Question on training LightGlue.

[frankkim1108]

Thank you for making this code public.

During training LightGlue, it seems like the loss is calculated between pred and gt's log_assignment.
However, to the best of my knowledge, the log_assignment seems to be created from the descriptors from each images extracted from superpoint.
Also, the descriptor shape matches the keypoints number.
What if the model predict different amount of keypoints compared to the GT keypoints, resulting in different descriptor shape and also resulting in different log_assignment shapes.
If the log_assignment shape from pred and gt are different wouldn't there be an error?

Thank you for your time.