At Embedded Technology 2010, Tohto C-Tech demonstrated real-time 3D face reproduction, using a GPGPU and stereo cameras.

The demonstration shows video shot with a stereo camera rig, with 3D replication achieved by a matching method based on phase correlation. The matching processing was done at high speed by using GPGPU technology, to enable 3D modeling of changes in facial expression in real time.

(Phil Lelyveld comment: this is similar to hacker work with the Kinect, as seen here.)

“If processing is done using just a CPU, it takes too long, so it can’t be achieved in real time. So we’ve utilized a GPU to increase the processing speed.”

A GPU is a broadly applicable device, with several hundred cores running in parallel and a wide memory bandwidth. POC matching processing is done in parallel as threads on the GPU. In this way, faces are reproduced in real time.

“With a stereo camera, two cameras take pictures simultaneously, and the pictures can be reproduced as 3D data, based on the discrepancy between them. 3D data can be obtained using a single stereo camera, but when a face is taken from the front, it isn’t possible to reproduce the sides. So we take pictures from two viewpoints, enabling us to reproduce faces all-round.”

“As a specific application, we think this system could be used to identify people’s faces. Changes in size and expression are difficult to detect from flat pictures, but by comparing 3D data, we think faces could be identified quickly.”

See the full story here: http://www.diginfo.tv/2010/12/13/10-0259-r-en.php