Research Projects
These are some of the projects I have worked on recently. For published results, the thumbnail images link to a detailed project description page including links to papers, posters, presentations, videos, and more.
3-D Deformable Vehicle Model
3-d shape recovery from passive imagery is a difficult
problem. Purely bottom-up approaches usually encounter
far to many ambiguities to produce accurate and dense
3-d surface data. Working top-down it is possible to
match specific 3-d models to images, but this process
is not useful for computing the shape of unknown objects.
This project finds middle ground by fitting a deformable
shape model to images. The model is specific to a class
of objects (vehicles in this case) and contains information
about the prior probablity of each shape.
Inexpensive Handheld Laser Scanning
Scanning the shape of objects in 3-d typically
requires expensive scanning systems that are
difficult to use and require constant recalibration.
The goal of this project is to build a simple
hand-held scanning system that addresses these
problems. Our design uses inexpensive off-the-shelf
components (a laser pointer, a pair of cameras, and
a glass rod) to capture highly accurate 3-d point
clouds. We use a new planar light constraint that
offers improvements over previous work in this area.
Read more ...
Background and Shadow Models
This project is about building video background models.
The code is based on my C++ statistical modeling
generic library (i.e. templates in the style of STL).
Standard background modeling usually can not separate
a moving foreground object from its moving cast shadow.
This work explores using 3-d models to predict the shadow
and help learn appearance based shadow and background
models. The working example is video of traffic
surveillance.
Read more ...
Plausible Physics in Augmented Images
This project combines plausible physics simulation and
augmented reality. Starting with a series of images of
a static scene taken from multiple unknown viewpoints,
the system automatically reconstruct an augmented world.
This allows users to virtually toss objects into the world.
The virtual objects interact with each other and with
real world objects according to plausibly realistic
phyics simulations. Virtual objects also cast shadows
onto, and may be occluded by, the real scene.
Read more ...
Epipolar Curve Tracking in 3-D
A translating object that does not undergo rotation
with respect to a camera induces a strong
constraint on tracking. All 3-d points on such an
object with diverge from, or converge to, a single
image vanishing point. This point is the common epipole
of any pair of frames. Given this constraint, it is
possible to overcome the aperature problem and uniquely
correspond points along curves. This project applies
the problem to tracking and 3-d reconstruction of
curves on vehicles.
Read more ...