Will you release the results on CityScapes?

[PkuRainBow]

Great work!

I am wondering the reproduced performance on the CityScapes. It would be great if you could share the related results.

[Tete-Xiao]

We're afraid that we're short-handed so we can't do that right now.

It shouldn't be hard. You may take a try.

[PkuRainBow]

@Tete-Xiao Thanks for your quick reply. I want to try the ResNet101, but it seems that the url you provide is valid .....

http://sceneparsing.csail.mit.edu/model/pretrained_resnet/resnet101-imagenet.pth

As the official use 7x7 conv instead of the setting like below:

        self.conv1 = conv3x3(3, 64, stride=2)
        self.bn1 = SynchronizedBatchNorm2d(64)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(64, 64)
        self.bn2 = SynchronizedBatchNorm2d(64)
        self.relu2 = nn.ReLU(inplace=True)
        self.conv3 = conv3x3(64, 128)
        self.bn3 = SynchronizedBatchNorm2d(128)
        self.relu3 = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)

[PkuRainBow]

@Tete-Xiao Thanks for your repo. I could reproduce the PSP to achieve 78.6 mIOU on the CityScapes with the hyperparamters you provide on my own code base.

[tonysy]

@PkuRainBow Hi, could you share your training configuration? I try to train PSP on CIityScapes, but get 68 mIOU. I use crop instead of resizing the image and the crop size is 512*512. Thanks a lot!

[Tete-Xiao]

@PkuRainBow Good to hear that! We're glad to know our toolbox is useful.

[PkuRainBow]

@tonysy My crop size is 769X769

[boundles]

@PkuRainBow Hi, could you share your training configuration, such as learning rate, data augmentation?

[PkuRainBow]

@boundles We are preparing for a new work, which is simple and performs better than PSP. We will release our paper and code in the future months. Please pay attention to my git repo.