A PyTorch implementation of semantic segmentation according to Multi-Scale Context Aggregation by Dilated Convolutions by Yu and Koltun. Pretrained weights are obtained from dialation-Tensorflow and thanks to Andrea Palazzi for his work from which this work has been inspired.
How To
- Download the pretrained weights from cityscapes
- Keep that in the data folder
- Run CAN_network.py for the demo.
Notes I have kept the output image size 1024x1024x3 due to memory constraint. If you want the size to be 1024x2048x3, have enough memory and change the sizes appropriately in the code.