search

YangtaoWANG95 / TokenCut

(CVPR 2022) Pytorch implementation of "Self-supervised transformers for unsupervised object discovery using normalized cut"
MIT License
304 stars 34 forks source link

wants to improve the failed images comeout from this code #13

Closed srilekhapanda closed 2 years ago

srilekhapanda commented 2 years ago

I have run this code for unsupervised saliency detection for ECSSD dataset where found some of the images are not generating the proper mask. Varying the tau and taking eigen vector without taking the absolute value getting some improved results.

What could be the generalised solution for this type of images?

XiSHEN0220 commented 2 years ago

Hi, @srilekhapanda

Thanks for your interest in the project !

I think the adaptive tau might be something good to try. Maybe one naive solution is to build a list of tau as candidates and search for the one that produces the most separable eigenvector.

To be honest, it is not evident to have a simple/general cutting strategy. If you find something that is elegant and works better, do not hesitate to make a pull request.

Best,

srilekhapanda commented 2 years ago

Thanks for your response XiSHEN. I have one more question that for some of the images without taking the absolute value of the eigen vector getting improved results. Is it the right approach to do?

XiSHEN0220 commented 2 years ago

We only leverage abs eigenvector to determine the seed, which for me, is reasonable to do. Please refer to this line

To get the bounding boxes, we still use the raw eigenvector. Please check this part

srilekhapanda commented 2 years ago

Okay Xishen thank you again for clarification. Have a good day.