SalScan is a universal framework to develop, compare and visualize Saliency models.

This module has only been tested under Python 3.8.

Use the following command to install dependencies:

pip install -r requirements.txt

And install module in edit mode with the following command:

pip install -e .

Before launching any model, you will need a dataset. A specific class need to be created for each dataset you will use. But before creating your own class, check if your dataset is available in SalScan.Dataset. For now, CAT2000, Le Meur and Toronto datasets are supported.

If you want to create your own dataset class here are some points of attention:

• The name of your class needs to have the same name as its containing file and it must end with the Dataset keyword: dataset_nameDataset
• Your class will take as an input the path of the dataset. This path must point to the original files provided by its author. Thus, anyone with this class will be able to use it in this framework without any further processing.
• Your class must inherit from AbstractDataset.
• Your class need to implement all methods as in already existing datasets in order to work.

Go check the code from already existing datasets and AbstractDataset to develop your own.

A dataset is initialised with its path:

from SalScan.Dataset.CAT2000Dataset import CAT2000Dataset

cat2000 = CAT2000Dataset(path="/path/to/cat200/dataset")

You can then populate your dataset by scanning the directory and load data:

cat2000.populate()

Note that this is done automatically in sessions.

A model can be run in standalone but with SalScan you can use sessions. A session is a unique association of a dataset, a model, parameters and metrics. You will need the three to run a session. Sessions are generic and does not need to be re-implemented or edited. Heres is how you can use them:

Here goes how to use models in standalone.

Metrics available.

This work has been started during Alexandre Milisavljevic PhD thesis and extended by Alessandro Mondin through the support of Ittention.

@PHDTHESIS{mili2020,
    url = "http:https://www.theses.fr/2020UNIP5136",
    title = "Visual exploration of web pages : understanding its dynamic for a better modelling",
    author = "Milisavljevic, Alexandre",
    year = "2020",
}