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

Unable to reproduce the result as paper reported #8

Open
greatwallet opened this issue Mar 27, 2022 · 1 comment
Open

Unable to reproduce the result as paper reported #8

greatwallet opened this issue Mar 27, 2022 · 1 comment

Comments

@greatwallet
Copy link

greatwallet commented Mar 27, 2022

Hello, thank you for your inspiring work! I tried to re-run the code on all of the datasets, but the results were not as promising as those repored by paper.

For CUB and DeepFashion, I did not modify any of the codes, but the metrics on CUB were poor.
image

And for Pascal-Part, I modified the training hyper-params according to your supp file. And also for the fairness of evaluation, I trained a foreground segmentator (a DeepLabV2-ResNet50-2branch) and used the predicted mask at evaluation. I conducted training upon Car, Cat and Horse. The results on Horse were OK, but those on Cat and Car were really not good.

image

Could you please provide analysis upon why the re-run results on CUB were not good? Also, for Pascal-Part, are there any training details that is left out in papers, so that I did not reproduce the results? (BTW, could you provide supervised mask of PascalPart?)

@FredericOdermatt
Copy link

Hey,

I just retrained the model on the CUB dataset with a batch size of 3 instead of 6 and I got the following results FYI

  FG-NMI1 FG-ARI1 Full-NMI Full-ARI
paper 46.0 21.0 43.5 19.6
loading provided weights 46.03 21.03 43.52 19.58
own training run (batch size 3 instead of 6) 44.10 19.90 41.28 18.70
this github issue 39.62 17.72 39.12 18.11

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