Contact: Remco Veltkamp
3D Shape Recognition
Although the human visual system is very good at the interpretation of
images, the amount of digitally acquired and stored data is so large that we
need computerized analysis and retrieval methods to handle it. Many
applications exist in industry, company archives, internet, and hospital
environments. Two examples are verification of misuse of a company logo among
one of the millions of images on the Web, and examination and retrieval of
the millions of digital medical scans made every year. Characteristic to our
research is that fundamental research is followed by experimental
verification. The approach we take is strongly algorithmic, and we put
emphasis on the analysis of shape of patterns that are present in images,
sound, video, and 3D scenes.
The topics we work on are feature extraction, segmentation, matching of
images, and organization of images in index structures for efficient
retrieval, with a focus on medical imaging and shape-based multimedia. Our
research in medical imaging is concerned with alignment of 3D scans (CT, MRI,
etc.). Our research in multimedia is concerned with the algorithmic aspects
of shape analysis. Some aspects involved are the representation,
decomposition, approximation, and deformation of shape, the transformation of
one shape into another, the measuring how similar two shapes are, and the
organization of shapes in (index) search structures.
The research focusses on the following topics:
For shape-based image matching, features are extracted from images, such as
edge and corner points, curves, and regions. The set of feature of one image
is then translated, rotated, and scaled so as to minimize some similarity
function with respect to the features from the other image. We develop
similarity functions that are robust against noise, occlusion, blurring, and
small deformations, as well as algorithms to compute them efficiently.
We are primarily interested in multimeda retrieval based on the certain
pattterns that are present in the content. For example patterns of shape in
images and video, or patterns of themes in music. Recognition and matching of
the shape of such patterns is often a computationally expensive task. For
large collections this makes it unrealistic to sequentially match all
multimedia objects in the database with the query. Therefore we investigate
indexing search structures that make it possible to use accurate but complex
similarity measures and computationaly expensive matching algorithms, while
the search is still efficient.
In reverse engineering, robot vision, machine part database retrieval and
many other applications, one of the primary tasks is to match and classify a
given 3D object shape. Laser scanning devices, web repositories, and private
industrial database provide a huge amount of 3D shapes in the form of point
clouds or polyhedral surfaces. We perform research on the matching two point
clouds, a point set with a polyhedral surface, and two polyhedral surfaces.
Two research issues are how to do partial matching, and how to achieve a
accurate match without an good initial estimate.
Our research in the area of medical imaging focuses on image processing and
image analysis techniques for the tasks of image segmentation and image
registration. Currently, specific projects involve the accurate segmentation
of the aortic lumen and aortic thrombus using a model-based approach, and the
improvement and extension of current medical image registration techniques
based on an information-theoretic approach.