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

The code on Ade20k #15

Open
Yinyf0804 opened this issue Feb 19, 2022 · 4 comments
Open

The code on Ade20k #15

Yinyf0804 opened this issue Feb 19, 2022 · 4 comments

Comments

@Yinyf0804
Copy link

Thanks for your great work "Open-World Entity Segmentation". Could you provide the evaluation code on ADE-20K dataset, when using models trained on COCO?

@qqlu
Copy link
Owner

qqlu commented Feb 21, 2022

We use the class-agnostic COCO API to evaluate both COCO and ADE20K. Thus, you could only transform the ADE20K annotation format to COCO annotation format.

@Yinyf0804
Copy link
Author

Thanks for your advice. I use the model trained on COCO (r50 1x 29.8) and evaluate it on ADE20K val, get the results 20.09 box AP and 13.78 seg AP. So i wonder if the AP box are used as AP_e in Table4(a) ?

@qqlu
Copy link
Owner

qqlu commented Feb 22, 2022

We still use AP_e to evaluate ADE20K. At first, we use 3x models in Table4(a). Second, we clear some annotations in ADE20K to keep consistency with COCO. That's because the definitions of COCO and ADE20K for things and stuff are different. This part of the code should be released after the acceptance of the paper.

It is also a weakness of our paper that the annotation should be kreept consistent among different datasets. We are solving this problem in another way.

@Yinyf0804
Copy link
Author

Thanks for your reply. But the caption of Table4 mentions that "(3) Ours with 12 epochs in (a)", and Sec 6.2 mentions that "PanopticFCN and DETR are trained for 12 and 300 epochs". I wonder if you actually use 1x models in Table4(a)?

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