From:                     Kendra Smith

Sent:                      Thursday, April 06, 2000 7:46 PM

To:                         M?crosöft Research Tech Talk, Sem. Notice

Cc:                         Kendra Smith

Subject:                 UW-CSE Colloq / 4-25-00 / Leung / UC-Berkeley / Object Recognition with Material and Shape

UW-CSE Colloq / 4-25-00 / Leung / UC-Berkeley / Object Recognition with Material and Shape

 

*NOTE* This lecture will be broadcast live via the Internet. See

http://www.cs.washington.edu/news/colloq.info.html for more information.

 

UNIVERSITY OF WASHINGTON

Seattle, Washington 98195

 

Department of Computer Science and Engineering

Box 352350

(206) 543-1695

 

COLLOQUIUM

 

SPEAKER:      Thomas Leung, UC-Berkeley

 

TITLE:          Object Recognition with Material and Shape

 

DATE:           Tuesday, April 25, 2000

 

TIME:           3:30 pm

 

PLACE:                    134 Sieg Hall

 

HOST:           Linda Shapiro

 

ABSTRACT:

 

One of the fundamental problems in computer vision is recognition. The two

major sources of information for object recognition are material and

shape.  Material, commonly referred to as texture, is about what an object

is made from.  Shape, or geometry, is about what form an object takes.

 

I will first present studies aimed at understanding different aspects of

texture.  By modeling texture as repeating 2D patterns, the problems of

detection, grouping, and surface shape recovery are solved.  A technique

is also derived which can synthesize a wide range of real-world

textures.  However, a lot of natural texture is not only due to 2D

patterns but also 3D surface height variation.  I will present two models

which take into account the surface relief.  The first model makes simple

assumptions about the surface property to derive intuitive analytical

expressions about visibility and shading. The second approach is a

learning framework which acquires a texture model through a collection of

images.  The basic intution is that texture can be modeled by a universal

vocabulary of prototypes, called 3D textons.  The 3D textons are similar

to phonemes in speech processing or alphabets in English.  Using the

learned 3D texton vocabulary, useful texture recognition and synthesis

results are demonstrated.

 

I will also present a model for object shape.  Objects are defined as a

collection of features.  The configuration of these features specify the

shape of the object.  Shape classes arise when the configuration changes

from instance to instance.  For example, the locations of the facial

features (eyes, nose, and mouth) can vary with the expression (laughing or

crying) or identity of the person.  I will present a method which captures

such variations probabilistically.  The algorithm is demonstrated on the

problem of face detection from photographs.

 

Refreshments to follow.

 

Email: talk-info@cs.washington.edu

Info: http://www.cs.washington.edu