Pretrained model with ResNet50 backbone.

[skyshoumeng]

Thank you for the great work, could you provide the pretrained model with ResNet50 backbone. Thank you very much.

[danielS91]

The networks with ResNet50 backbone only serve as a reference for our approach. As shown in our experiments, the performance gain ist quite small while increasing inference time notably.

However, you can easily train it on your own by passing --encoder resnet50 to our train.py. Note that the used ResNet50 is from TorchVision and, thus, uses downsampling in the spatial convolution instead of the first point-wise convolution in the bottleneck block (see: our code, this is also known as ResNet v1.5, see GitHub TorchVision).

[skyshoumeng]

OK, I got it, thank you very much.