Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Results on HPatches between pre-trained model and this Github #99

Open
miumiu94 opened this issue Jan 16, 2023 · 0 comments
Open

Results on HPatches between pre-trained model and this Github #99

miumiu94 opened this issue Jan 16, 2023 · 0 comments

Comments

@miumiu94
Copy link

miumiu94 commented Jan 16, 2023

Hi, thanks for your great work!

I have a question. You showed results on HPatches that compare the evaluation from different models in the table. The first row shows the evaluation from pre-trained model which is from SuperPointPretrainedNetwork, and the third row is from this GitHub.

I notice that repeatability(in Detector metric) from this GitHub is better than that from pre-trained model, but matching score(in Descriptor metric) is lower distinctly. The result above mean that model from this GitHub can detect feature keypoint better, but match worse because descriptors have a poor descriptive ability. I observe that you use different batch size, learning rate, lamda-loss(desc loss/detect loss) from pre-trained model in the training details. I tried to adjust same parameters mentioned above with pre-trained model, but match performed even worse.

So, I wonder why the result mentioned above occurred, and how to make the training reach agreement with the pre-trained model.

Looking forward to receiving your reply!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant