for Parametrization-Free Animation Reconstruction from Video

Abstract

We present a dense 3D correspondence finding method\nthat enables spatio-temporally coherent reconstruction of\nsurface animations from multi-view video data. Given as input\na sequence of shape-from-silhouette volumes of a moving\nsubject that were reconstructed for each time frame individually,\nour method establishes dense surface correspondences\nbetween subsequent shapes independently of surface\ndiscretization. This is achieved in two steps: first, we obtain\nsparse correspondences from robust optical features\nbetween adjacent frames. Second, we generate dense correspondences\nwhich serve as map between respective surfaces.\nBy applying this procedure subsequently to all pairs\nof time steps we can trivially align one shape with all others.\nThus, the original input can be reconstructed as a sequence\nof meshes with constant connectivity and small tangential\ndistortion. We exemplify the performance and accuracy of\nour method using several synthetic and captured real-world\nsequences.