[Registration] SuperGlue, Predator, REGTR

Reference

DReg-NeRF: Deep Registration for Neural Radiance Fields

Superglue: Learning feature matching with graph neural networks

Predator: Registration of 3d point clouds with low overlap

REGTR: end-to-end point cloud correspondences with transformers

Matching 2D image correspondences: SuperGlue, Predator, REGTR

Inspired by SuperGlue [32], which is a deep learning method for matching 2D image correspondences, Predator [16] and REGTR [42] adopted the self-attention and cross-attention mechanisms from SuperGlue to learn the correlation for pairwise low overlapping point clouds.

The ground-truth overlapping scores are computed from dense point clouds and used to mask out the correspondences outside the overlapping regions.

SuperGlue, Predator, REGTR 모두 self-attention, cross-attention mechanism 사용

• In the self-attention layers, $Q = K = V$ represents the same feature tensor in each block.
  • The self-attention mechanism enables the network to learn the relationship inside the same feature points
• In the cross-attention layers, the keys, and the values are the feature tensors from the other block.
  • the cross-attention mechanism enables the communication of the different feature points