Dense Relation Network: Learning Consistent and Context-Aware Representation For Semantic Image Segmentation. Modification of DRN source code
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.
ordering_op-inl.h
into incubator-mxnet/src/operator/tensor
softmax**
into incubator-mxnet/src/operator/contrib
sh init.sh
to build some libs for dataloader and detection taskdata/cistycapes
, you can use soft link to add the datasetln -s <dataset> ./data/cityscapes
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 '[email protected]' The model release on 'https://pan.baidu.com/s/14_zNi_m7hjv-sMWjY0D1Hw'
Thanks the author anyway!