search

magicleap / SuperGluePretrainedNetwork

SuperGlue: Learning Feature Matching with Graph Neural Networks (CVPR 2020, Oral)
Other
3.31k stars 672 forks source link

For megadepth training, how to deal with the points without gt depth? #72

Closed archershot closed 3 years ago

archershot commented 3 years ago

When i reproduce train on megadepth, i find that there are somes parts of superpoints without gt depth. So i cannot check whether it has the correspondece point in another image. How should i deal with these points? ignore or treat them as no match?

archershot commented 3 years ago

And should these points add to superglue for training?

sarlinpe commented 3 years ago

These point should still be given as input to SuperGlue, but they should be ignored when computing the loss.

archershot commented 3 years ago

These point should still be given as input to SuperGlue, but they should be ignored when computing the loss.

thank you very much!

sarlinpe commented 3 years ago

Please close this issue if your problem is solved.