Supervision by Fusion: Towards Unsupervised Learning of Deep Salient Object Detector

Pytorch inference.

We add a Pytorch inference demo for Pytorch users! We convert the caffe inference prototxt and weights into pytorch model. The conversion codes are modified from caffemodel2pytorch.

• Usage: Change the image directory in demo_pytorch.py, them run the script, you'll get saliency maps in data/output.

Paper

Read our paper here .

Requirements

• Make sure you have installed caffe and matcaffe to run the demo code.
• Since the training procedure contains a modified loss layer, if you want to run codes of training part, you'll have to install this modified caffe version.

Installation and demo

1. Clone the repository

   git clone https://github.com/zhangyuygss/SVFSal.caffe.git

2. cd into the directory you cloned, let's call it SVFS_ROOT

3. Download our pretrained caffemodel, put it into SVFS_ROOT/data.

   You can download the pretrained model from here.

4. Modify caffe path in file demo.m:

   Change this line:

   addpath(genpath('/your/path/caffe-master/matlab/'));

   to your own caffe path.

5. Run demo.m

   After running demo code, you should find the output saliency map from SVFS_ROOT/data/output. Put your images into SVFS_ROOT/data/images if you want to run on your own dataset.

   If you would like to download our saliency map results, you can also find them here via google drive or here.

Training and evaluation

• Script prepare_and_train.m contains one round training of our framework, it calls corresponding script to do fusion, data preparation and start training. However, we don't recommend you to run training procedure since the codes are not well modified and you have to change some path in the script to make it run on your machine. We only release the training part as a reference for you if you want to check some details in the codes.
• Evaluation code can be found in SVFS_ROOT/evaluate/evaluate_net.m, also the code serves as a reference. If you have any question about the codes, feel free to ask me.

Citation

Cite our paper with:

@article{zhang2019synthesizing,
  title={Synthesizing supervision for learning deep saliency network without human annotation},
  author={Zhang, Dingwen and Han, Junwei and Zhang, Yu and Xu, Dong},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={42},
  number={7},
  pages={1755--1769},
  year={2019},
  publisher={IEEE}
}
@INPROCEEDINGS{DZhang2017SVFSal,
	author = {Dingwen, Zhang and Junwei, Han and Yu, Zhang},
	title = {Supervision by fusion: towards unsupervised learning of deep salient object detector},
	booktitle = {ICCV},
	year = {2017}
}