|
Image Correspondences
Modified Feature-Based
Matching Algorithm
The original feature-based
image-matching algorithm, FBA, proposed in [Weng93] is
extended and modified to improve its performance. Enhancements have been made
in improving the accuracy of the original algorithm and enable it to handle a
wide range of image resolutions and contrast. All attributes used for the
matching process are normalized and thus are handled in a uniform manner.
This facilitates the determination of their relative weightages.
Extensive experiments are
conducted to examine the impact of the various parameters on the performance
of the modified feature-based algorithm, MFBA. Although there is no ideal set
of values for all input image types, the exercise provides insight to the
effect of each of the parameters. Nevertheless, without a priori information
on the input images, a general set of parameters is proposed. This set can be
customized if knowledge on the input image is available. In general, the weightage for a certain attribute should be increased if
it is known that it dominates over others in the image. The MFBA is tested
with both synthetic and real gray scale images. The
results achieved are convincing. Similar convincing results are obtained even
when MFBA is tested with bi-level line art drawings.
Image Matching:
Porting to FPGA Platform
In order to determine motion
from time-varying image sequences, it is necessary to establish
correspondences between images. That is, to identify image elements in
different images that corresponds to the same element of the sensed scene.
This is one of the most challenging steps in motion and structure analysis.
Although it is possible to process many images simultaneously to establish
the respective correspondences, the underlying basic problem is to establish
the correspondences between two images. In the finite motion approach, the
images are treated as samples of scene taken at discrete times, and discrete
features such as edges, corners and regions are selected as tokens that are
to be matched. In this project, student will examine an existing hierarchical
image matching system. Program codes will be provided. The student is then
expected to develop a suitable VHDL model and port it to a FPGA platform in
order to expedite the matching process.
Shape Tracking and
Recovery using Model-Based Invariant
Establishing shape
correspondences between related images and recovery of planar objects in long
sequence are the subjects of this paper. Establishment of point correspondences
for large image motion is a demanding task requiring time-consuming
iterations. In addition, 3D objects are perceived very differently from
different viewpoints; as such, their recognition is beyond the ability of 2D
feature-based system. This issue is addressed by the use of 3D model-based
shape invariant. The shape of an object in an image sequence is tracked by
means of a hybrid image-based and a model-based shape invariant feature. The
proposed scheme also sports a corrective and self-diagnosis step whereby the
results obtained are refined and verified before they are used as sources for
subsequent tracking down the sequence. This is particularly important for a
long sequence as the effect of accumulated error will cause tracked points to
drift further and further away from the desired location in the later frames
as error propagates. We assume that the tracked object stays fully on the
view throughout the sequence. At the end of a successful tracking, the 3D
structure of the object tracked is recovered without additional computation.
A Generic Linux-based
Render Farm for Image/Video Processing and Computer Graphics Applications (Seah Hock Soon, Dilip Krishnan, and Zhang Jian
Feng)
Many of our research algorithms
are extremely computationally expensive. Simulations and experimentation
would greatly benefit from the development of a generic render-farm that is
not only flexible and open to further modification, but is also relatively
inexpensive. This project revolves around the research and development of
such a generic render farm, based on inexpensive, off-the-shelf Linux
processors and interconnecting components. The complete system would involve
the development of an Application Programming Interface (API) and a
Linux-based system that optimizes I/O and processing times based on user
inputs.
|