Research Interests 
 Visionbased HCI, visual tracking, gesture analysis
 Graphics, texture synthesis, multiresolution meshes and morphing
 Statistical learning, pattern recognition
 Learning human motion, realistic animation
 High dimensional data analysis
 Realtime visual tracking system

Research Projects 
3D ModelBased Hand Tracking 
 Modelbased approach
 Modeling articulate motion constraints
 Tracking articulated motion
 Tracking rigid motion
 HCI applications

ModelBased Approach
The modelbased approach for hand tracking uses a 3D hand model
to estimate the current hand configurations. The idea is to compare
image features produced by the model projection and that extracted
from real hand images. The hand state is recovered from the model
configuration that generates the best match. With a well initialized
hand model, this approach can produce a very accurate estimate.
In our experiments, we have tried a 2D patch model and a 3D cylinder
model both in 3D space. [More]

Learning Articulate Motion Constraints
For the case of articulated hand tracking, one of the main problems
assoicated with the modelbased approach is the high degrees of
freedom (DOF) involved in specifying a hand configuration. The hand
shape is descirbed by its joint angles which has roughly 21 parameters.
Therefore, to estimate the correct hand motion and configuration
is equivalent to a search in the high dimensional space. This is
an impossible task given the current computing technology. Fortunately,
natural hand motion is highly constrained, and previous work has
shown that by incorporating constraints obtained from biomedical
studies, the computational complexity can be greatly reduced. However,
there are many constraints that are hard to learn or represent in
a closed form; thus, we propose to learn the constraint model using
both semiparametric and nonparametric approaches. From our initial
empirical observations. we noticed that the trajectories of hand
motions when moving between basis states are roughly linear. (HUMO00)
[More]

Tracking Articulated Motion
The motivation for learning the motion constraints is to reduce
the DOF involved in tracking finger motion. However, learning the
compact parameterization of the constraint model is only half of
the work. To successfully track the finger motion, we must address
both of the two key issues:
 The representation of the feasible configuration space
 An efficient tracking algorithm associated with this space structure.
For the case of semiparametric representation, we used the linear
motion trajectory observed as the auxiliary importance function
and implemented the sequential importance sampling filter, which
can successfully track many common motions. For the case of nonparametric
configuration space representation, we proposed a stochastic NelderMead
simplex search algorithm which is a general tracking algorithm that
combines the topdown statistical approah and the bottomup direct
search approach. (ICCV01, FG04) [More]

Tracking Rigid Hand Motion
The complete hand motion is determined by its
 global motion parameters: {R.
t} in 3D space (6DOF)
 local articulation parameters: joint angles
(21 DOF)
Treating the hand as a rigid object, there are many algorithms
that can be used to recover the global motion parameters of an articulated
object. We have implemented both the statistical topdown approach
using stochastic Nelder Mead simplex search, and the bottomup direct
search approach using algorithms such as ICP. The global and local
parameters are estimated separately and then combined in an iterative
manner until the final estimation converges. For the case of tracking
specific gestures, we could also apply the learning approach to
reduce the problem complexity by learning a more compact parameterization
of the feasible space. (WMVC02, FG04) [More]

HCI Applications

Intern Projects  RealTime
Tracking Systems 
 Improved the performance of realtime AdaBoost face detector
with particle filtering
(Mitsubishi Electric Research Lab intern project, 2001)
 Implemented a realtime hand detecting and tracking system
(IBM T.J. Watson Research intern project, 2002)


Graphics Course Projects 
RealTime Rendering Systems 

ViewDependent Terrain
A system is built for realtime interactive viewing of a complex
terrtain data set. A viewdependent refinement scheme based on
Lindstrom's work is implemented for generating polygons to render.
The data used for this project is a Digital Elevation Map (DEM)
of the Grand Canyon [4097 x 2049] produced by the USGS. The source
for this dataset is the Georgia Tech
Large Models Archive.


Sphere Hierarchies and Splatting
For this project, a
QSplat like system is built. A hierarchical spatial data structure
is implemented and applied to the realtime splat rendering application.
Both axisaligned partition and oriented partition are explored.
Additionally, viewdependent refinement is implmented for the rendering.
This image shows the bounding spheres of each vertex at the most
refined level 

Video Textures
Based on the work of
Arno Schödl, et al , a new video sequence can be synthesized
from recorded video data base.
This 6 second sequence of neverendingpageflipping
action is generated from a data set of 20 frames.
( MPG, 885k
) 
Ray Tracing 

Ray tracer
A ray tracer is implemented that supports refraction, algebraic
surfaces, and constructive solid geometry.
Click here for a larger image 

Procedural shading
This scene is constructed by procedural shading, which uses various
combinations of the Perlin noise function.
Click here for a larger image 

Particle system
A simple simulation of particle system
( AVI, 638k
)
