search

nv-tlabs / GSCNN

Gated-Shape CNN for Semantic Segmentation (ICCV 2019)
https://nv-tlabs.github.io/GSCNN/
Other
916 stars 200 forks source link

Question about paper. #1

Closed lc82111 closed 5 years ago

lc82111 commented 5 years ago

Thanks for your share.

Several questions:

  1. Does Eq.(6) miss a minus sign?
  2. How to obtain the gradient of an image for the input of the shape stream? soble?
  3. The regular steam outputs multi-scale feat-maps. How to obtain the attention maps at a resolution of HxW always?
tovacinni commented 5 years ago

Thanks for your interest in our work!

  1. Yes, I believe Eq.(6) is missing a minus sign. We will fix this, thanks for pointing it out!

  2. We use canny edges.

  3. We use bilinear upsampling.

2/3. We plan on releasing the code soon, so hopefully these questions would be answered more accurately once that happens.

arc144 commented 5 years ago

Thanks for your paper and reply. May I also ask some questions?

  1. How do you use the ASPP module to fuse the regular stream with shape stream? Do you apply ASPP only to regular stream, upsample it and then concatenate with shape stream followed by 1x1 conv?

  2. In the shape stream figure there is a conv1x1 leading to another conv1x1 (from gradients) does it mean that you concat the results of both convolutions?

  3. Finally, if I'm working in a binary context (foreground/background) may I simply replace the CE-losses for BCE and replace the argmax/gumbel softmax with a simple treshold?

Thanks a lot. Will the code be released this month?

Nutakki2259 commented 5 years ago

Hi, When can i get the code for this project?

tovacinni commented 5 years ago

Thanks for your interest in our work!

@arc144

1/2. We have released the code now so I think the code would better answer that question.

  1. I believe so! We have tried our network with some private datasets in a binary context and results do seem to improve. Thresholding is one strategy, but you could also consider having 2 classes (foreground / background).

@Nutakki2259

The code has been released.