CHICAGO — Howa machine recognizes a man from a treesounds like pretty much a mundane,theoretical question. How a machinedifferentiates an enemy combatant from a”friend” or any other object on thebattlefield, however, is definitely the sortof question the Defense Advanced ResearchProjects Agency (DARPA) should be asking.Similarly, how a car identifies a pedestrianon the street — under any and all weatherconditions — is a significant challenge tothe development of advanced driverassistance systems.
Despite recent advances in the embeddedvision area, vision algorithms still are anexperimental field, studded with stubbornproblems. How to weed through a growingnumber of vision algorithms, to best aid thetime-consuming task of generating testobjects necessary for better visionalgorithms, has been DARPA's focus in recentyears.
Under a program called Visual MediaReasoning (VMR), DARPA contracted twoprivate companies — SRI International andNext Century Corporation — and completedthe development of two general-purposevision system development tools. One offersthe automated evaluation of vision algorithmperformance over a massive parameter space.The other enables generation of syntheticimage content for use in training andtesting detection and recognitionalgorithms.
Mike Geertsen, DARPA program manager, willpresent an overview of the VMR program andthese enabling tools at the Embedded Vision Summitscheduled for Oct. 2 in Boston.
Jeff Bier, a founder of the Embedded Vision Alliance,explained that the most remarkable thingabout the latest development is that “thesetools will be released as an adjunct to theOpenCV open-source computer vision softwarelibrary in late 2013 or early 2014.”
He quipped, “You rarely hear 'DARPA' and'open-source' in the same sentence.”
Although there are many shades of”open-source,” Bier told me that DARPA'sidea is to make these general-purpose toolspublicly available in a source code form.
DARPA's open-source move underscores thereality that computer vision is still adeveloping field, with lots of people tryingdifferent ideas and technologies.
If you are in search of the best visionalgorithms, your challenge is not onlyassessing “a whole bunch of differentalgorithms,” but also looking into eachalgorithm featured with “10 different knobs”to turn, Bier explained. In other words,”You could end up running through 100algorithms with 1,000 parameter settings.”
True, computer vision is no longer anacademic theory. There has been asignificant growth in vision algorithms,application developers, and theircommunities. But that very growth hasresulted in “a cluttered landscape ofalgorithms,” according to Bier.
Under DARPA's VMR program, SRI hasdeveloped automated performancecharacterization tools, providing anefficient means of assessing the performanceof different algorithms across imagingdomains. The tools identify how well analgorithm will perform for a given image ata chosen parameter setting, and theparameters for a particular algorithm andimage.
Meanwhile, Next Century Corp., under theVMR program, has developed tools to generatesynthetic images that can be used for thedevelopment of vision algorithms.
Why synthetic images?
Noting that 3D synthetic images used invideo games these days are very realistic,Bier explains that until now, visionalgorithm developers had to collect vastnumbers of images that required manualannotation. In order to develop visionalgorithms, such annotated data has to coverthe range of object variations, pose, andenvironment situations, so that a computervision system can perform successfully inoperational situations.
“By generating these images synthetically– where we know a priori what theyare, we can supplement the vast data neededfor vision algorithm development.”
DARPA going open-source with these visiontools is certainly a novelty to a lot ofpeople. More important, there will be a lotof eager developers lining up to get theirhands on them, Bier predicts.
This blog was also published on EETimes.