Chatroulette’s Pervert Detector is Broken: Here’s How to Fix It

Chatroulette is back! Unfortunately, the intriguing, often scandalous and much-hyped video-chat site that pairs visitors with random conversation partners from anywhere in the world is also still horribly broken.

According to one correspondent who braved the re-launched site, it took four full minutes of hitting the “next” button to find a chatter who wasn’t showing his bits to the camera. Ahem.

It didn’t have to be this way: Andrey Ternovskiy, the Russian wunderkind who started Chatroulette, has been courted by pretty much everyone in Silicon Valley, apparently, and promised in early June to re-launch the site with some kind of automatic image filter that would keep the pervs off the site–“software that can quickly scan video to determine if a penis is being shown,” according to TechCrunch.

What’s baffling is not that the world is full of anonymous men ready to drop their pants–it’s that Ternovskiy appears not to have implemented any of the broad array of technologies available to him that could prevent these men from showing up in your chat queue.

Indeed, a 2009 paper from a handful of researchers in Germany, Detecting Pornographic Video Content by Combining Image Features with Motion Information, outlines the most popular ways to automatically detect inappropriate content, many of which are mature technologies that have been around for years.

Skin Detection (Keyframe Analysis)

This is a no-brainer: people who are having sex on camera are not wearing clothes. “Simple features including the average skin probability, the ratio of skin pixels, and the size of the largest skin region are used for classification,” says Jansohn et al.

This technique doesn’t even require that video itself be analyzed - just a random selection of keyframes (still images) extracted from the feed.

Bag-of-visual-words

A pre-programmed vocabulary of “visual words” is trained into the system - they’re something like characteristic shapes, patches of color, etc. - and then the system evaluates the frequency with which these visual words show up. Apparently, when you plot the frequency of visual words, they cluster distinctively in pornographic material.

Like Skin Detection, this technique is also applied just to keyframes. So, hey, it’s not even that computationally intensive!

Periodicity Detection (Motion Analysis)

“It seems reasonable to assume that sex scenes in video can be characterized by a periodic motion pattern,” observe Jansohn and colleagues.

“Figure 1: Periodicity detection (PERWIN): a video scene with its mean motion signal in x-direction (top) and the classifier score (bottom). In the beginning of the video scene (left), clothes are taken off. Later, during sexual intercourse, periodic motion occurs, and scores indicate a higher probability for pornography.”

Google Does It, so Why Can’t Ternovskiy?

Skin detection is one of the primary methods of automatic filtering behind Google’s SafeSearch feature, as outlined in this 2006 paper (pdf) by Google researchers on the subject. Not everyone has the world’s largest data centers running the world’s cleverest parallelization algorithms, but surely, if Ternovskiy and his backers were serious about turning Chatroulette into something other than the butt of jokes on Southpark, they’d implement at least the most basic algorithms for detecting nudity.

Granted, key-frame analysis won’t always cut it in the anything-goes world of Chatroulette, where users can disrobe almost as fast as you can click the “next” button, so periodicity detection might also be required.