Skip to content

Latest commit

 

History

History
17 lines (15 loc) · 1.08 KB

cv_20.md

File metadata and controls

17 lines (15 loc) · 1.08 KB

Eulerian video magnification for revealing subtle changes in the world

Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, William Freeman (2012)

Key points

  • Reveal temporal variations (both periodic and non-periodic) in videos that are impossible to see with the naked eye
    • Lower spatial frequency can be exaggerated more without introducing artifacts
  • Spatial decomposition --> temporal filtering --> amplification for certain temporal frequency band of interest (e.g. heart rate)
    • Both color variation and small motion
  • Temporal filtering needs sufficient spatial pooling to rise above noise
  • Brightness constancy assumption used
  • Lagrangian: trajectory of particles tracked over time
    • Computationally expensive due to accurate motion estimation and hard to make artifact-free
    • Better for enhancing fine point feature motion and large amplitudes
  • Eulerian: properties of a voxel evolve over time
    • Motion is not explicitly estimated, but exaggerated by amplifying temporal color changes at fixed locations
    • Better for smoother structures and smaller amplitudes