This repository is the implementation of "Narrow while deep network for real-time semantic segmenation"
We have tested our code on Ubuntu 16.04, Pytorch 1.0
Requirements:
Pytorch 1.0
You need first to download the Cityscapes dataset by yourself, since the Cityscape uses 19 class label for semantic segmentation task, you also need to convert the original 33 class label image to 19 class lbel image using the code provided by the Cityscapes team.
If you used Cityscape before, you can ignore this step.
All the parameters are manually set in the Seg_NDnet/train.py.
The parameters you need to change may be the path of the Cityscape dataset.
Or you can just put the Cityscape in the "opt" memu in Ubuntu
You can train the model after cd to the menu Seg_NDnet with:
python train.py
If there are any other error messages about "python package can not find", use "pip intall xx" to setup
After the training is completed, set the "RESTORE_FROM" in "Seg_NDnet/infer_citys.py" with the path at which your model saved, and then run
python infer_citys.py
this command will make prediction over the val set one by one and save the prediction to the path you specified
We use the code provided by the Cityscape team to compute miou of test;
The code is located at 'Cityscape-master/cityscapescripts/evaluation/evalPixelLevelSemanticLabeling.py'
python evalPixelLevelSemanticLabeling.py
There are two parameters you need to specify
line 52 PRED_PATH='the path of saved prediction (gray label)images'
line 667 groundTruthImgList = glob.glob('/opt/Cityscapes/gtFine_trainvaltest/gtFine/val//_gtFine_labelIds.png'(the path of gt images))