List of preliminary topics and sessions:
In general sessions will start with an introduction and survey of
the sub-area, then after this the main part of the session is made
up of technical talks on recent topics.
Recent work on geometry
Most of multi-view geometry and structure-from-motion algorithms has
been formulated for perspective (pinhole) cameras. Other camera types
are interesting for different applications, e.g. surveillance or
navigation: omni directional catadioptric cameras (cameras looking at
curved mirrors), fisheye cameras, and more exotic or specific devices
such as pushbroom or x-slit cameras, etc. A basic unified imaging
model for all these technologies consists in describing a camera by a
lookup-table between its pixels and their lines of sight in 3D. In
this session we address recent works on how such a general imaging
model can be calibrated, how generic structure-from-motion methods can
be devised and how multi-view geometry can be formulated for it.
The session will close with a discussion of possibilities and issues
concerning practical applications.
Variational, PDE-based methods in Vision
Recently a variety of computer vision and image analysis problems
have been studied using variational methods and PDE's, especially
surface reconstruction and segmentation. This allows separate study of
the problem and possible solution methods (unlike when algorithms
often were directly formulated for a specific task), and opens up the
vast field of knowledge and methods for differential equations to
computer vision applications. This session will cover both basic
principles and their application to a variety of currently studied
problems including edge and boundary extraction, segmentation, stereo,
shape reconstruction algorithms and surface fitting with applications
in general computer vision, medical image analysis and computer
graphics. Solution methods will range from traditional to more recent
level set methods. The session will also cover recent developments in
variational methods and general PDE-methods.
Morning: Tutorial and basic concepts.
Afternoon: Selected recent works.
Machine Learning and Human Motion Tracking
As computer vision systems are developed for increasingly complex
tasks and environments it becomes difficult to a-priori program all
details. Methods from Machine Learning has been merged into Computer
Vision on a variety of levels. Learning can be applied directly in
vision algorithms to e.g. find patterns and their variation. On a mid
level, a vision algorithm can be trained to particular applications by
parameter learning, and finally learning can be used on a system level
to guide during run time which submethods/procedures should be applied
to particular data. In this session we study recent learning methods
in conjunction with a variety of vision problems, such as
segmentation, object recognition, tracking, hand-eye coordination, as
well as integration of learning to guide execution in systems.
Applications, 3D models from 2D video, Virtual Heritage
Recently video processing and graphics rendering has become possible
on standard consumer PC's. In collaboration with artists from the
Banff New Media Institute participants will explore the creative use
of capture, modeling and rendering. A laboratory setup will be
provided with several computers and cameras for the capture of
geometry and appearance using shape-from-silhouette and
structure-and-motion methods. Captured 3D geometry and appearance
models are integrated and processed in Maya, and can be rendered
either into movie animations, or used interactively with our real time
As an appealing application we will focus on the capture of natural
and cultural heritage into 3D digital models, so called Virtual
Heritage. Vision is an appealing sensing modality. It is passive,
avoiding damage to sensitive artifacts, and it scales well and can
capture from small objects to whole landscapes. Recently, numerous
practical heritage capture projects have been carried out. This
session will expose and contrast experiences with different methods
applied to real world capture and modeling problems.
Participants are also be invited to bring their own capture
projects, which can be presented as talks, demos or posters.