This repository provides the Pytorch implementation of TranSalNet: Towards perceptually relevant visual saliency prediction published in the Neurocomputing paper.

Overview:

• Python 3.8
• Pytorch 1.7.1
• Torchvision 0.8.2
• OpenCV-Python 4.5.1
• SciPy 1.6.0
• tqdm 4.56.0

TranSalNet has been implemented in two variants: TranSalNet_Res with the CNN backbone of ResNet-50 and TranSalNet_Dense with the CNN backbone of DenseNet-161.
Pre-trained models on SALICON training set for the above two variants can be download at:

• TranSalNet_Res: Google Drive / BaiduYun (access code: 1234)
• TranSalNet_Dense: Google Drive / BaiduYun (access code: 1234)

It is also necessary to download ResNet-50 (for TranSalNet_Res) and DenseNet-161 (TranSalNet_Dense) pre-trained models on ImageNet. These models can be download at:

• ResNet-50: Google Drive / BaiduYun (access code: 1234)
• DenseNet-161: Google Drive / BaiduYun (access code: 1234)

The pre-trained models should be downloaded and put in the folder named pretrained_models in the code folder first, then the following example codes can be used smoothly.
We have prepared two Jupyter Notebook files (.ipynb) for usage of TranSalNet.

• Testing: testing.ipynb. It can be used to compute and obtain the visual saliency maps of input images.
  By default, the test image and the corresponding output are in the folder named testing, and the models are loaded with parameters pre-trained on the SALCON training set.
• Fine-tuning or Training from scratch: training&fine-tuning.ipynb

  Data prepare for fine-tuning and training:
  │ dataset/
  ├── train_ids.csv
  ├── val_ids.csv
  ├── train/
  │   ├── train_stimuli/
  │   │   ├── ......
  │   ├── train_saliency/
  │   │   ├── ......
  │   ├── train_fixation/
  │   │   ├── ......
  ├── val/
  │   ├── val_stimuli/
  │   │   ├── ......
  │   ├── val_saliency/
  │   │   ├── ......
  │   ├── val_fixation/
  │   │   ├── ......
  

In the above two .ipynb files, it is possible to choose whether TranSalNet_Res or TranSalNet_Dense is used, depending on the needs and preferences.

Please note: The spatial size of inputs should be 384×288 (width×height).

If this work is helpful, please consider citing:

@article{TranSalNet,
title = {TranSalNet: Towards perceptually relevant visual saliency prediction},
journal = {Neurocomputing},
year = {2022},
issn = {0925-2312},
doi = {https://doi.org/10.1016/j.neucom.2022.04.080},
author = {Jianxun Lou and Hanhe Lin and David Marshall and Dietmar Saupe and Hantao Liu},
}