PornHub introduces a machine vision system for automatic recognition of faces, poses and other video attributes.

The classification of the image on the grounds and recognition of the actress on the face of the frame from the video. Illustration: PornHub

Porn industry has always been the engine of technological progress. It continues to remain so now. For example, one of the most visited sites in the world PornHub (daily audience of 80 million people) is preparing to introduce a machine vision system, writes TechCrunch . The system automatically processes and distributes millions of videos into rubrics.

For easy navigation through the site, all videos are classified according to various criteria, including the name of the actress, the type of storyline, the subject of the video, and so on. Previously, all the work on classification was performed by people - moderators. But the number of amateur and professional videos is rapidly increasing. And now PornHub, instead of hiring new moderators and paying them to watch NSFW, launches a machine vision system that will work around the clock and does not require a salary.

“Ultimately, we want to give our fans everything that they prefer to see, and our new model will be able to produce more accurate results. We hope that thanks to this they will come back to see even more, - said Corey Price, vice president of PornHub. “We are talking about the constant updating of our platform, providing our fans with access to the latest technologies.”

Recognition of the position of the actress (posture) in the frame. Illustration: PornHub

The computer vision system can recognize specific actors and actresses in the frame, as well as their positions and other attributes. The video is tagged with relevant tags and is available for easy searching. As shown in the demonstration, the system works literally in real time. It is possible that the recognition and tagging will be carried out directly during the download of content on the site. When demonstrating a new system, the identity of one actor was recognized even when shooting from the side.

So far, the model has been trained on 50,000 thematic videos, including amateur ones. “We plan to scan the entire library in early 2018,” said Price. “In short, the technology will recognize various sex positions and categories and will be able to affix appropriate tags.” The entire PornHub library contains 5 million videos.

Machine Vision Model at PornHub

Recognizing poses in sex is a relatively new application of the machine vision system. It seems that scientific papers on this topic have not yet emerged, so PornHub is in a certain sense a pioneer in this area of ​​scientific and technological progress. But in the recognition of faces is nothing special: such systems have existed for a long time. Here, the work is facilitated by the fact that the number of actresses is limited (porn actresses on PornHub are more than 10,000), so the system should work more accurately than the recognition of arbitrary people on the metro via FindFace and other similar services. Cory Price says that the system works fairly accurately.

Among other things, the system of machine vision will help fight video piracy. Pirates often steal the intellectual property of PornHub and place it on other sites like YouPorn, xvideos, xHamster, etc. Now you can implement a system like ContentID that runs on YouTube.

Machine Vision Model at PornHub

You can imagine what new features might appear on PornHub in the future, when their machine vision system becomes more advanced. Perhaps the site will be able to upload an arbitrary photo of any familiar or unfamiliar girl - and PornHub will find a video with an actress who most closely resembles her. In principle, such functionality can be launched right now, technologies quite allow. Moreover, the Megacams video chat site already provides this service . In principle, you can make a browser plugin so that on the pages of the girls on VKontakte or on Facebook there are links to relevant videos right away.

And vice versa - under the group video immediately put down links to relevant profiles of people on Facebook, Twitter, LinkedIn, etc., where they are already under real names.

Perhaps, as you learn, the system will learn how to make text annotations for the plots. Who knows, all of a sudden, to save time, it’s more convenient for someone not to watch the entire movie, but simply to read the description of further events in text form.

Well, given the fact that the system now accurately recognizes the postures on video, we expect new interesting statistics from PornHub for each country and even for each city from where an amateur video is being downloaded.

Maybe in the future, AI will learn to compile its own videos from fragments of different videos, using the high rating of the audience as a positive incentive when teaching the neural network. There are many ideas on how to apply machine vision in this area.


All Articles