gaviotas/amn

Official PyTorch implementation of "Threshold Matters in WSSS: Manipulating the Activation for the Robust and Accurate Segmentation Model Against Thresholds", CVPR2022

amn's Issues

the step of make_sem_seg_pass is very time-consuming?

When I ran the code, I noticed that the step of make_sem_seg_pass is very time-consuming, especially for COCO with 80K images. Is it normal?

Is there any way to accelerate the pseudo-label generation?

The performance for 'Refined seed for AMN (CAM + CRF)'

Hey, thank you for your impressive work.

I try to download your 'Refined seed for AMN (CAM + CRF)', it is the pseudo mask for train_aug dataset.
I don't see if you published the evaluation function for train_aug, so I used the 'eval_sem_seg.py', it could see test the performance for 'train' and 'val' set.
But the miou of 'train' set is only among 37.78%.

I also used 'evaluation.py' in https://github.com/YudeWang/SEAM to get the miou of 'train_aug', the result is among 33.23%.

I am not sure if I made any mistakes during the evaluation, just let me know if I did and thank you again for your work.

Results on COCO

Thanks for your wonderful work in WSSS. Could you release the script on COCO and upload the corresponding checkpoint?

question

Hello, the idea of your article is very inspiring to me, but when running the code, it may be due to the different environment and computer configuration, resulting in the final result is not very ideal, do you have any suggestions for this problem?

the process

First of all, thank you for your code. I ran the code according to the steps in the file generate_pseudo_mask.sh. Is it the complete experiment process to run all the comments in the file generate_pseudo_mask.sh?

the segmentation

Hello, I trained the segmentation network according to the link to deeplabv2 you gave me, using The train_aug.txt (10582) pseudo mask trained the segmentation network, and then verified with val.txt (1449), the result was only 0.03. Among the 21 classes, only the background is recognized, and the results of other classes are 0. Do I need to use a pseudo mask when verifying?

Results on VOC12

Hi, thanks for the very organized repository. I have been trying to replicate the results achieved on the training set, however, the max I achieved with a single run of the entire pipeline is an mIOU of 70.7. Could you suggest reasons as to why that might be happening?

Thanks

train model

hello ,thanks for your code,Will you make the training code public?
I want to train it

the label

label smoothing

hi，thanks for your code.I would like to ask where the code of label smoothing as a training technique to subside the
noise in initial seed is? I hope you can tell me， thanks！