2021-09-01 This is a pytorch implementation of our FCL-Net, 2021, version 1.0.

2021-11-11 Update paper information, and plan to add docker soon, version 2.0.

​ This is the implements of our work FCL-Net: Towards accurate edge detection via Fine-scale Corrective Learning in Pytorch. For more details, please refer to our paper.

​ Integrating multi-scale predictions has become a mainstream paradigm in edge detection. However, most existing methods mainly focus on learning to effectively extract and fuse multi-scale features while ignoring the deficient learning capacity at fine-level branches, limiting the overall fusion performance. In light of this, we propose a novel Fine-scale Corrective Learning Net (FCL-Net) that exploits semantic information from deep layers to facilitate fine-scale feature learning. FCL-Net mainly consists of a Top-down Attentional Guiding (TAG) and Pixel-level Weighting (PW) module. The TAG adapts semantic attentional cues from coarse-scale prediction into guiding the fine-scale branches by learning a top-down LSTM. The PW module treats each spatial location's importance independently, promoting the fine-level branches to detect detailed edges with high confidence. We evaluate our method on three widely used datasets, BSDS500, Multicue and BIPED. Our approach significantly outperforms the baseline and achieves competitive ODS F-measure of 0.826 on BSDS500 benchmark.

Here gives some examples of edge detection results, comparing with existing methods in Figure (a). As shown in Figure (b), our method greatly improves fine-scale feature learning and detects more detailed edges with accurate location.

(a)	(b)

We report ODS and OIS for comparison with other previous impressive works. Moreover, we reproduce HED, RCF and BDCN, and report the performance compared with the original paper.

• Pytorch>=0.3.1
• Tensorboard
• AttrDict
• For more details about the environment of conda, please refer to requirements.txt.

1. Clone this FCL-Net repository.

2. Prepare datasets

   • to download current famous edge detection dataset, you can refer to https://github.com/MarkMoHR/Awesome-Edge-Detection-Papers to prepare data. PASCAL Context dataset is available here, but needs to extract edges from segmentation masks. For convenience, RCF provides extracted annotations for NYUD and Pascal Context, as well as preprocessed datasets like BSDS500.
   • data augmentation: HED and BIPED have provide augmented dataset or code for augmentation, mainly including scaling and rotation. If you want to know more details about data augmentation, you can refer to the HED paper.
   • put the data into \data folder, and prepare the image list referring to train_pair.lst.

3. Download ImageNet pretrained parameters

   • Here we use the pretrained vgg16_bn model to initialize the backbone model. To download parameters provide by Pytorch, you can refer to link. And you need to put it into ./pytorch_net/models/;
   • BDCN also provides the source code to convert a VGG16-caffe model into pytorch version. If you are interested in this, you can refer to BDCN.

4. Training

   • configure the parameters for training in ./pytorch_net/config, we provide configuration files for our work. We also provide the configuration files for our reproduced HED, RCF and BDCN by pytorch.

     standard_HED.yaml  # for HED
     standard_RCF.yaml  # for RCF
     standard_BDCN.yaml  # for BDCN
     standard_BAN.yaml  # for BAN, now we only build the network according to the paper, the training code will be added in the future.
     standard_FCL.yaml  # for our work, FCL-Net

     This reproduced BDCN has an ODS of 0.809, which is a little bit lower than original source code.

   • submit your task;

     sbatch FCL_submit.sh

5. Evaluation

   • To evaluate the model, please refer to Testing HED part in https://github.com/s9xie/hed;
   • Note that you need to use Piotr's Structured Forest matlab toolbox available here https://github.com/pdollar/edges, and remember to do NMS first if needed and put files in ./matlab_code;
   • we provide an example here ./evaluation/eval_epoch_fcl.m of Matlab.

6. To draw P-R curves and compare with other works

   • you can refer to https://github.com/yun-liu/plot-edge-pr-curves for details;

7. Our pretrained models.

   Model	Link	ODS
   HED	baiduyun	0.790
   RCF	baiduyun	0.807
   BDCN	baiduyun	0.809
   BDCN-official	baiduyun	0.810
   FCL	baiduyun	0.815

   Password for baiduyun: repr ;

• The Docker-file for testing purposes will be available soon.

• We reveal the evaluation results after NMS of our method and our reproduced version of RCF, BDCN as well as the baseline for comparison with other work in \evaluation\eval;

Note that our reproduce version of BDCN employs the original code released by the authors here https://github.com/pkuCactus/BDCN. However, we didn't reach the performance caused by subtle difference in data augmentation as referred to the author He. If you want to know more details about BDCN, you can refer to project.

• We have searched for the optimal setting for the loss weights of side outputs and final output, and the experiment results are listed below for reference.

@article{XUAN2022248,
    title = {FCL-Net: Towards accurate edge detection via Fine-scale Corrective Learning},
    author = {Wenjie Xuan and Shaoli Huang and Juhua Liu and Bo Du},
    journal = {Neural Networks},
    volume = {145},
    pages = {248-259},
    year = {2022},
    issn = {0893-6080},
    doi = {https://doi.org/10.1016/j.neunet.2021.10.022},
    url = {https://www.sciencedirect.com/science/article/pii/S0893608021004135},
}

[1]. Our implementation is based on this project by chongruo, and we also refer to wonderful projects of Liu and He. Thank you for their wonderful works and all contributors;

[2]. When doing experiments, we also emailed Liu and He. Thanks very much for their kind responses and helpful advice.