This repository contains the Tensorflow implementation of our paper "CAGNet: Content-Aware Guidance for Salient Object Detection". The paper can be found at: [Sciencedirect] [arXiv]
1- Clone the repository
git clone https://github.com/Mehrdad-Noori/CAGNet
cd CAGNet
2- If you want to train the model, download the following dataset and unzip it into data folder.
3- To run the training, set the arguments or use the default settings:
python train.py --backbone_model 'ResNet50' --batch_size 10 --save_dir 'save'
The backbone_model can be one of the following options: VGG16, ResNet50, NASNetMobile or NASNetLarge
4- To generate saliency maps:
python predict.py --model '/path/to/trained/model' --input_dir /path/to/input/images/directory --save_dir 'save'
You can also download and use our pre-trained models
5- Evaluation code
You can use this toolbox to compute different saliency measures such as E-measure, Weighted F-measure, MAE, PR curve ...
We provide the pre-trained model and pre-computed saliency maps for DUTS-TE, ECSSD, DUT-OMRON, PASCAL-S, and HKU-IS datasets.
- CAGNet-V (VGG16 backbone): [pre-trained model] - [saliency maps]
- CAGNet-R (ResNet50 backbone): [pre-trained model] - [saliency maps]
- CAGNet-M (NASNet Mobile backbone): [pre-trained model] - [saliency maps]
- CAGNet-L (NASNet Large backbone): [pre-trained model] - [saliency maps]
Please feel free to contact me, or raise an issue if you encounter any problems.
@article{mohammadi2020cagnet,
title={CAGNet: Content-Aware Guidance for Salient Object Detection},
author={Mohammadi, Sina and Noori, Mehrdad and Bahri, Ali and Majelan, Sina Ghofrani and Havaei, Mohammad},
journal={Pattern Recognition},
pages={107303},
year={2020},
publisher={Elsevier}
}