Technische Berichte

We evaluate the performance of Import Vector Machines (IVM),a sparse Kernel Logistic Regression approach, for the classification of hyperspectral data. The IVM classifier is applied on two different data sets, using different number of training samples. The performance of IVM to Support Vector Machines (SVM) is compared in terms of accuracy and sparsity. Moreover, the impact of the training sample set on the accuracy and stability of IVM was investigated. The results underline that the IVM perform similar when compared to the popular SVM in terms of accuracy. Moreover, the number of import vectors from the IVM is significantly lower when compared to the number of support vectors from the SVM. Thus, the classification process of the IVM is faster. These findings are independent from the study site, the number of training samples and specific classes. Consequently, the proposed IVM approach is a promising classification method for hyperspectral imagery.

We propose a framework for object tracking in image sequences, following the concept of tracking-by-segmentation. The separation of object and background is achieved by a consecutive semantic superpixel segmentation of the images, yielding tight object boundaries. I.e., in the first image a model of the object’s characteristics is learned from an initial, incomplete annotation. This model is used to classify the superpixels of subsequent images to object and background employing graph-cut. We assume the object boundaries to be tight-fitting and the object motion within the image to be affine. To adapt the model to radiometric and geometric changes we utilize an incremental learner in a co-training scheme. We evaluate our tracking framework qualitatively and quantitatively on several image sequences.

@ARTICLE{yang*11:evaluation, author = {Yang, Michael Ying and Förstner, Wolfgang}, title = {Feature Evaluation for Building Facade Images - An Empirical Study}, journal = {Department of Photogrammetry, Institute of Geodesy and Geoinformation, University of Bonn, TR-IGG-P-2011-02}, year = {2011}, }

High-dimensional data structures occur in many fields of computer vision and machine learning. Transformation between two high-dimensional spaces usually involves the determination of a large amount of parameters and requires much labeled data to be given. There is much interest in reducing dimensionality if a lower-dimensional structure is underlying the data points. We present a procedure to enable the determination of a low-dimensional, projective transformation between two data sets, making use of state-of-the-art dimensional reduction algorithms. We evaluate multiple algorithms during several experiments with different objectives. We demonstrate the use of this procedure for applications like classification and assignments between two given data sets. Our procedure is semi-supervised due to the fact that all labeled and unlabeled points are used for the dimensionality reduction, but only few them have to be labeled. Using test data we evaluate the quantitative and qualitative performance of different algorithms with respect to the classification and assignment task. We show that with these algorithms and our transformation approach high-dimensional data sets can be related to each other. Finally we can use this procedure to match real world facial images with cartoon images from Springfield, home town of the famous Simpsons.

We describe ground truth data that we provide to serve as a basis for evaluation and comparison of supervised learning approaches to image interpretation. The provided ground truth, the eTRIMS Image Database, is a collection of annotated images of real world street scenes. Typical objects in these images are variable in shape and appearance, in the number of its parts and appear in a variety of congurations. The domain of man-made scenes is thus well suited for evaluation and comparison of a variety of interpretation approaches, including those that employ structure models. The provided pixelwise ground truth assigns each image pixel both with a class label and an object label and offers thus ground truth annotation both on the level of pixels and regions. While we believe that such ground truth is of general interest in supervised learning, such data may be of further relevance in emerging real world applications involving automation of man-made scene interpretation.

@techreport{ korc-forstner-tr09-etrims, author = "Kor{\v c}, F. and F{\" o}rstner, W.", title = "{eTRIMS} {I}mage {D}atabase for Interpreting Images of Man-Made Scenes", number = "TR-IGG-P-2009-01", month = "April", year = "2009", institute = "Dept. of Photogrammetry, University of Bonn", url = "http://www.ipb.uni-bonn.de/projects/etrims_db/" }

This report points out the role of sequences of samples for training an incremental learning method. We define characteristics of incremental learning methods to describe the influence of sample ordering on the performance of a learned model. Different types of experiments evaluate these properties for two different datasets and two different incremental learning methods. We show how to find sequences of classes for training just based on the data to get always best possible error rates. This is based on the estimation of Bayes error bounds.

@TECHREPORT{wenzel09*:role, author = {Susanne Wenzel and Wolfgang F{\"o}rstner}, title = {The Role of Sequences for Incremental Learning}, institution = {Department of Photogrammetry, University of Bonn}, year = {2009}, number = {TR-IGG-P-2009-04}, month = {October} }

@TECHREPORT{Roscher2009, author = {Roscher, R., Foerstner, W.}, title = {{Multiclass Bounded Logistic Regression -- Efficient Regularization with Interior Point Method}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2009}, number = {TR-IGG-P-2009-02} }

This is a technical report for presenting a documentation on the segmentation and the graph construction of a Hierarchical Markov random eld (HMRF). The segmentation is based on multiscale analysis and watershed regions as presented in [Drauschke et al., 2006]. The region's development is tracked over the scales, which denes a region hierarchy graph. This graph is used to improve the segmentation by reforming the regions geometrically more precisely. This work is taken from [Drauschke, 2009]. Furthermore, we determine a region adjacency graph from each image partition of all scales. The detected image regions, their adjacent regions and their hierarchical neighbors are saved into an xml-fille for a convenient output.

