Andrew-Qibin/DSS

Deeply Supervised Salient Object Detection with Short Connections

What's new!!!

A new fast approach is now available. Here is the PyTorch implementation. Here is the link to the project page.

Network architecture and more details

Please refer to our paper.

Usage

Please install Caffe first. I think you may find a great number of tutorials talking about how to install it.

cd <caffe_root>/examples
git clone https://github.com/Andrew-Qibin/DSS.git

Before you start, you also need our pretrained model.

wget http://mftp.mmcheng.net/Andrew/dss_model_released.caffemodel

You can also download it from here (google drive). The results produced by this model are slightly different from the ones we reported in our paper (with higher F-measure score and also higher MAE score).

If you want to train the model, please prepare your own training dataset first. The data layer we used here is similar to the one used in HED. You can also refer to the data layer used in Deeplab or write your own one.

You may also find our data layer here. Notice that if you use caffe, please cite their paper.

Then, run

python run_saliency.py

If you want to test the model, you can run

ipython notebook DSS-tutorial.ipynb

About the CRF code we used, you can find it here. Notice that please provide a link to the original code as a footnote or a citation if you plan to use it.

Visual comparison with previous start-of-the-arts

From left to right: Source, Groundtruth, Ours, DCL, DHS, RFCN, DS, MDF, ELD, MC, DRFI, DSR.

Useful links that you might want

MSRAB: including 2500 training, 500 validation, and 2000 test images. (This is also our training set.) The source images can be found here.
MSRA10K: You can also use this dataset for training as some works did.
Evaluation Code (Windows): The cold is based on MS Visual Studio.
Evaluation Code (Ubuntu): This code is based on C++ and with a python wrapper for python users.

We add the resnet version of our model into this repo. Also, a larger set of training data can be found in the lists dir. ResNet version caffemodel can be found here (google drive).

If you want to compare your results with ours, you may download them from here (Baidu Drive) or (Google Drive).

If you think this work is helpful, please cite

@article{HouPami19Dss,
  title={Deeply Supervised Salient Object Detection with Short Connections},
  author={Hou, Qibin and Cheng, Ming-Ming and Hu, Xiaowei and Borji, Ali and Tu, Zhuowen and Torr, Philip},
  year  = {2019},
  volume={41}, 
  number={4}, 
  pages={815-828}, 
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}
}