wanglixilinx / dsrl Goto Github PK

Are you going to open the code soon?

Thanks,

您好，拜读了文章，有一个问题还不敢确定。请问一下FA的两个输入是只有Segmentation分支和SR分支的瓶颈层特征是吗？

I try to realise the FA loss after your answers。But I met some questions in relation graph 。
my test code is
x = np.random.random((256, 64, 64))
y = np.random.random((256, 64, 64))
y = torch.from_numpy(x).to(device).float()
x = torch.from_numpy(x).to(device).float()
out_feature = torch.bmm(x.permute(1,2,0), x.permute(1,0,2))
out2_feature = torch.bmm(y.permute(1,2,0), y.permute(1,0,2))
print(out_feature.shape)

I want to follow your last answer to write the code, " S= torch.bmm(F.transpose, F), Here the shape for F is W' H' x C', the shape for F.transpose is C' x W'H', so the shape of similarity matrix S is W'H' x W'H'. "
so I write this : out_feature = torch.bmm(x.permute(1, 2, 0), x))
but it will raise error, so i try the method that will not raise error.but the end shape is 64,64,64. Please tell me what wrong I make. Thank you very much!

hi, thanks a lot for your paper, is the code will be open?

Does open source need to last half a year?

The code seems missing?

Hi there. I was wondering that SSSR path outputs features with 19 channels and SISR outputs features with only 3 channels(RGB i guess?), so how do you compute the FA loss between them?

(I was trying to realize your idea using Deeplab v3+ as backbone, and Deeplab v3+ actually only used a simple interpolation to upsample the last_conv features(with 19 channels, and in your papar you said that you added another layer of interpolation to make sure output is 2x bigger than input). Meanwhile i used another 3 groups of ConvTranposed and Conv to build the SISR path, so finally SISR will output an image with 3 channels and 2x bigger than input. Since i'm not sure how to compute FA loss with features that have different channels, I currently choose to use 19 channels SSSR last_conv features and another 19 channels features in the halfway of SISR's decoder to compute the FA loss, but the result is like a disaster.)

When the code can be released?Thanks.

When we will see the code?

最近在研究你们的paper，很棒的idea！ 请问下SSIR中的Decoder是怎么实现的？ 是简单的双线性上采样还是用什么方法实现的？

Will you open the code?

Hi! Recently I read your CVPR paper, it is very interesting idea. But I have a confusing part is that on the baseline. Why the baseline are so low ? How about the results on Stronger Baseline.

the size of output feature map is WHC, and the number of its similarity matrix should be C, so for the last loss L, do you sum C different L_fa loss?

I recently read your paperand I have a question that how is “ semantic segmentation similarity matrix or SISR similarity matrix” in FA part realized? Looking forward to your reply as soon as possible

The code har under legal sweep for 5 months, when the code can be released?Thanks.