First Order Motion Model for Image Animation

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