Arbicon-Net Arbitrary Continuous Geometric Transformation Networks for Image Registration

estimate geometric transforms between iamge pairs

• method
  • shared CNN -> feature map for each input
  • 4D coorelation tensor: using cosine similarity
  • then conv(symmetrically) -> aggregate neighbor information, normalize, MLP -> transformation descriptor
  • duplicate, coordconv -> F, done
• loss
  • strong: L2 loss
  • weak: defined on feature map