First Order Motion Model for Image Animation

less than 1 minute read

Published:

decouple apperance and motion information by a self-supervised formulation

  • a set of learned keypoints and local affine transformations
    • thus called “first order” method
  • an occlusion-aware generator: infer background pixels from the context which is not included in source image
  • trained by reconstruct a training video
    • predict a dense motion field
    • approx. transformations from sets of sparse trajectories (didnot get it…)
  • losses
    • reconstruction loss based on perceptual loss
      • and on diff. number of resolutions