Visualizing Dynamic Images and Eye Movements with CARPE
The DIEM project is an investigation of how people look and see. DIEM has so far collected data from over 250 participants watching 85 different videos. All of our data is freely available for research and non-commercial use as restricted by a CC-NC-SA 3.0 Creative Commons license. The data together with CARPE will let you visualize where people look during dynamic scene viewing such as during film trailers, music videos, or advertisements. The project was made possible by generous funding from the Leverhulme Trust and the Economic and Social Research Council of the UK (Prof. John M. Henderson, Principal Investigator).
CARPE, or Computational and Algorithmic Representation and Processing of Eye-movements, allows one to begin visualizing eye-movement data in a number of ways.
There are a number of different visualization options:
- low level visual features that process the input video to show flicker or edges;
- heat-maps that show where people are looking;
- clustered heat-maps that use pattern recognition to define the best model of fixations for each frame;
- peek-through which uses the heat-map information to only show parts of the video where people are looking.
Have a look at a montage of 4 example visualizations, all of which were produced with CARPE:
This post will help you get started. Before we begin, make sure you meet the system requirements:
Windows XP+ or Vista+ (Not tested on Windows 7)
1+ GB RAM
Updated graphics card w/ 128+ MB (It is very important that you have updated your graphics card as CARPE uses the latest graphic card functionality)
Note that if you are a developer, you may want to play with the source code to try getting CARPE working on your system, whether it be OSX or Linux, as CARPE uses all platform-independent-code and libraries with the exception of reading and writing files for the Windows OS.
Installing the dependencies:
CARPE requires a number of dependencies that are all contained within the following package: Download [36 MB] and unzip the package hosted on our Google CODE repository and install each file.
Installing the binary:
CARPE can be installed by downloading and unzipping the package hosted on our Google CODE repository: CARPE.7z [80 MB]
If you are having problems unzipping these files, please install a 7z client: http://www.7-zip.org/
CARPE’s folder structure:
/bin - main executable file: CARPE.exe
/bin/data/video - input eye-movement video files
/bin/data/audio - input eye-movement audio files
/bin/data/event_data - input eye-movement tracking data
/bin/data/output - output recorded movie visualization
/bin/data/stats - output GMM heatmap statistics
With CARPE.7z comes an example video file, in /bin/data/video/50_people_brooklyn_1280x720.mp4, example eye-tracking data in /bin/data/event_data/*50_people_brooklyn_1280x720*.txt, and example audio file in /bin/data/audio/50_people_brooklyn_1280x720.wav
Initially, the video and audio files were merged as a single file. For the purposes of eye-tracking, they have been split into 2: video and audio. This ensures sample accurate measurements. A number of participants eye-movement files are available. These were recorded at 1000 Hz and sampled down to video rate.
After installing all the dependencies, begin using CARPE by running the executable:
A dialog window should appear asking you to open the eye-tracking move file. Navigate to
CARPE should now load the eye-tracking files in
/bin/data/event_data into memory and then display the main window.
Select any options that you would like (the default is for producing clustered heatmaps), and then press the ‘space bar’ to begin. You can press the ‘space bar’ or ‘x’ keys to pause/play the movie. Further, you can also press ‘z’ to advance a frame backwards, or ‘c’ to advance a frame forwards. Pressing ‘p’ will display/hide the Options panel.
A slider on the bottom of the window lets you advance to any point in the movie by using the mouse. You can also click the play/pause button to switch between play or pause.
Yes, exiting CARPE requires its own section. Always press the ESC button to exit CARPE. If you are unable to exit CARPE, you will then have to force quit CARPE by closing the console window. We are still working on resolving this issue.
If you use CARPE in your publication, we would appreciate it if you could cite the following work:
Parag K. Mital, Tim J. Smith, Robin Hill, John M. Henderson. “Clustering of Gaze during Dynamic Scene Viewing is Predicted by Motion” Cognitive Computation, Volume 3, Issue 1, pp 5-24, March 2011.