https://zju3dv.github.io/loftr/ 视频讲解很好

Detector-Free, semi-dense

Method

LoFTR Module

• detail

Coarse level matches

differentiable matching layer: a dual-softmax operator [34, 47]

1. The score matrix S between the transformed features is first calculated by

   

2. apply softmax on both dimensions of S to obtain the probability of soft mutual nearest neighbor matching. Then the matching probability Pc, confidence matrix, is obtained by:

   

3. Match Selection

   

   1. mutual nearest neighbor (MNN), 文中没解释，按常规理解。

Coarse-to-Fine Module

zoom to 1/2 resolution, apply coarse level matches on each patch pairs defined by i^, j^

compare feature of i^ with features of neighbors of j^ by correlation.

Supervison

The final loss consists of the losses for the coarse-level and the fine-level: L = Lc +Lf

Discussion