Suppress and Balance: A Simple Gated Network for Salient Object Detection

Xiaoqi Zhao*, Youwei Pang*, Lihe Zhang, Huchuan Lu, Lei Zhang
⭐ arXiv » [Slides]

The official repo of the ECCV 2020 (oral) paper, Suppress and Balance: A Simple Gated Network for Salient Object Detection.

Results(RGB/RGB-D SOD, VOS)

RGB SOD: Google Drive / BaiduYunPan(4ha3) RGB-D SOD: Google Drive / BaiduYunPan(9her) VOS: Google Drive / BaiduYunPan(0hue)

Related Works

• (ECCV 2020) A Single Stream Network for Robust and Real-time RGB-D Salient Object Detection: https://github.com/Xiaoqi-Zhao-DLUT/DANet-RGBD-Saliency
• (ECCV 2020) Hierarchical Dynamic Filtering Network for RGB-D Salient Object Detection: https://github.com/lartpang/HDFNet
• (CVPR 2020) Multi-scale Interactive Network for Salient Object Detection: https://github.com/lartpang/MINet

Trained Model

You can download the trained VGG-16/ResNet-50 GateNet model at BaiduYunPan(s3l2).

Requirement

• Python 3.7
• PyTorch 1.5.0
• torchvision
• numpy
• Pillow
• Cython

Training

1.Set the path of training sets in config.py
2.Run train.py
We also release the new baseline, Gated FPN in [model_GatedFPN_newbaseline.py]. You can use it for future reserach.

Testing

1.Set the path of testing sets in config.py
2.Run generate_salmap.py (can generate the predicted saliency maps)
3.Run generate_visfeamaps.py (can visualize feature maps)
4.Run test_metric_score.py (can evaluate the predicted saliency maps in terms of fmax,fmean,wfm,sm,em,mae). You also can use the toolkit released by us:https://github.com/lartpang/Py-SOD-VOS-EvalToolkit.

We also recommend you to use another code released by Jun Wei https://github.com/weijun88/F3Net. Based on the training strategy of this code, our performance can be further improved( but for fair comparisons, we do not use the multi-scale training trick), the performance as follows:
(ResNet-50-GateNet)
dataset: DUTS MAE: 0.0375 maxF: 0.8899 avgF: 0.8199 wfm: 0.8225 Sm: 0.8889 Em: 0.9133
dataset: HKU-IS MAE: 0.0317 maxF: 0.9353 avgF: 0.9010 wfm: 0.8873 Sm: 0.9187 Em: 0.9558
dataset: ECSSD MAE: 0.0355 maxF: 0.9486 avgF: 0.9217 wfm: 0.9049 Sm: 0.9269 Em: 0.9513
dataset: DUT-OMRON MAE: 0.0536 maxF: 0.8157 avgF: 0.7497 wfm: 0.7365 Sm: 0.8386 Em: 0.8697
dataset: PASCAL-S MAE: 0.0648 maxF: 0.8882 avgF: 0.8400 wfm: 0.8213 Sm: 0.8625 Em: 0.8958
In addition, this performance is based on a model that reduces the number of output channels of Fold-ASPP by half, so the entire parameter and model size can be greatly reduced. We believe that Fold-ASPP can still be optimized in the future.

BibTex

@inproceedings{GateNet,
  title={Suppress and Balance: A Simple Gated Network for Salient Object Detection},
  author={Zhao, Xiaoqi and Pang, Youwei and Zhang, Lihe and Lu, Huchuan and Zhang, Lei},
  booktitle=ECCV,
  year={2020}
}