We have been working hard to design and implement a working,
interactive Augmented Reality (AR) system based on system setup which
allows a user to pan and tilt a flat panel monitor behind which a video
camera is mounted (see image on the right). Virtual objects are then
rendered into the live video stream in real-time respons to user
rotations, and taking into account real scene geometry for occlusions
and real scene illumination for rendering. We will demonstrate the
current state of the system, including how the system is geometrically
calibrated to a world coordinate system, how we record environment maps
to use for rendering reflective and refractive objects, and
also how we render shadows cast by the virtual objects onto the real
scene geometry.
There will be some kind of tasteful treat at the event ... a cake
or something like that, to celebrate the "inauguration" of the system.
Traditionally, many geometric transformations are carried
out using
matrix and vector algebra. Common examples of such transformations are
translations and rotations. While translations are both easy to grasp
and implement, rotations can be notoriously difficult to handle. The
causes of this lie both in the human mind and in mathematics. An
example of the mental difficulties associated with rotations occur in
situations where several coordinate systems, angles, and axes come into
play and things get mixed up. The mathematical issues, however, are
more serious since they cannot all be resolved using any of the
traditional techniques. Apart from explaining some of the common issues
with the traditional approaches, this colloquium will focus on a
mathematical tool, the quaternion, which - if applied correctly -
solves all of the problems associated with other rotation techniques.
The fundamentals of quaternions as well as algebra and applications to
rotation problems will be covered in this brief introduction.
A way to recognize
building and other rigid structures are with use of features, which is
something AAU working on with the IPCity project. 
The canopy structure of a plant holds valuable
information for assessment of its health status. Camera based 3D
Reconstruction of non-rigid biological objects such as plants is a
difficult problem. It is also difficult to test it because ground truth
is difficult to obtain. My work is based on ray-traced plant images
(for low level quantitative measures) and real images of individual
plants (for high level quantitative and qualitative evaluation). I will
present different approaches to estimating disparity maps that are good
enough to extract information of the canopy. Methods cover window based
correlation and energy minimizing methods. The disparity maps are
segmented into individual parts and each part is fitted a NURBS surface
so that the total model allows overlapping parts. Finally, I will show
some annotated reconstructions of real plants.
The health of a
vegetation canopy is closely coupled to its structure,
which is often represented by two statistical descriptors: the leaf
area index and the leaf angle distribution. These parameters are used
by biologists and agronomists and are commonly estimated by inverting
optical measurements with simplified canopy models. Ground-based
sensors are often used for multi-angular gap fraction measurements.
Factors to be taken into account in such measurements include mixed
pixels as well as mixed light scattering components, e.g., double
scattering and specular reflection. This study investigates the
potential of an alternative approach based on multi-angular
measurements of scattering components, e.g., sunlit and shaded
vegetation, sunlit and shaded soil, specular reflection, etc.. In
earlier work, it was shown that if the spectra of the illuminants and
the materials are known it is possible to separate multiple scattering
components using spectral unmixing. The idea in this method is to
couple the abundances of the scattering components to vegetation
structure parameters. The expected component abundances are predicted
using stochastic simulation in a Poisson canopy model with an
ellipsoidal leaf angle distribution which is specified by a single
shape parameter. The model predictions are matched against the observed
component abundances, and a numerical method is used to search for the
leaf area index and mean leaf angle that give the best match. The
method is evaluated against ground truth data using recursively
ray-traced images.