Can we train XFeat on a custom dataset?

[blackmamba-ops]

No description provided.

[guipotje]

Hi @blackmamba-ops, yes, it is possible to train XFeat on a custom dataset. You just need to adapt the dataloader for your dataset.

[Lannist]

Hi @blackmamba-ops, yes, it is possible to train XFeat on a custom dataset. You just need to adapt the dataloader for your dataset.

How should I create this dataset, any Tutorial？

[guipotje]

All you need is to construct a data loader that delivers a correspondence map in the form xy_source (N,2) and xy_target (N,2), where each i-th line in the first set is the source pixel coordinates that corresponds to the i-th target pixel. This can be obtained via homography for synthetic training, or depth maps / dense optical flow. Here is the exact line where the code expects this correspondence map.