PyTorch Implementation of SalGAN: Visual Saliency Prediction with Generative Adversarial Networks

This repository contains a PyTorch implementation of SalGAN: Visual Saliency Prediction with Generative Adversarial Networks by Junting Pan et al,. The model learns to predict a saliency map given an input image.

I hope you find this implementation useful.

Results

TODO:

Example Generations

TODO: Predict.ipynb

Training

The code requires a pytorch installation.

Before you train the model, preprocess the dataset by running preprocess.py to resize the ground truth images and saliency maps to 256x192.

To train the model, refer to main.py.

Pretrained Model

TODO:

Data

We used the SALICON dataset for training.

Reference:

https://imatge-upc.github.io/saliency-salgan-2017/