DeepUnifiedCRF_ICCV19

Main code for the ICCV2019 paper "Structured Modeling of Joint Deep Feature and Prediction Refinement for Salient Object Detection". Arxiv Link

Caffe Installation

Please install Caffe with the continous CRF layers based on the work ContinuousCRF-CNN².

cd <caffe_root>/examples
git clone https://github.com/danxuhk/ContinuousCRF-CNN.git

Training & Testing

1. The training code can be found in run_sal.py.
2. The solver can be found in solver.prototxt.
3. The network protocol can be found in train.prototxt.
4. The corresponding files for pre-training are listed in the pretrain folder.
5. The testing code can be found in test.m or test.py. We used the matlab code in our own experiement.

Pre-trained Model

The model can be downloaded here.

Note that this released model is newly trained and is slightly different from the one reported in our paper.

Saliency Maps

The resulted saliency maps on eight datasets can be downloaded here.

Citation

[1] Our paper:
@inproceedings{xu2019deepcrf, 
  title={Structured Modeling of Joint Deep Feature and Prediction Refinement for Salient Object Detection}, 
  author={Xu, Yingyue and Xu, Dan and Hong, Xiaopeng and Ouyang, Wanli and Ji, Rongrong and Xu, Min and Zhao, Guoying}, 
  booktitle={ICCV}, 
  year={2019}
} 

[2] ContinuousCRF-CNN:
@article{xu2018monocular,
  title={Monocular Depth Estimation using Multi-Scale Continuous CRFs as Sequential Deep Networks},
  author={Xu, Dan and Ricci, Elisa and Ouyang, Wanli and Wang, Xiaogang and Sebe, Nicu},
  journal={IEEE TPAMI},
  year={2018}
}