First Order Motion Model for Image Animation
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
- reconstruction loss based on perceptual loss