Stable View Synthesis

less than 1 minute read

Published:

geometric scaffold from SfM and MVS

  • on-surface feature aggregation: directional feature vector processed to produce a new vector
  • SfM and MVS: colmap
  • for each point, query the set of images where the point is visible
  • encoder: imagenet pretrained resnet18, batchonrm frozen
  • fc graph of features and a graph attention network
  • a sqeuence of unet for rendering