how to generate noisy saliency maps during training?

[JiangAndy]

I have read your paper and your code, I have some questing to ask your help ?

1. I'm confusing about the noisy saliency maps, I want to know how to generate noisy saliency maps during train.
2. in your code, there is a sal_label in your dataloder.py, so whta's the sal_label?
3. the method described in the paper is unsupervised, but in the train precedure, labels are used, so I'm really confusing. can you explain these above questions? thank you!

[kris-singh]

Sorry for the late reply. This is a re-implementation of the paper.

If you read the paper carefully, you would find that paper uses labels from unsupervised methods as pseudo labels for the training procedure. You need to first use unsupervised models to get noisy saliency maps, which you then use to train the Network. Hope that answers your questions.

[GerardWalsh]

Which implementations did you use to generate the noisy labels? It seems quite hard to find implementations of the 4 methods the authors chose to generate groundtruth.

[kris-singh]

I implemented 2 methods of the 4 proposed methods. The paper is agnostic of any unsupervised saliency detection methods, so you are free to use any. If I get the time I will try to upload the implementations as well.

[GerardWalsh]

Thanks, that would be great to have your implementations. @JiangAndy I used OpenCV's saliency detection module.