Bottom-Up Visual Attention: Interactive Demonstration

We present here an interactive demonstration of the performance of our model for bottom-up, task-independent selective visual attention. This demonstration is written in JavaTM, which you need to enable in your browser in order to use the demo.

You may want to check this screen grab of the applet running flawlessly on a Silicon GraphicsTM IndyTM workstation with Netscape NavigatorTM 2.0 and compare it to what you obtain with your web browser.

  • The Traffic Signs database was kindly provided to us by Werner Ritter at Daimler-Benz. Each 512x384 image is evaluated by the model in approximately 10-15 seconds on a PC computer.
  • The Military Vehicle database is the "Search 2" image database which Dr. Lex Toet at TNO-Human Factors Research Institute, The Netherlands was kind enough to send us; this database of high-resolution images (6144x4096 pixels) has been designed to test human as well as model target detection performance with complex natural scenes. Each 70Mb-image takes 15 minutes and 484Mb of RAM to evaluate on a 500MHz DEC/Alpha workstation.
  • The Pedestrians database was kindly made available to us by Constantine Papageorgiou, who works with Tommy Poggio at MIT/CBCL as part of the MIT/Harvard/Caltech MURI project; with this database, we are investigating whether a rapid selection of a few interesting image locations might speed-up Constantine's pedestrian detector algorithm.
  • The Pedestrians - Trained database was evaluated using a specialized version of our model, in which supervised learning allowed us to tune the weights of the different feature maps in the model. The first, third, fifth, etc. images were used as training set, and the remaining images as test set.
  • The Good Housekeeping database was submitted to us by Peter Walker, Director of the Neural Nets Group at McCann-Erickson, McCann-Erickson House, 36 Howland Street LONDON W1A 1AT (McCann is one of the world's largest advertising agencies, part of IPG. HQ in NY).
  • The other image databases were either collected from the web or shot using a consumer-electronics video camera.

Let's Try It! (Netscape only)
No parameter tuning! All images processed with same model.

New!   Internet Explorer Version!
In Microsoft's Internet Explorer, the Java function java.awt.Graphics.drawImage(Image, x, y, w, h, ImageObserver) does not work properly. So we compiled an alternate version of the applet that pre-scales all images, and then calls the simpler function java.awt.Graphics.drawImage(Image, x, y, ImageObserver), which appears to have been correctly implemented by Internet Explorer. This is quite slow, especially when you try to scroll the thumbnails left or right, but appears to work under Internet Explorer.

Copyright © 2000 by the University of Southern California, iLab and Prof. Laurent Itti