The Johns Hopkins University

Whiting School of Engineering

Department of Electrical and Computer Engineering

Categorizing Video Sequences of Non-Rigid Dynamical Objects

Seminar By

Avinash Ravichandran

Graduate Research Assistant

Electrical and Computer

Abstract:

Non-rigid dynamical objects are common in our day to day life. Some examples include waves rippling on the surface of a lake, a flag fluttering in the wind etc. These objects constantly change their appearance and simple assumptions such as the geometry being fixed or the object being Lambertian are no longer true. This makes solving existing vision problems for such objects very challenging. Of particular interest to us, is the problem of categorizing video sequences of such objects.

Categorization of objects plays an important part in the visual understanding of our environment, and it deals with identification of regions of the image/video that correspond to objects of interest. A vast literature exists for rigid object categorization, but little has been done for non-rigid dynamical objects. Existing work has been able to model non-rigid objects as the output of a linear dynamical system, enabling a compact representation of these video sequences. Using such a representation, there have been preliminary attempts to categorize these video sequences, by borrowing metrics from the control systems community. However, none of the existing work handles any invariance to properties such as scale, illumination or camera pose between the test and training datasets. These classes of transformations are of particular importance to most vision based categorization applications.

In this work we propose to categorize video sequences of such objects also known as dynamic textures with a method that is invariant to changes in scale, illumination and camera pose. In addition, we would like to categorize dynamic textures when there is clutter present in the sequence. Our approach relies on exploiting properties of the model parameters and combining them with existing concepts from object categorization of rigid bodies.

We will present in the talk, preliminary results on extracting appearance information i.e. the camera pose from video sequences of dynamic textures. We will show how to exploit the model parameters to extract the appearance information. This will serve as the preliminary step on which we can build our categorization framework.

Thursday, December 6, 2007

4:00 p.m.

Barton 117

Refreshments will be served at 3:45 p.m.

FOR DISABILITY INFORMATION

CONTACT: Candace Abel (410) 516-7031 cabel@jhu.edu