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

Could you offer the pretrained model on bsds500? #100

Open
stupidman98 opened this issue Dec 13, 2021 · 3 comments
Open

Could you offer the pretrained model on bsds500? #100

stupidman98 opened this issue Dec 13, 2021 · 3 comments

Comments

@stupidman98
Copy link

I noticed that you said the edge on bsds500-trained model is thick.I want to see.thank you

@xavysp
Copy link
Owner

xavysp commented Dec 15, 2021

Hi, to see this difference, quantitatively, its hard to say. we need a new way, or metric for evaluation. For a perceptual judgement, you can find our banner. When the models are trained with BSDS the predictions are thick, if the model is trained with BIPED, as you can see, the predicted edges are thinner.

let me know what you think.
Xavier

@stupidman98
Copy link
Author

In my opnion,weighted celoss may lead net to predict more edge around gt for gain less loss,because the cost from false positive is cheap.Focal loss maybe a good choice,which is briefer than CATS,and it effects in my experiments.

@xavysp
Copy link
Owner

xavysp commented Dec 16, 2021

I agree with you. However, when I started in Edge detection, I started tackling the most important problem, which is Dataset for edge detection. Once this problem is solved or at least near to be solved. We can start on optimisation, loss function, and/or deep learning (DL) architecture. At the end, the objective is to make a DL models for edge detection as simpler as the ones for image classification.

Nice to chat with you and keep working!

Xavier

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

2 participants