This is our code for modeling top-down attention.
Please also see my research page for illustrative videos.

For now, we are sharing data for two games:
1) 3D Driving school and
2) Hot Dog Bush.

Data includes:

1- Data (features for each frame including gist, bottom-up saliency map, actions and events) as well as 
corresponding eye position (one for each frame) and saccades (for frames that a saccade occured).
There is one .zip file for each game.


2- A video file for each game where you can extract frames and calculate more features if you like (inclusing 1).
(One for each game). Currently we share data of 5 subjects over 3DDS as all data is very large.

3- Annotations (a polygon drawn around objects of the hotdog bush).

4- Code.zip. This files contains our space-based classifiers as well as Dynamic Bayesian Network (DBN) models
for prediction of the next saccade or the attended object. After unzipping this file, you will see another
README.txt file explaining how to run the code. This code is mainly for the HDB game but with some minor
changes, you should be able to run it over the 3DDS game (for instance by feeding the 3DDS features to classifiers).


Note: Our documentation is not yet complete. In case of facing any troubles please email one of us:
Ali Borji:  borji@usc.edu OR
Dicky n. Sihite: sihite@usc.edu

We will appreciate if you can share your opinions in this regard with us.

Sincerely.