Ehemaliger Mitarbeiter des Institutes
Professur für Photogrammetrie
Nussallee 15
53115 Bonn


Wissenschaftlicher Mitarbeiter

am Institut tätig

von 2006 bis 2010


Neue Seite

Diese Seite wird seit September 2010 nicht weiter gepflegt. Bitte nutzen Sie meine neue Seite am Forschungszentrum Jülich.



I am mainly working on image feature detection and matching under difficult conditions. Image feature matching is the basis for many computer vision applictions, especially for automatic orientation of unordered sets of images, image registration and object recognition. Especially in case of poor texture or repeated structures in the images, state of the art procedures become unreliable due to a very low amount of features and due to shortcomings in the mostly heuristic feature matching approaches. My goal is to overcome both problems and thereby improve the results of such computer vision applications.

Completeness of Coding with Image Features

I found that it is crucial to combine different feature detectors to get as much information as possible. Good results can be achieved with two or three detectors only in case they are highly complementary. Therefore we developed a measurement scheme for the completeness of a set of detectors with respect to the information contained in an image, which gives direct insight into the complementarity of detectors. It has been published in an upcoming issue of the International Journal of Computer Vision (IJCV) and in a shorter version on BMVC'09.

The completeness is expressed by the distance between two distributions over the image (left): A reference, represented by local entropy over scales (middle), and a feature coding density, represented by a Gaussian mixture distribution based on sets of features (right). My presentation is available as a video on

Scale Invariant Junction Keypoints

The completeness measure showed that one important detector has been missing so far: A Scale Invariant Junction or Corner Detector. We developed a framework for finding such points, being especially distinctive, interpretable and accurate. The detector is called SFOP, you find more information and a matlab implementation on the project website.

Incorporating 2D geometric consistency into feature-based matching

When using multiple feature types, the matching procedure has to be modelled with care in order to take the different properties of features into account. For example, matching line segments based on descriptors is much weaker than matching affine regions. It even has different constraints, as one line segment may well match multiple segments in other images. Furthermore, in case of poor texture and repeated structure, the similarity of feature descriptors alone is hardly reliable. In such cases, the mutual spatial layout of features is a stable queue for rigid scenes. I am therefore developing a framework which integrates descriptor-based matching with consistency of spatial layout for different feature types and feature descriptors. The framework is based on a Markov Random Field formulation, utilizing a sound statistical derivation of the underlying energies, and solved using an LP relaxation scheme. Spatial constraints are modelled using uncertain oriented projective geometry. As a result, I obtain a robust and fast procedure for matching different features types simultaneously under difficult conditions, which consistently avoids heuristic thresholds and data-driven parameter tuning.

Other interests

Apart from my main work, my interests include

  • Markerless Tracking and Augmented Reality
  • Software development (esp. C++, Python, Java) and software engineering techniques
  • Unix and Mac OSX
  • Jazz Music


The following datasets show indoor scenes with sparse texture. They have been used in some of my publications and my thesis Robust Wide-Baseline Stereo Matching for Sparsely Textured Scenes. Click on the thumbnail images for downloading the datasets.
Blank-12. Very sparse texture, multi-planar 3D structure, approx. 90 percent image overlap. 12 images at a resolution of 1203x800 pel.
Blank-22. Very sparse texture, multi-planar 3D structure, approx. 90 percent image overlap. 22 images at a resolution of 752x500 pel.
Class. Moderately sparse texture, multi-planar 3D structure, fisheye lens, approx.~60 percent image overlap, 8 images at a resolution of 752x500 pel.

Curriculum Vitae



Here is a list of theses that I supervised.

  • Miriam zur Mühlen & Birgit Abendroth
    Genauigkeitsbeurteilung und Untersuchungen der Zuverlässigkeit von optischen Onlinemessungen
    Diplomarbeit in Zusammenarbeit mit aicon 3D systems.
    (Accuracy Assessment and Investigations on the Reliability of Optical Online Measurements, a joint work with  aicon 3D systems)
    Diploma thesis
  • Rebekka Schultz:
    Orientierung einer Kamera in einer Legolandszene
    (Orientation of a camera in man-made scenes)
    Bachelor thesis
  • Marko Pilger:
    Stabile skalierungsinvariante Fenster zur Bestimmung quadratgradientenbasierter Punktmerkmale
    (Stable scale invariant windows for locating square-gradient based point features)
    Diploma thesis
  • Stephan Steneberg:
    Robuste Relative Orientierung kalibrierter Kameras mit Bildkanten
    (Robust relative orientation of calibrated cameras using edge features)
    Diploma thesis in collaboration with the Working Group Active Vision, University of Koblenz


  • Übung zur Vorlesung "3D Koordinatensysteme", WS2009/2010
    Lab assistant, "3D Coordinate Systems", winter term 2009
  • Übung zur Vorlesung "Projektive Geometrie und Bildfolgenanalyse", SS2009
    Lab assistant, "Projective Geometry and Analysis of Image Sequences", summer term 2009
  • Übung zur Vorlesung "Photogrammetrie II", WS2008/2009
    Lab assistant, "Photogrammetry II", winter term 2008
  • Übung zur Vorlesung "Photogrammetrie I", SS2008
    Lab assistant, "Photogrammetry I", summer term 2008
  • Übung zur Vorlesung "Objekterfassung aus Punktwolken", WS2007/2008
    Lab assistant, "Object registration from point clouds", winter term 2007/2008
  • Übung zur Vorlesung "Objekterfassung aus Punktwolken", SS2007
    Lab assistant, "Object registration from point clouds", summer term 2007

Other resources

My personal website at contains more information about computer vision research, programming, and my family.