GPI Seminar Series: Ganesh Sundaramoorthi

Ganesh Sundaramoorthi (GMSV/KAUST), Deformable Shape Tracking by Region Matching,
Wednesday October 2nd, 12:00 - 14:00, Seminar Room D5-007

Deformable Shape Tracking by Region Matching

Abstract:
In this talk, I will discuss the problem of determining the precise shape of a deforming object from a video sequence. The problem is central to computer vision and remains a challenge due to the interactions of many different nuisances that generate infinite variations of the appearance and shape of the object in image. Classical solutions to determining the precise shape of the object have primarily built on the idea of discriminating the image into foreground and background based on simple image statistics, but these methods are severely inadequate for objects with complex appearance in cluttered backgrounds. An alternative approach, which we propose, is to treat the tracking problem as one of matching, i.e., where the current estimate of the object is matched to the next image. This allows one to tailor advanced optical flow and image matching methods, that have been only recently proposed and are able to cope with complex object appearance, to the deforming shape tracking problem. I will discuss how to tailor these algorithms to shape tracking. Further, in the context of tracking, one may revisit the design of advanced image matching (i.e., feature matching) algorithms, and design features that have better invariance/discriminability properties, and are therefore able to outperform current matching techniques. I will illustrate the primary ideas and illustrate results on challenging video, and show state-of-the-art performance (including out performance of Adobe's latest After Effects Software)