Giters

qubvel / segmentation_models.pytorch

Segmentation models with pretrained backbones. PyTorch.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

PSPNet implementation

thawro opened this issue · comments

commented

I found a few inconsistencies with the original paper

  1. There is no dilation strategy for encoders for PSPNet in smp, while authors of the paper used dilated strategy for the encoder (i.e. ResNet):

"...we use a pretrained ResNet [13] model with the dilated network strategy [3, 40] to extract the feature map."

  1. In the smp there is a dropout layer before segmentation head, while authors of the paper do not mention dropout at all (maybe because batch normalization is used)
  2. The convolutional layer before segmentation head uses kernel_size=1 and the convolutional layer inside segmentation head uses kernel_size=3. Shouldnt that be kernel_size=1 for segmentation head and kernel_size=3 for conv before head?

Are those modifications intentional? If so, what was the reason?

commented

This issue is stale because it has been open 60 days with no activity. Remove stale label or comment or this will be closed in 7 days.

commented

This issue was closed because it has been stalled for 7 days with no activity.