Research Projects
Visual Computing


The three pillars of Visual Computing are

The first benefits from techniques and results of image analysis and synthesis. This page lists results, both current and of the last years, from research projects in the field of Visual Computing.

Since 2005 the project "Dynamic Learning for Geometrical und Graphical Object Acquisition" is funded by the German Research Foundation (DFG). The following of the listed projects have been worked on within this framework:

Below you can download two large image databases of densely sampled object views.



Computer Vision /
Human-Computer Interaction


Autonomous Systems / Dynamic Learning

Dynamic Learning for Object Acquisition
subm.: SPPRA 2010
EANN 2009
Trans. Comp. Intell. Sys. 2(2), 2008
ECAI 2008
BMVC 2008
Int. J. ISTA 2(2/3), 2007
ICVS 2006
ALaRT 2005



SCOUT - SCanning OUTstandingly
ARCS 2005





Object Recognition

Learning Sparse Object Representations
ESANN 2002
ATR 1999



Pose Estimation
Machine Vision and Applications 16(1), 2004
ICVS 2003
Image and Vision Computing 20(5-6), 2002
Image and Vision Computing 20(4), 2002
BMVC 1999



Object Recognition from Noisy Images
SIP 2003





Image Analysis


Tracking / Video Analysis

Adaptive Object Tracking in Dynamic Environments
Trans. Comp. Intell. Sys. 1(1), 2009

An object tracking system is introduced which is applied to interacting persons in seminar rooms. Persons are allowed to move freely within a group of other persons and thus can be occluded. A flexible way of representing a moving person allows for real-time tracking in difficult situations such as sudden camera movements or a dynamically changing background. In addition, an interactive selection of new target persons is possible.



Segmentation

Image Segmentation Based on Height Maps
CAIP 2007

We introduce a new method for image segmentation. It is based on a height map generated from the input image, which characterizes the image content in such a way that the application of the watershed concept provides a proper segmentation of the image. This method is able to provide better segmentation results on difficult images such as images of natural objects, than without the intermediate height map generation.



Low Level Image Analysis

Similarity Measure for High-Dimensional Feature Spaces
ICANN 2007

Feature matching in high dimensional feature spaces is a commonly used technique for object recognition. One major problem is to find an adequate similarity measure for the particular feature space. We propose a method to obtain such a similarity measure. Though it is based on the underlying topology of the feature space, no knowledge about it is needed a priorily, as the topology is generated by a growing neural gas.

Banana Wavelets - Filters for Curved Features
Computer Vision and Image Understanding 77(1), 2000
ARVO 1998
ESANN 1997
STIPR 1997



Feature Point Detection in Multiframe Images
CPRW 2000
Pattern Recognition Letters 20(2), 1999
IVCNZ 1998





Image Synthesis


Visualization

Aesthetic Primitives of Images
subm.: Int. J. Social Robotics, 2009
IV 2007

Images play an important role in visualization. As users are more willing to adopt a product if it evokes pleasurable feelings the aesthetic appeal of images is important. We derive six dimensions of visual aesthetics from the modularities of the human visual system and explore, inspired by principles of the visual arts and cognitive neuroscience, which pecularities of these dimensions are adequate for an aesthetic impression.

Squaring the Circle in Panoramas
ICCV 2005

Pictures taken by a rotating camera cover the viewing sphere surrounding the center of rotation. In order to obtain a flat panorama the spherical image has to be projected onto a picture plane. We show that in the same panorama multiple projections may coexist which can be chosen locally depending on the image content. Such multi-view projections can produce more compelling results than one single global projection.



Computational Geometry -
From Images to Models

Freehand Acquisition of Objects
subm.: J. Comm. Comp.
INTERACT 2009
WIO 2006

We introduce a tool which allows an untrained user to take three images of an object freehand with a simple consumer camera. From these images a 3d model of the visible parts of the object is reconstructed. We propose solutions for three weaknesses of the state-of-the-art reconstruction pipeline: an improved SIFT-based feature detection, a two-stage RANSAC process facilitating a faster selection of relevant object points, and a novel texture mapping.

Bundle Adjustment Joined by Evolutionary Algorithms
VIE 2006

Bundle adjustment is used to improve results of multiple view reconstruction to obtain 3d structure from images. It minimizes the reconstruction error by numerical methods, but there is no guarantee that the detected minimum is a global one. This disadvantage can be diminished by evolutionary algorithms. We apply a combination of bundle adjustment with an evolutionary algorithm to the 3d reconstruction of objects.

Image Rectification -
Catching the Infinity Cases
ICINCO 2007

Image rectification is used to reconstruct 3d structure from two images. The common algorithm fails if one epipole lies inside an image. Another scheme known as polar rectification fails if one epipole lies at infinity. But autonomous systems can produce image pairs with such difficult positions of epipoles. We present an extension to polar rectification which allows for image rectification with epipoles even at infinity and inside an image.



View Reconstruction

Interpolation of Novel Object Views
BMVC 2001
ENAIS 2001





Computational Photography / Digital Art

Dark Days - New York
SIGGRAPH 2008
IV 2007
CGIV 2007

The topic of Dark Days is the isolation of the modern human beeing in today's big cities. The reality of the city inhabitants undergoes an estrangement by coarse grain and blur. The walking humans at crossroads in New York are not visible in the negatives with the naked eye. Only an extreme enlargement makes them visible. A source of inspiration have been the Swiss sculptor Alberto Giacometti and the Russian filmmaker Andrej Tarkowskij.

Dark Days - Prague
IV 2008
CGIV 2008

The topic of Dark Days is the isolation of the modern human beeing in today's big cities. The reality of the city inhabitants undergoes an estrangement by coarse grain and blur. The walking humans at streets in Prague are not visible in the negatives with the naked eye. Only an extreme enlargement makes them visible. A source of inspiration have been the Swiss sculptor Alberto Giacometti and the Russian filmmaker Andrej Tarkowskij.



Downloads

download image databases of densely sampled object views

For two objects 2500 grayscale images of size 128x128 in tiff-format can be downloaded, which cover the upper half of the viewing sphere.



back to Gabi's homepage

last update: 2009-10-15