This is a PyTorch reimplementation of Holistically-nested Edge Detection (HED). The code is evaluated on Python 3.6 with PyTorch 1.0 (CUDA9, CUDNN7) and MATLAB R2018b.
Clone the repository:
git clone https://github.com/xwjabc/hed.git
Download and extract the data:
cd hed
wget https://cseweb.ucsd.edu/~weijian/static/datasets/hed/hed-data.tar
tar xvf ./hed-data.tar
Train:
python hed.py --vgg16_caffe ./data/5stage-vgg.py36pickle
The results are in output
folder. In the default settings, the HED model is trained for 40 epochs, which takes ~27hrs with one NVIDIA Geforce GTX Titan X (Maxwell).
Evaluate:
cd eval
(echo "data_dir = '../output/epoch-39-test'"; cat eval_edge.m)|matlab -nodisplay -nodesktop -nosplash
The evaluation process takes ~7hrs with Intel Core i7-5930K CPU @ 3.50GHz.
Besides, based on my observation, the evaluated performance is somewhat stable after 5 epochs (5 epochs: ODS=0.788 OIS=0.808 vs. 40 epochs: ODS=0.787 OIS=0.807).
Evaluate the my pre-trained version:
python hed.py --checkpoint ./data/hed_checkpoint.pt --output ./output-mypretrain --test
cd eval
(echo "data_dir = '../output-mypretrain/test'"; cat eval_edge.m)|matlab -nodisplay -nodesktop -nosplash
The result should be similar to ODS=0.787 OIS=0.807.
Evaluate the official pre-trained version:
python hed.py --caffe_model ./data/hed_pretrained_bsds.py36pickle --output ./output-officialpretrain --test
cd eval
(echo "data_dir = '../output-officialpretrain/test'"; cat eval_edge.m)|matlab -nodisplay -nodesktop -nosplash
The result should be similar to ODS=0.788 OIS=0.806.
This reimplementation is based on lots of prior works. Thanks to Saining for the original Caffe implementation. Thanks to @meteorshowers for a PyTorch implementation where I adopt most of the code from. Thanks to @jmbuena for a fixed version of Piotr's Toolbox. Thanks to Berkeley Institute for Data Science which provides a mirror of BSDS500 dataset (the original link to the dataset seems broken).