GFINet

Dependencies

>= Ubuntu 16.04 
>= Python 3.6
>= Pytorch 1.3.0
OpenCV-Python

Preparation

• download the official pretrained model ([Baidu drive](https://pan.baidu.com/s/1zRhAaGlunIZEOopNSxZNxw code：fv6m)) of ResNet-50 implemented in Pytorch if you want to train the network again.
• download or put the RGB saliency benchmark datasets ([Baidu drive](https://pan.baidu.com/s/1kUPZGSe1CN4AOVmB3R3Qxg code：sfx6)) in the folder of dataset for training or test.

Generate the thicker edge mask

After preparing the data folder, you need to use the mask_edge.py to generate the thicker edge mask for training. Run this command

python data2/mask_edge.py

Generate the dilated mask

After preparing the data folder, you need to use the mask_edge.py to generate the dilated mask for training. Run this command

python data2/mask_regione.py

Training

you may revise the TAG and SAVEPATH defined in the train.py. After the preparation, run this command

python train.py

make sure that the GPU memory is enough (the original training is conducted on a one NVIDIA RTX 2080Ti (11G) card with the batch size of 24).

Test

After the preparation, run this commond to generate the final saliency maps.

 python test.py 

We provide the trained model file ([Baidu drive](https://pan.baidu.com/s/1KdP0doBCiIme4y_j4Y4OPQ code:uht6)), and run this command to check its completeness:

cksum model-20210718 

you will obtain the result model-20210718.

Evaluation

We provide the evaluation code in the folder "eval_code" for fair comparisons. You may need to revise the algorithms , data_root, and maps_root defined in the main.m.

Citation

We really hope this repo can contribute the conmunity, and if you find this work useful, please use the following citation:

@article{DBLP:journals/tnn/ZhuLG23,
author = {Ge Zhu and
Jinbao Li and
Yahong Guo},
title = {Supplement and Suppression: Both Boundary and Nonboundary Are Helpful for Salient Object Detection},
journal = {{IEEE} Trans. Neural Networks Learn. Syst.},
volume = {34},
number = {9},
pages = {6615--6627},
year = {2023},
url = {https://doi.org/10.1109/TNNLS.2021.3127959},
doi = {10.1109/TNNLS.2021.3127959},
timestamp = {Sun, 24 Sep 2023 15:45:36 +0200},
biburl = {https://dblp.org/rec/journals/tnn/ZhuLG23.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}