How to apply the algorithm of loftr to a complete system？

[FujiwaraZayako]

I'm sorry to bother you with this question, and I sincerely ask for your answer. I compared superglue with loftr. This indirect method of superglue can directly extract points from a single picture, which is close to the Orb method of most slams, and the feature points detected in these independent frames will be used in other steps. However, loftr skip this step is the end-to-end direct method. Although the results can calculate the motion attitude, it is difficult to integrate the existing slam and navigation system without a related method for single-frame feature points. I have two questions. How to make loftr generate motion trajectories? How can I apply the algorithm of loftr to a complete SLAM or visual navigation? If you are willing to answer, please send a private letter to me: [email protected] 。 Thank you very much.

[zehongs]

Hi, please refer to our PAMI version: Semi-Dense Feature Matching with Transformers and its Applications in Multiple-View Geometry. https://ieeexplore.ieee.org/abstract/document/9956767

[FujiwaraZayako]

Hi, please refer to our PAMI version: Semi-Dense Feature Matching with Transformers and its Applications in Multiple-View Geometry. https://ieeexplore.ieee.org/abstract/document/9956767

Is the relevant part of this article open source in your code base? I searched and couldn't find it.

[zehongs]

Not yet.