OVERVIEW

Source code for DCME for instance segmentation.

Pease check ISISA for an improved version.

TRAIN

• Download and extract cityscapes

• Download, compile and install caffe

• Set the variables in parameters.json

• set PYTHONPATH to point to caffe_layer

    source set_python_env.sh
  

• Resize cityscapes images/annotations

    python scripts/resize_dataset.py
  

• SegNet

    ./caffe/build/tools/caffe train --solver models/solver_segnet.prototxt
  

    python ./caffe/build/install/python/train.py --solver models/solver_segnet.prototxt
  

• FCN

  ./caffe/build/tools/caffe train --solver models/fcn/solver_fcn.prototxt --weights weights/fcn8s-heavy-pascal.caffemodel
  

TEST

python scripts/test_segmentation.py --model models/segnet_inference.prototxt --weights weights/segnet_iter_50000.caffemodel --test_file dataset/cityscapes_val.txt
python scripts/test_segnet.py --test_file dataset/cityscapes_val.txt --gpu 0
python scripts/test_fcn.py --test_file dataset/cityscapes_val.txt --gpu 0

TOOLS

• Plot training loss:

  python caffe/tools/extra/parse_log.py log_file processed_log

  python scripts/plot_loss.py --log processed_log