Lagrangian motion magnification with double sparse optical flow decomposition

Microexpressions are fast and spatially small facial expressions that difficult to detect. Therefore, motion magnification techniques, which aim at amplifying hence revealing subtle in videos, appear useful for handling such expressions. There basically two main approaches, namely, via Eulerian or Lagrangian techniques. While the first one magnifies implicitly by operating directly on image pixels, approach uses optical flow (OF) techniques extract magnify pixel trajectories. In this study, we propose a novel local of micro-motions. Our contribution is 3-fold: first, fine tune recurrent all-pairs field transforms (RAFT) OFs deep learning faces adding ground truth obtained from variational dense inverse search (DIS) OF algorithm applied CASME II video set micro This enables us produce videos an efficient sufficiently accurate way. Second, since micro-motions both space time, approximate sparse components time leading double decomposition. Third, use decomposition specific areas face, where introduce new forward warping strategy using triangular splitting grid barycentric interpolation RGB vectors corners transformed triangles. We demonstrate feasibility our various examples.