This is the official source code of paper "A Doubly Decoupled Network for Edge Detection" (DDN), which is accepted by Neurocomputing.

Deep learning-based edge detectors are increasingly scaled up to achieve higher performance. However, this growth comes at the cost of higher computational demands and a greater risk of model overfitting. In this paper, a Doubly Decoupled Network (DDN) is proposed to alleviate these problems by decoupling the data and features separately. Firstly, we decouple image-edge pairs and build learnable Gaussian distributions. The images are mapped to learnable Gaussian distributions, and the corresponding edges are mapped to samples of the distributions. More robust features are generated by normalizing the latent variables and sampling from the Gaussian distributions. Secondly, we compress the redundant calculations by decoupling image features. Shallow features need not only high resolution to provide accurate spatial information, but also diversity to provide sufficient cues for deep features. We decouple image features into spatial and semantic features and encode them separately. Shallow features no longer provide cues for semantic features and their diversity can be drastically compressed. The proposed DDN achieves the state-of-the-art accuracy across multiple edge detection benchmarks, while the computational cost is similar to that of VGG-based methods.

Download the project

git clone git@github.com:Li-yachuan/DDN.git

Install the requirement packages Torch-related packages need to be installed manually. DDN is quite robust to the environment. If the same environment cannot be configured due to CUDA version issues, it can work with a relatively close basic one

pip install -r requirements.txt

BSDS500: following the setting of MuGE, and the data can be download from the link

put the unziped data to any path (such as "/data/users/liyachuan/dataset/BSDS")

And Change data_pth in config/BSDS-DDN_M36.yaml to data_pth: /data/users/liyachuan/dataset/BSDS

DDN uses the pretrained CAformer as the backbone, so the ckpt of CAformer have to be download before train DDN.
DDN-M36 is based on caformer_m36, and the ckpt can be found here
DDN-S18 is based on caformer_s18, and the ckpt can be found here
DDN-B36 is based on caformer_b36, and the ckpt can be found here

Download the ckpt and link it in the config file 'config/BSDS-DDN_M36.yaml'(or the other config file you used) ckpt.

python main.py  --output bsds --cfg config/BSDS-DDN_M36.yaml

python main.py  --output bsds-test --cfg config/BSDS-DDN_M36.yaml --mode test --resume [path of ckpt] --mg --ms

mg means generating multi-granularity edges ms means generating multi-scale edges

The pretrained ckpt of DDN is here, if you donot want to train DDN, you can test the ckpt directly.

Following MuGE, the single-granularity evalution depend on the Matlab package.

And the multi-granularity edges evalution by the commond

python eval_muge_best/best_ods_ois.py [multi-granularity edges dir]

more detail can be found in MuGE

NYUDv2 following RCF, and config the data_pth in file "config/NYUD-DDN_M36.yaml"
BIPED following DexiNed, and config the data_pth in file "config/BIPED-DDN_M36.yaml"

The edges of these two datasets are in png format, which is different from the mat format of BSDS. This does not affect their usage.

@inproceedings{liu2017richer,
  title={Richer convolutional features for edge detection},
  author={Liu, Yun and Cheng, Ming-Ming and Hu, Xiaowei and Wang, Kai and Bai, Xiang},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  pages={3000--3009},
  year={2017}
}

@inproceedings{poma2020dense,
  title={Dense extreme inception network: Towards a robust cnn model for edge detection},
  author={Poma, Xavier Soria and Riba, Edgar and Sappa, Angel},
  booktitle={Proceedings of the IEEE/CVF winter conference on applications of computer vision},
  pages={1923--1932},
  year={2020}
}

@inproceedings{zhou2024muge,
  title={Muge: Multiple granularity edge detection},
  author={Zhou, Caixia and Huang, Yaping and Pu, Mengyang and Guan, Qingji and Deng, Ruoxi and Ling, Haibin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={25952--25962},
  year={2024}
}

@article{li2024new,
  title={A new baseline for edge detection: Make Encoder-Decoder great again},
  author={Li, Yachuan and Pomab, Xavier Soria and Xi, Yongke and Li, Guanlin and Yang, Chaozhi and Xiao, Qian and Bai, Yun and LI, Zongmin},
  journal={arXiv preprint arXiv:2409.14976},
  year={2024}
}