HED in pytorch

This work is an implementation of paper Holistically-Nested Edge Detection.

Performance

Input Image	dsn1	dsn2	dsn3	dsn4	dsn5	Fusioned Output (dsn6)

On BSDS500

Method	ODS (Fusion/Merged)	OIS (Fusion/Merged)	AP (Fusion/Merged)
Our Implementation	0.78731/0.78280	0.80623/0.80356	0.78632/0.83851
Original Paper	0.782/0.782	0.802/0.804	0.787/0.833

As mentioned in the paper, Fusion refers to the fusion-output(dsn6) and Merged means results of combination of fusion layer and side outputs.

How to Run

Prerequisite:

• Pytorch>=0.3.1
• Tensorboard
• AttrDict

Training/Testing

The coda/data structure

$ROOT
  - ckpt           # save checking points
  - data           # contains BSDS500
  - matlab_code    # test code
  - pytorch-HED    # current repo

To prepare for data, please refer to Training HED part in https://github.com/s9xie/hed

For training

python submit.py

Create your custom configuration file (xxx.yaml) in ./config, and modify config_file in submit.py.

Our implementation is a little different form the original caffe version. We used vgg architecture with BN layers, and also more data argumentations.

For testing, please install the Piotr's matlab toolbox first. Please refer to https://github.com/s9xie/hed.

References

@InProceedings{xie_HED,
author = {"Xie, Saining and Tu, Zhuowen"},
Title = {Holistically-Nested Edge Detection},
Booktitle = "Proceedings of IEEE International Conference on Computer Vision",
Year  = {2015},
}

Related Projects

[1]. Original Implementation by @s9xie

[2]. hed by @xwjabc

[3]. hed-pytorch by @meteorshowers

[4]. hed(caffe) by @zeakey