SCRN

Code repository for our paper "Stacked Cross Refinement Network for Edge-Aware Salient Object Detection", ICCV 2019 poster. Paper and supplementary material are available.

Change Log

2020.11.4：We update the predicted saliency maps of 20 algorithms on SOC test and validation sets!

Framework

Experiments

Results on traditional datasets
Results on SOC (attribute-based performance, structure simalarity scores), more comparison can be found in SOC Leaderboard

Usage

Requirements
- pytorch 0.40+
- scipy
Clone the repo

git clone https://github.com/wuzhe71/SCRN.git 
cd SCRN

Train/Test
- Train
  - Download datasets: DUTS, DUT-OMRON, ECSSD, HKU-IS, PASCAL-S, THUR15K, SOC
  - Set your dataset path, then
```
python train_SCRN.py
```
  - We only use multi-scale traing for data agumentation, and the lr is set as 0.002. If you change to single-scale training, the lr should better change to 0.005.
- Test
  - Download the pre-trained model from google drive or baidu yun (code: ilhx), and put it in './model/'. This model is only trained on the training set of DUTS and tested on other datasets, including SOC and test set of DUTS. Set your dataset path, then
```
python test_SCRN.py
```
  - You can also download the pre-computed saliency maps from google drive or baidu yun (code: 8mty).

SOC saliency maps

In the paper, we compare SCRN with nine methods on SOC validation set. Here we provide saliency maps of 20 SOD methods on both test and validation sets (google drive or baidu yun (code: wnjp)): DSS、NLDF、SRM、Amulet、DGRL、BMPM、PiCANet-R、R3Net、C2S-Net、RANet、CPD、AFN、BASNet、PoolNet、SCRN、SIBA、EGNet、F3Net、GCPANet、MINet.

Citation

@InProceedings{Wu_2019_ICCV,
author = {Wu, Zhe and Su, Li and Huang, Qingming},
title = {Stacked Cross Refinement Network for Edge-Aware Salient Object Detection},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {October},
year = {2019}
}

Contact Us

If you have any question, please contact us (wuzh02@pcl.ac.cn).