semantic-segmentation-pytorch
semantic-segmentation-pytorch copied to clipboard
On using non-dilated Resnet weights for dilated Resnets
For the dilated ResNets (resnet50-dilated, resnet101-dilated, etc), the same pretrained weights as the non-dilated ResNets are used. Even though only the last two layers are different (so all earlier layers should be okay in terms of weights), wouldn't we expect significantly degraded performance if we naively use regular ResNet weights, but with an operation which removes striding and adds dilation?
Am I missing something?