This repository holds a Keras porting of the ICLR 2016 paper by Yu and Koltun. It holds the four semantic segmentation pretrained networks that you can find in the original repo (Caffe).

Just use the DilationNet function in dilation_net.py to get the model. To see an example, run predict.py.

Please note that the porting works on with the Theano dim ordering. Tensorflow backend should since if needed, the function convert_all_kernels_in_model is called. However, it is not tested.

Cityscapes model disclaimer: I didn't manage to convert the final upsampling layer (deconv with grouping), so I replaced it with Upsampling + Convolution.