"This is a coming-of-age moment in this field," says Perona. "By choosing among existing machine vision techniques, we were able to put together a system that is much more capable than anything that had been demonstrated before."
The team fed the computer the characteristic details of what each individual behavior looks like on video. A lunge, for instance, begins with a shortening of the fly's body as the fly rears up; the fly then makes a quick darting movement, closing to within a few centimeters of another fly.
Once the computer had mastered these details, the researchers then compared the computer's analysis of a piece of video to the analysis produced by a human. "We looked at how many instances the computer caught, and how many it missed," says Anderson. "By looking at the errors the computer made, we were able to further refine our descriptions to create an even more accurate system."
In the end, Anderson notes, this back-and-forth resulted in a program that is "actually better than humans at detecting some of the instances of the various behaviors."
"Where previous experiments had been carried out on 100 to 1,000 frames of video, we carried out our experiments on 100,000 frames of video," Perona adds. "And while previous experiments showed numerous errors in tracking, we get very few. We are able to give accurate performance figures."
The next step, says Anderson, is to try to extend this automatic behavior-detection system to mice--a more difficult task when you're dealing with a fuzzy-edged creature like a mammal, but one that is important if we hope to some day link the genes behind fruit-fly behaviors with the genes that may cause s
|Contact: Lori Oliwenstein|
California Institute of Technology