The paper achieves 82% mIoU on the Cityscapes test part, while as far as I know, several papers in ICCV 2019 get a better mIoU, it shall be better to revise the diction about the proposed SOTA conclusion.

hi, can we use this way in the strip pooling that

your way----
(1)strip pooling-conv2D(3,1)-expand-
(2)strip pooling-conv2D(1,3)-expand-
=fusion((1),(2))--conv2D(1,1)

can we use this ,
(1)strip pooling-expand-
(2)strip pooling-expand-
=fusion((1),(2))--conv2D(3,3)
don't use the conv2D(3,1) and conv2D(1,3) in the middle, instead , use conv2D(3,3) after they fusion??, i am confused with it ,can you explain it ? thanks.

Hi,
Thank you for open-sourcing the amazing work.
In the SPHead module, the pool_size parameter
er is set to 20, 12..
Can I ask why these two values? If the input image is 480x480 as in ade20k, the feature map produced by the ResNet backbone is 15x15 (32 down) which is smaller than 20x20. It actually becomes an upsampling operation rather than pooling.

Hello, in resnet.py, if I only want to add SPM module in resnet50, I just need to add it in the exact location. Do I need to remove the average pooling and FC layer behind layer3 and layer4? If I don't want to remove the FC layer, can I?

HI, i am trying to design a simple module of strip pooling for my network. Would you please give me suggestions is it ok or i have to make changes in some places thanks in advance.

class StripPooling(nn.Module):
def init(self, pool_sizes):
super(StripPooling, self).init()
self.pool_sizes = pool_sizes

def forward(self, x):
    h, w = x.shape[2:]
    k_sizes = []
    strides = []
    for pool_size in self.pool_sizes:
        k_sizes.append((int(h / pool_size), int(w / pool_size)))
        strides.append((int(h / pool_size), int(w / pool_size)))

    sp_sum = x

    for i in range(len(self.pool_sizes)):
        out = F.avg_pool2d(x, k_sizes[i], stride=strides[i], padding=0)
        out = F.upsample(out, size=(h, w), mode="bilinear")
        sp_sum = sp_sum + out

    return sp_sum

Thank for your contribution！And I have some questions I want to ask you。
Why is there only 19 categories in cityscapes_colors.txt and no background categories?
Why is it not necessary to calculate the loss function of the background class in calculating the loss function?
In the code, you set the value of the background to 255 and set ignore_label = 255. I think you ignore the calculation of the loss function of the background, and the output channel of the network is also 19 types, and there is no background prediction. Is this prediction chart accurate? It is also possible that I have overlooked the details of some code processing, and please provide guidance accordingly.

What is the visualization method in the article？

Hi, there,

I tried to adapt Strip Pooling as 3D and insert to my model, but after about 20-50 iterations in a epoch, I got nan loss.
How you saw this situation?

What's the risk when run StripPooling? Is it because overflow?

But it's quite confused that the loss didn't go increase or decrease, but suddenly "nan".

Thanks a lot!

def init __ (self, nDenselayer, channels, growth, norm_layer = nn.BatchNorm2d, ** kwargs):
         super (RDB1, self) . init __ ()
         modules = []
         channels_buffer = channels
         for i in range (nDenselayer):
             modules.append (make_dense1 (channels_buffer, growth, 3, norm_layer, self._up_kwargs))

Hi，thanks for your code ,your job is amazing.
I meeting a problem when I'm training the model. The value of “MainLoss ” is shaking just in the beginning，can you give me some advice？
I only modify the value of "batch_size" to 2, because I only have one gpu.
Thanks.

Hi, thank you for your interesting work.
But I'm confused why the strip pooling module(SPM) structure is different from the paper. In the paper, the end of SPM is a sigmoid function and the output of SPM elementwise multiply x, instead of elementwise summation shown in the source code (SPNet/models/customize.py).

Are these pre trained SPnets or generic pre-trained resnet-50 and resnet-101 models?

Thanks 🍡

Hi Andrew

When I try to download the resnet50 and resnet101 pretrained models, download failed.
the web page shows as below

It is a very nice job. However, I can't find the schematic illustration strip pooling module implementation in this codebase, which is depicted in Figure.2 in the paper. Would you please provide me an explanation ?

Thank you for your contribution. When running test.py and train.py, the following error occurred. I want to ask why it is and how to modify it.The error is as follows:
AssertionError: Non-existent key: ade20k
I hope I can get your guidance, and I will thank you very much！
My email is [email protected]

Thanks for the valuable contribution!
I am trying to reproduce the results for cityscapes, I used the configuration file cityscapes_spnet101.yaml
The validation mIoU can only reach to ~73.x %.
my experiment runs on 4 GPUs, sync_bn is enabled.

The only thing I changed is the batchsize, i am using 12 sample in a batch.

Could you give me some hints?

excuse me,I tried to download many times but failed.Is there any other way to get it？