Robotics and Intelligent Machines Center (RIM) Seminar-Sing Bing Kang

Add to Calendar
Date:
February 6, 2013 12:00 pm - 1:00 pm
Location:
Marcus Nanotechnology Building, Room 1116

Sing Bing Kang, principal researcher at Microsoft Research and adjunct assistant professor in the School of Interactive Computing

Unless otherwise noted, all seminars are held in room 1116 in the Marcus Nanotechnology Building from 12-1 p.m. Seminars are open to the public.

Sing Bing Kang, principal researcher at Microsoft Research and adjunct assistant professor in the School of Interactive Computing at Georgia Tech, presents "2D-to-3D Video Conversion" as part of the RIM Seminar Series.

Abstract
There is a renewed interest in stereo movies due to the general improvement in stereo films as well as technological advances in 3D display technologies. However, the stream of new stereo films is slow due to the additional overhead of using multiple cameras and the expertise required to use stereo in a compelling manner. It follows that a solution to this content problem (which is seen as inhibiting wider adoption of stereo technology) is to convert recent and legacy 2D films into stereo. In this talk, I will describe two approaches to convert regular video to stereoscopic video. The first approach is in the form of an interactive system we call “Depth Director.” Depth Director assists rotoscoping with temporally-coherent segmentation techniques, extracts depths using structure from motion, and uses what we call “depth templates” to enforce real-world 3D constraints into the system. Through Depth Director’s novel interactive user interface, the user is given as much or as little control as desired to overcome accuracy issues with automatic computer vision algorithms and to manipulate the relative “depth contrast” between objects. The second approach is automatic and makes use of non-parametric depth sampling. We use local motion cues to improve the inferred depth maps, while optical flow is used to ensure temporal depth consistency. For training and evaluation, we use a Kinect-based system to collect a large dataset containing stereoscopic videos with known depths.