Fast Saliency

This repository contains the code for our two fast saliency detection papers: "SAMNet: Stereoscopically Attentive Multi-scale Network for Lightweight Salient Object Detection" (IEEE TIP) and "Lightweight Salient Object Detection via Hierarchical Visual Perception Learning" (IEEE TCYB).

We use Python 3.5, PyTorch 0.4.1, cuda 9.0, and numpy 1.17.3 to test the code. We recommend using Anaconda.

Citations

If you are using the code/model/data provided here in a publication, please consider citing:

@article{liu2021samnet,
  title={{SAMNet}: Stereoscopically Attentive Multi-scale Network for Lightweight Salient Object Detection},
  author={Liu, Yun and Zhang, Xin-Yu and Bian, Jia-Wang and Zhang, Le and Cheng, Ming-Ming},
  journal={IEEE Transactions on Image Processing},
  volume={30},
  pages={3804--3814},
  year={2021},
  publisher={IEEE}
}

@article{liu2020lightweight,
  title={Lightweight Salient Object Detection via Hierarchical Visual Perception Learning},
  author={Liu, Yun and Gu, Yu-Chao and Zhang, Xin-Yu and Wang, Weiwei and Cheng, Ming-Ming},
  journal={IEEE Transactions on Cybernetics},
  volume={51},
  number={9},
  pages={4439--4449},
  year={2020},
  publisher={IEEE}
}

Precomputed saliency maps

Precomputed saliency maps for 12 widely-used saliency datasets (i.e., ECSSD, DUT-OMRON, DUTS-TE, HKU-IS, SOD, THUR15K, Judd, MSRA-B, MSRA10K, PASCALS, SED1, SED2) are available in the SaliencyMaps folder. Note that if a compressed file is larger than 100 Mb, we divided it into two files.

Testing and training

Before running the code, you should first link the saliency datasets to the ROOT folder:

ln -s /path/to/saliency/datasets/ Data

Hence, the generated Data folder contains all datasets. We have put data lists in the folder of Lists. The pretrained models are put in the Pretrained folder.

Testing SAMNet

For example, we use the following command to test SAMNet on the ECSSD dataset:

python test.py --file_list ECSSD.txt --pretrained Pretrained/SAMNet_with_ImageNet_pretrain.pth --model Models.SAMNet --savedir ./Outputs

The generated saliency maps will be outputted into the folder of ./Outputs/ECSSD/.

Training SAMNet

By default, we use the DUTS training set for training:

python train.py --pretrained Pretrained/SAMNet_backbone_pretrain.pth --model Models.SAMNet --savedir ./Results

Testing HVPNet

python test.py --file_list ECSSD.txt --pretrained Pretrained/HVPNet_with_ImageNet_pretrain.pth --model Models.HVPNet --savedir ./Outputs

Training HVPNet

python train.py --pretrained Pretrained/HVPNet_backbone_pretrain.pth --model Models.HVPNet --savedir ./Results

Useful links

Here are the project pages of our other saliency detection papers:

DNA: Deeply-supervised Nonlinear Aggregation for Salient Object Detection (IEEE TCYB 2022)

EDN: Salient Object Detection via Extremely-Downsampled Network (IEEE TIP 2022)

Regularized Densely-connected Pyramid Network for Salient Instance Segmentation (IEEE TIP 2021)

A Simple Saliency Detection Approach via Automatic Top-Down Feature Fusion (Neurocomputing 2020)