Mikhail Scherbina

How do I reduce the number of training rays?

https://github.com/thohemp/6DRepNet360/issues/2#issuecomment-1798445673

Same

https://github.com/RuojinCai/doppelgangers/pull/8

When I try to train, I sometimes get warnings like: ``` Corrupt JPEG data: 126 extraneous bytes before marker 0xd9 ```

I got an AP of 84.3 when training without flip aug on superpoint + lightglue. Initialized from you checkpoint, trained for 20 epochs. I made it so that Learning rate...

Sure thing, here are the keypoints produced with superpoint. Below is your imlementation of read loftr_matches. I added some debug code for visualization at the bottom of the fuction. ```python...

So, obviously, it has smaller number of keypoints than disk / loftr

My model also does not filter the cup matches, they are all 1: 

Disk + lightglue + your loftr weights also produces the same but better result: