ETH-CVG

/

lightglue_disk

like 13

Follow

ETH Zurich - Computer Vision and Geometry Lab 33

Hello everybody!

I was trying to use the "LightGlueForKeypointMatching" class using Disk as the keypoint extractor but
when loading as mentioned on the documentation:

'''
from transformers import LightGlueForKeypointMatching

model = LightGlueForKeypointMatching.from_pretrained("ETH-CVG/lightglue_disk", trust_remote_code=True)
'''

it fails with:

KeyError: 'disk'

I have the most recent transformers and kornia version in my env (kornia==0.8.1 and transformers==4.53.2) and my Python version is 3.9.13

I tryed to load it using the stevenbucaille Disk keypoint extractor:

"
model = LightGlueForKeypointMatching.from_pretrained("stevenbucaille/disk", trust_remote_code=True)
"

But results were horrible even with simple images, due to using a model of type "disk" to instantiate a model of type "lightglue" and some weights of LightGlueForKeypointMatching not being initiallized from the model checkpoint at stevenbucaille/disk. I also used the AutoModel loading withand without the trust_remote_code and the same KeyError: "disk" appeared:

'''
from transformers import AutoModel

model = AutoModel.from_pretrained("ETH-CVG/lightglue_disk", trust_remote_code=True)
'''

Any help will be more than welcomed, maybe I am doing something wrong or maybe this is an issue that needs to be opened.

Thanks in advance to everyone participating in this discussion :)