Generate custom dataset

[nfyfamr]

Dear author, thanks for sharing the great work! I am creating a private dataset to test matching on the DKM. I tried to follow the instructions in https://github.com/Parskatt/DKM/blob/main/docs/training.md. But the d2-net generates scene info file as npz, while it seems yours use npy file. On top of the file extension, the npz file doesn't contain 'pairs' key, which is used in the training. I am wondering how you generated your scene info file. Could you share your processing code? Thanks!

[ZhouShunlong]

I also encountered this problem. Are there any good solutions?

[Parskatt]

Hi I thought I put my modifications here but maybe not. The only difference is that I save all pairs with overlap > 0 instead of an overlap matrix. This is due to the matrix being extremely sparse for lsrge scenes, so basically we save space. Regarding npy vs npy its just using np save instead of savez iirc.