ScanGAN360

Code and model for the paper "ScanGAN360: A Generative Model of Realistic Scanpaths for 360º Images".

Requirements

This work was developed using:

* python 3.7.4
* pytorch 1.2.0
* cudatoolkit 10.0.30
* opencv 4.1.2

You can install an environment with all required dependencies using scangan360.yml file in Anaconda.

Inference

The current version of the repository includes a basic, yet functional version to generate scanpaths from a 360º image using the ScanGAN360 model.

Usage

python main.py --mode inference 

This will read an image image_path = "data/test.jpg" and generate a set of scanpaths that will be saved in path_to_save = "test/". You can modify both those paths, and the number of generated scanpaths n_generated. Each of the images will contain 25 different scanpaths.

Training the model

Training is now available. [Updated June 15th]

python main.py --mode train 

Make sure you have correctly updated utils.py, including all the directories required. Also, check the data folder to download the required images and processed gaze data.

Checkpoints and models are saved periodically in the assigned folder.