backseason/dfi

Code for our IEEE TIP 2020 paper "Dynamic Feature Integration for Simultaneous Detection of Salient Object, Edge and Skeleton"

Dynamic Feature Integration for Simultaneous Detection of Salient Object, Edge and Skeleton

This is a demo PyTorch implementation of our IEEE TIP 2020 paper.

We also provide an Online Demo.

Prerequisites

• Pytorch 0.4.1+
• torchvision
• opencv

Demo usage

1. Clone the repository

git clone https://github.com/backseason/DFI.git
cd DFI/

2. Download the pretrained model

dfi.pth GoogleDrive | BaiduYun (pwd: wkeb) and move it to the pretrained folder.

3. Test (demo)

The source images are in the demo/images folder. By running

python main.py

you'll get the predictions under the demo/predictions folder. The predictions of all the three tasks are performed simultaneously.

4. Pre-computed results and evaluation results

You can find the pre-computed predictions maps of all the three tasks and their corresponding evaluation scores with the following link: Results reported in the paper GoogleDrive | BaiduYun (pwd: 7eg3)

5. Contact

If you have any questions, feel free to contact me via: j04.liu(at)gmail.com.

If you think this work is helpful, please cite

@article{liu2020dynamic,
  title={Dynamic Feature Integration for Simultaneous Detection of Salient Object, Edge and Skeleton},
  author={Jiang-Jiang Liu and Qibin Hou and Ming-Ming Cheng},
  journal={IEEE Transactions on Image Processing},
  year={2020},
  volume={},
  number={},
  pages={1-15},
  doi={10.1109/TIP.2020.3017352},
}

dfi's Issues

PR曲线

Hi, 想问下，论文中的PR曲线怎么画的，我对比了其他论文，你跟他们画法不一样，能否分享下代码呢
你在 backseason/PoolNet#36 提到用 https://github.com/Pchank/caffe-sal 这里的代码，但我测试过，还是跟你论文的不一样，所以有点疑问

how can i get onnx format DFI model?

excuse me, can you provide the onnx format DFI model?

Dataset

could you share with me the dataset(baidu disk)? some dataset are difficult to find

training dataset

Thansk for your great work!
I find that the number of training images in saliency detection, edge detection and skeleton detection is different. Thus, I am curious about how to learn the three tasks simulataneously?

特征可视化

作者您好，非常感谢您的分享。
有个问题想请教您一下，您是如何获取和绘制中间层特征的可视化图的。可否分享一下呢？
非常感谢

Implementation details

Thanks for your amazing work.
After reading your paper and code, I have few question about some implementation details.

1. I notice that you set the bias to false in almost all convolution layers except the prediction layers. Can you tell me the reason?

2. Instead of using Batchnorm (BN), you use Groupnorm (GN) . Did you compare these two methods and how much improvement GN brings?

3. While in most blocks relu layer is applied after the GN layer, I notice that in some blocks you don't use the relu layer. For example:
   
   I am curious about how to judge whether a relu layer is necessary.

training detail

你好，作者，这篇文章非常有创新性，尤其是这种训练模式，请问可以分享一下训练的代码吗，我想充分理解这种过程，谢谢作者。

论文的提问

在表Ⅲ中，w/o是什么意思，正片论文都没有提及，作者可以解释一下吗，谢谢！

DFI RESULTS OF HKU-IS DATASET

Hi, Can you please provide the whole results(4,447 predictions) of the HKU-IS dataset ?