sunxjtu / bdcn

The code for the paper Bi-Directional Cascade Network for Perceptual Edge Detection( CVPR2019 ),and add the re_Dice loss from the paper Learning to Predict Crisp Boundaries (ECCV 2018)

Bi-Directional Cascade Network for Perceptual Edge Detection(BDCN)

Learning to Predict Crisp Boundaries(ECCV 2018)

This code is update thecode into python3,and add the re_Dice loss which mentioned in Learning to Predict Crisp Boundaries(ECCV 2018).we apply it into face edge detection task,the experiment has achieved good results.

The first paper proposes a Bi-Directional Cascade Network for edge detection. By introducing a bi-directional cascade structure to enforce each layer to focus on a specific scale, BDCN trains each network layer with a layer-specific supervision. To enrich the multi-scale representations learned with a shallow network, we further introduce a Scale Enhancement Module (SEM). Here are the code for this paper.

The second paper mainly proposed a useful loss to help the network learning a crisp edge .

the experiment results

(left:the original results,right:add the re_Dice loss results)

Prerequisites

• pytorch >= 0.2.0(Our code is based on the 0.2.0)
• numpy >= 1.11.0
• pillow >= 3.3.0
• python3

Train and Evaluation

1. Clone this repository to local

git clone https://github.com/pytorch/pytorch.git

2. Download the imagenet pretrained vgg16 pytorch model [vgg16.pth](link: https://pan.baidu.com/s/10Tgjs7FiAYWjVyVgvEM0mA code: ab4g) or the caffemodel from the model zoo and then transfer to pytorch version. You also can download our pretrained model for only evaluation. The google drive link.

3. Download the dataset to the local folder

4. running the training code train.py or test code test.py

Pretrained models

BDCN model for BSDS500 dataset and NYUDv2 datset of RGB and depth are availavble on Baidu Disk.

The link https://pan.baidu.com/s/18PcPQTASHKD1-fb1JTzIaQ
code: j3de

The pretrained model will be updated soon.