Face Swapping 2.0 - Now Mostly Prettier

In July, I made a post regarding new technology from Columbia University that swapped faces in photos and videos with generic amalgems drived from a photo database. The technique, which had potential privacy benefits...and even certain Hollywood applications...suffered from one major drawback. The modified faces were really really really ugly.

See below (From Columbia University Research Paper):



Tough to stomach, I know, but these results were bound to be approved someday, and thankfully that day is here. The New York Times is reporting that computer scientists in Israel have taken a more beauty-centric approach to face swapping that simply replaces faces in images and video with more idealized representations of the same faces. Their software applies an algorithm covering over 200 different facial measurements that brings a target face closer to the ideal. In most cases, this approach goes well beyond simple airbrushing, dramatically altering a person's appearance and perceived identity.

See below (From Lars Klove for the NYT):



I am not sure either technology is perfectly suitable for privacy protection in video surveillance just yet, but given the choice between the two approaches above, for aesthetic reasons, I'd go with the latter.

Trace & Extract 3D Objects from Video

From Australian Centre fro Visual Technologies:

VideoTrace is a system for interactively generating realistic 3D models of objects from video—models that might be inserted into a video game, a simulation environment, or another video sequence. The user interacts with VideoTrace by tracing the shape of the object to be modelled over one or more frames of the video. By interpreting the sketch drawn by the user in light of 3D information obtained from computer vision techniques, a small number of simple 2D interactions can be used to generate a realistic 3D model. Each of the sketching operations in VideoTrace provides an intuitive and powerful means of modelling shape from video, and executes quickly enough to be used interactively. Immediate feedback allows the user to model rapidly those parts of the scene which are of interest and to the level of detail required. The combination of automated and manual reconstruction allows VideoTrace to model parts of the scene not visible, and to succeed in cases where purely automated approaches would fail.

2 Years Post Neven Deal...Google Launches Face Rec

Via TechCrunch:

In the anticipated release of Google’s new and improved Picasa, the company will offer facial recognition technology to help you identify friends and family in your pictures without requiring you to tag them by-hand each time you see them.

Launching at noon PDT today, Picasa’s facial recognition technology will ask you to identify people in your pictures that you haven’t tagged yet. Once you do and start uploading more pictures, Picasa starts suggesting tags for people based on the similarity between their face in the picture and the tags you already put in place for them.

The facial recognition technology comes to Picasa thanks to an acquisition Google made in 2006 of Neven Vision, a company that specialized in matching facial detail with images already found in a centralized database. Picasa’a facial recognition technology works in much the same way.

More from the Folks at Image Metrics

This video is impressive...and a bit creepy:

Amazing Video Enhancement Technology

Researchers at the University of Washington recently released a video showing a series jaw-drapping video enhancements now possible using various image analytic and modeling techniques. In one demo of interest, simply by mixing a few high-resolution photographs of a scene with a lower quality video stream, the scientists were able to dramatically enhance the quality and resolution of the complete video...by 4x.

The technology could also be used to seemlessly remove private content from surveillance video, truely making proctected images of people or objects invisible. That's probably welcome news to some, but disconcerting to law enforcement professials who already have significant concerns about the reliablity of photos presented them. Video has been generally thought of as much harder to manipulate...no more.

Computer Vision Resarch Goes Virtual

Sometimes the real world just isn’t real enough. That’s often the case in computer vision application development where super smart PhDs seek to create algorithms and technologies to track and classify people or objects within a video stream. Believe it or not, some of the same neural networks that catch bad guys today got their start by tracking frantic scientists running around their labs, offices, and dorm rooms.

But ObjectVideo thinks there is a better way…at least to start. Using technology from the videogame Half Life 2 , they have built a Virtual Video Tool that can be used to create “virtual surveillance” cameras.

The ObjectVideo Virtual Video (OVVV) Tool generates realistic video from simulated cameras in an interactive virtual world. This tool is free and is based on a modification (aka 'mod') of Half-Life 2, a commercially available game from Valve Software. Our hope in distributing this tool is to stimulate computer vision research in areas that cannot rely on canned video (eg. active tracking) or when large quantities of ground truthed video is unavailable or impractical (multi-camera installations, public spaces, the list goes on!).

The fact that virtual cameras are generally thought to lack the video noise and other artifacts found in real-world cameras, doesn’t prevent this tool from providing real benefits to students and researchers. Today gaming engines are so realistic and of such high quality that the line between real and virtual is being blurred. And, as OV points out, virtual cameras provide another benefit that’s impossible to achieve with real world footage: ground truth data that can be incorporated into the training process. Because virtual cameras are built on models of scenes where ever person and object and color and angle are actually known, a researcher always knows, without guess or estimation, just how well their computer vision algorithms are deciphering a particular video stream.

Beyond that, ObjectVideo has created most of the environments, models, and camera option necessary to test every conceivable surveillance variation during the testing process. Even blur, noise, and even lens and PTZ effects can be simulated with relative ease.

Virtual surveillance video is not just a great tool for computer vision researchers, it’s also an incredibly interesting area of research in itself. The folks at Valve Software have my appreciation for opening their platform enough to enable this kind of work. But believe it or not, Valve’s Half Life 2 is already almost 4 years old. Maybe ObjectVideo’s next endeavor can be a Crysis mod. That would be something. And next year there will be something else…even better.

Evansville Airport Gets Face Rec

From the Evansville Courier & Press:

Security is the paramount issue among air travelers these days, but Evansville Regional Airport — like other commercial airports throughout the world — rely on videocassette recorder security systems whose nonspecific tapes can take hours, days and sometimes weeks to decipher when suspicious concerns arise.

That, however, is about to change at the local airport.

The Evansville airport is the first in the nation to get a new technological security system, known as a 3VR (Third-Generation Video Recorder), said Bob Working, the airport manager. The searchable surveillance system uses a Google-like analytical search engine for spewing out valuable information instantly. For example, it displays in an instant on a computer monitor clear images of thousands of faces for identification purposes.

Eric Moss, vice president/director of data services for Gaither Technologies/STC, demonstrates a 3VR security system as company President Steve Rudolph is recorded on video, in the screen at left. The unit features searchable facial recognition and motion-changing events.