number of maxpooling and upsampling in VGGSegnet and VGGUnet is not same about image-segmentation-keras

[divamgupta]

Having more upsampling layers would increase the number of parameters which are not initialized by any pretrained model. That would make it much harder to train.

[gangooteli]

If we use vggnet segnet pre-trained weights, it will be applied to VGG16 layers i.e. Encoder part(or Downsamling part). The upsampling part is learned from scratch. In that in implementation of VGGSegnet, we can add one more upsampling layer to get the image dim == output dim.
And we need note to use resize.
What is your thoughts on that ?

[divamgupta]

A better idea would be to use less encoder layers from VGGNet. You could also increase the upsampling layers with skip connections for easy training.