https://zju3dv.github.io/efficientloftr/

1. aggregated self- and cross-attention: reduce token size before each attention for efficiency
2. fusion network: fusing transformed coarse features with backbone features
3. A two-stage refinement is followed to obtain sub-pixel correspondence
4. Mutual-Nearest Neighbor

Method

Local feature extraction

• adopt RepVGG [11] as our feature backbone

Efficient Local Feature Transformation

an aggregated attention mechanism with adaptive token selection for efficiency.

Why Aggragated Attention Module?

1. VanillaAttention: applying it directly to dense local features is impractical due to the significant token size.
2. LinearAttention: reduced representational power

depth-wise convolution

…

Coarse-level Matching Module