search

yywbkn / MMViT-Segmentation

6 stars 2 forks source link

Provide weights #1

Open dscarmo opened 1 year ago

dscarmo commented 1 year ago

Hello, can you provide the trained weights used in your paper for us to reproduce your prediction capabilities? Thank you!

wubukeneng commented 1 year ago

The weights of the two-category semantic segmentation experiment can be downloaded by clicking the following link. https://cowtransfer.com/s/3cc67ec9274040

dscarmo commented 1 year ago

Thank you, are the weights really that small in size?

dscarmo commented 1 year ago

@wubukeneng I also have another question, how can I run the multiclass segmentation without needing labels or the original dataset? I want to run in unseen data that doesn't have labels. The test script requires pseudo labels to be present. This is the same case also for the InfNet code. Could you help me with that?

yywbkn commented 1 year ago

Hello, running multiclass segmentation without labels or the original dataset has always been a challenge. There are two methods that you can try to address this issue: unsupervised learning and semi-supervised learning.

If you have a small portion of labeled data, you can try using semi-supervised learning. The first step is to perform supervised training using the labeled data. Then, in the second step, you can train the unlabeled data using methods such as generating pseudo labels or applying consistency regularization. Since this problem has always been a challenge, the final effectiveness can only be determined based on the results of the specific experiments conducted.