shjo-april / mars Goto Github PK

Hi! I wonder how many GPUs and what kind of GPUs are required in general weakly-supervised semantic segmentation task including your work? Plus as introduced in your paper, a single RTX A6000 with 48GB is enough, so how long does it take for a single group of experiment? Thanks!

hi ! thank u for your nice work.im the fans of sangsung.could u please share with me code to email :[email protected]
thanks a lot.

Dear authors,

Thank you for posting this repo. We are just wondering if it is possible for you to post the actual code (model, training testing) to the server? Since we did not see anything here.

Thank you,

I wonder why the teacher network can supplement FN but not supplement the biased objects as FP again. Since the teacher network has the ability to distinguish between the two, why the teacher network's prediction results are not directly used as the final prediction.

Firstly, thank you for publishing such great work.

I have a question about the difference between Figure 2 (b) and (c) in the paper.

If I understand correctly, the results is separated by two feature cluster centroids made by K-means (as your setting, Kfg = 2).
For (c), result is only segmented in the region {(y, x)|Y^b_i(y, x) = c} which is come from the WSSS pseudo mask.
Since in WSSS case we could use the same pseudo mask.why (b) is not segmented in the same region but the whole image?

and you mention Nc is the number of images for the mini-batch, and I didn't see the certain value for your setting in the Section 4.

Thank you for your time and impressive work.