Each week we’re using this blog to profile some of the companies that participated in the 2009 Emerging Companies Summit, starting with the five winners of our “Ones to Watch” award. You can learn more about innovative companies that use NVIDIA’s GPU technology in the GPU Ventures Zone.

Anyone who’s spent time on sites with a lot of user-generated content (UGC) has experienced the incredibly random nature of most social-media advertising. You’re looking at a photo of a car, the ad you get is … liposuction? The problem: Most visual content is untagged (or poorly tagged), and advertisers often don’t have a clue what they’re appearing next to. So either they don’t advertise on social media at all – or they do and cross their fingers they don’t end up next to something inappropriate or ill-suited to their brand.

New York-based Milabra has created a visual-recognition technology that solves the problem of disjointed advertising and UGC, creating value for visual content. It uses NVIDIA’s Tesla parallel computing power for machine-vision analysis that quickly and accurately identifies what’s in an image. It then is able to serve up the most appropriate ad for that image.

Looking at a photo of a red car? Milabra can serve up an automotive ad with a red car in it. Even better, the car in the ad can match compositionally with the photo you’re looking at. When the content and ad message make sense, conversion rates go up – which is exactly what Milabra has been able to do for advertisers*.

“We’re like AdSense for images,” says Milabra CEO Sam Cox, who met recently for a video chat with NVIDIA VP of Business Development Jeff Herbst to talk about Milabra’s technology (scroll to the bottom of the hyperlinked web page for the video link).  The key is to get around what Cox calls “banner blindness” by making online ads more effective and enable publishers to monetize visual content.

Monetizing visual content is a grail that a number of startups are chasing – with good reason. Although UGC is consuming more of our online time, Cox says visual media represents the most depressed online ad market. Cox points out that Facebook gets about a billion photos uploaded to it monthly which costs about $10 million to host. UGC may be wildly popular – it’s also expensive.

Milabra’s solution is simple. Publishers integrated with Milabra tag their pages with a simple bit of HTML. When a consumer uploads an image, that image gets captured for a multi-point visual analysis that looks at more than 150 categories to identify the subject in near real time. The analysis can identify people, places, things, but also more abstract categories – say, nightlife – and can do so with a fine-grained filter. (Watch Cox’s ECS demo for examples of how the analysis works.)

The data about the image is slotted into standard IAB (Interactive Advertising Bureau) categories, and is sent to Milabra’s ad server, where a decisioning logic ensures that the right ad gets served up next to it. Protocols that guard against adult content and brand misalignment are in place to ensure brand safety before any ads get served.

Online advertising depends on high-volume and low cost. To fit into this framework, Milabra’s system must be fast and accurate. And this is where NVIDIA’s GPUs and Tesla come in. Milabra uses the NVIDIATesla S1070 computing system to power its visual analysis. With Tesla, identifying more than 150 points about an image takes less than 200 milliseconds (and dropping). Using a CPU alone that analysis takes more than a second.

“There’s such a performance increase from using the GPU,” says Cox, “[the platform] wouldn’t be economical without it.”

For content publishers, social media may not be economical without Milabra.

* Cox’s ECS demo covered results for two advertisers. In one, a travel ad from an advertiser who’d never advertised on social media before, Milabra’s platform delivered 14 million impressions in 14 days and had a .44% conversion rate on people booking hotel rooms through the ad. In the case of an automotive ad, Milabra delivered a .48% clickthrough rate as compared to the expected rate of .02%.