Semantic Segmentation and Edge Detection

Setup

# install mmcv and pytorch

# install other dependencies
pip install -r requirements.txt

# If you want to install potato globally
python setup.py develop

Datasets

python tools/convert_datasets/preprocess_cityscapes.py

Training

CUDA_VISIBLE_DEVICES=0,1 ./tools/dist_train.sh <path/to/config> <num_gpus>

Test

Evaluate and Visualize Segmentatation and Edges

CUDA_VISIBLE_DEVICES=0,1 ./tools/dist_train.sh <path/to/config> <path/to/ckpt> <num_gpus> ...

To evaluate F-boundary scores:

# only supports single gpu
CUDA_VISIBLE_DEVICES=0, python tools/test_boundary_fscore.py <path/to/config> <path/to/ckpt> ...

Evaluate Edges

# save edges
CUDA_VISIBLE_DEVICES=0,1 ./tools/dist_train.sh <path/to/config> <path/to/ckpt> <num_gpus> --save-edge ...

python tools/test_edge.py <path/to/config> <path/to/predictions> ...

Evaluating edges will take a very long time (depends on the hardware).