The author of this projcet doesn't release the config file and usages. I was working on semantic segmentation based on MXNet. So I have fix the errors and add usage for the original project.

• Python2.7
• OpenCV
• CUDA 8 or 9

• Clone the mxnet source code
• Put ordering_op-inl.h into incubator-mxnet/src/operator/tensor
• Put softmax** into incubator-mxnet/src/operator/contrib
• Follow the official instructions to build and install
• sh init.sh to build some libs for dataloader and detection task

• Put data into data/cistycapes, you can use soft link to add the dataset ln -s <dataset> ./data/cityscapes
• Use the model provided by autho to load params(Optional)
  • Model
  • It's not the pretrained model, so I just use it for test.
  • I will try to generate a pretrained model recently

Because the original paper is not public now, I can only use some magic number of option to run this code. I am working on understanding the model from the code.

python2 experiment/deeplab/drn_train.py --cfg experiment/deeplab/cfgs/resnet_v2_38_deeplab_dcn_gru_v7.yaml

TBD

main result on ade20k testing is 0.5635(symbol-v11)-single model main result on cityscapes testing is 82.4(symbol-v7) and 82.8(symbol-v13) - single model If you have question or some advice, email me 'zhuangyq@pku.edu.cn' The model release on 'https://pan.baidu.com/s/14_zNi_m7hjv-sMWjY0D1Hw'

Thanks the author anyway!