@TECHREPORT{drauschke09:documentation, author = {Drauschke, M.}, title = {{Documentation: Segmentation and Graph Construction of HMRF}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2009}, number = {TR-IGG-P-2009-03} }

Dieser Report erläutert die Bestimmung von 3D-Merkmalen von Bildregionen durch Zuordnung von Bildpunkten in diesen Regionen zu den Objektpunkten in einer Punktwolke. Die Umsetzung erfolgt in einer grafischen Benutzeroberfläche in Matlab, deren Bedienung in diesem Report veranschaulicht werden soll.

@TECHREPORT{roscher08:bestimmung, author = {Roscher, R.}, title = {{Bestimmung von 3D-Merkmalen von Bildregionen aus Stereobildern}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-05} }

@TECHREPORT{jensen08:schattenentfernung, author = {Jensen, L.}, title = {{Schattenentfernung aus Farbbildern mit dem Retinex-Algorithmus}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-01} }

This technical report gives a short review on boosting with alternating decision trees (ADTboost), which has been proposed by Freund & Mason (1999) and refined by De Comite et al. (2001). This approach is designed for two-class problems, and we extend it towards multi-class classification. The advantage of a multi-class boosting algorithm is its usage in scene interpretation with various kinds of objects. In these cases, two-class approaches will lead to several one class versus background (the other classes) classifications, where we must solve unappropriate results like "always background" or "two or more valid classes" for a sample.

@TECHREPORT{drauschke08:multi, author = {Martin Drauschke}, title = {Multi-class ADTboost}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-06}, month = {August} }

This technical report presents feature subset selection methods for two boosting classication frameworks: Adaboost and ADTboost.

@TECHREPORT{drauschke08:feature, author = {Martin Drauschke}, title = {Feature Subset Selection with Adaboost and ADTboost}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-04}, month = {March} }

Aufgabenstellung: Digitalisierte 16-Bit-Luftbilder sollen automatisch verbessert werden. Dazu haben wir in (1) und (2) den Multi-Dodging-Ansatz vorgeschlagen. In diesem Verfahren wird ein Bild in sich nicht überlappende Ausschnitte (Patches) zerlegt. Dann wird in jedem dieser Bildausschnitte eine Histogrammverebnung durchgeführt. Da dieses Vorgehen die Patchgrenzen im verbesserten Bild hinterlässt, wurde abschließend zwischen den Patches bilinear interpoliert. In dieser Arbeit wird untersucht, ob die Verwendung einer bikubischen Interpolation an Stelle der bilinearen zu besseren Ergebnissen führt.

@TECHREPORT{drauschke08:verbesserung, author = {Drauschke, M.}, title = {{Verbesserung des Multi-Dodgings mittels bikubischer Interpolation}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-07} }

The Stable Regions Image Processing Module is a low-level region detector. It delivers image parts of interest without any further interpretation. These image parts are all regions of an image which do not change much over a certain range in scale space of the image. The output of this IPM is a list of polygons of any shape and their rectangular bounding boxes, which both are saved into an xml-file.

@TECHREPORT{drauschke08:description, author = {Martin Drauschke}, title = {Description of Stable Regions IPM}, institution = {Department of Photogrammetry, University of Bonn}, year = {2008}, number = {TR-IGG-P-2008-03}, month = {March} }

In the field of robotics and computer vision recursive estimation of time dependent processes is one of the key tasks. Usually Kalman filter based techniques are used, which rely on explicit model functions, that directly and explicitly describe the effect of the parameters on the observations. However, some problems naturally result in implicit constraints between the observations and the parameters, for instance all those resulting in homogeneous equation systems. By implicit we mean, that the constraints are given by equations, that are not easily solvable for the observation vector. We derive an iterative extended Kalman filter framework based on implicit measurement equations. In a wide field of applications the possibility to use implicit constraints simplifies the process of specifying suitable measurement equations. As an extension we introduce a robustification technique similar to [Ting et.al 2007] and [Huber 1981], which allows the presented estimation scheme to cope with outliers. Furthermore we will present results for the application of the proposed framework to the structure-from-motion task in the case of an image sequence acquired by an airborne vehicle.

@TECHREPORT{steffen08:robust, author = {Richard Steffen}, title = {A Robust Iterative Kalman Filter Based On Implicit Measurement Equations}, institution = {Department of Photogrammetry, University of Bonn}, year = {2009}, number = {TR-IGG-P-2008-08}, month = {August} }

Ein großer Aufgabenbereich der Bildverarbeitung ist die Merkmalsextraktion. Hierbei ist es zunächst erforderlich, die Bilder durch eine Segmentierung in konsistente Landkarten zu überführen. Wir verwenden zur Segmentierung einen Wasserscheidenalgorithmus. Die Verwaltung der Landkarten sollte möglichst effizient erfolgen. Der vorliegende Report erläutert zunächst unterschiedliche Speicherstrukturen und geht auf einen möglichen Ansatz zur konsistenten Umwandlung von Rasterdaten in Vektordaten ein. In einem zweiten Tiel beschäftigen wir uns mit dem Aufbau einer Datenbank zur Verwaltung dieser Vektordaten von Matlab aus.

@TECHREPORT{herms07:aufbau, author = {Herms, K.}, title = {{Aufbau einer Datenbank unter Matlab zur Verwaltung von Bildsegmenten}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2007}, number = {TR-IGG-P-2007-05}, month = {August}, keywords = {Segmentation, Algorithmic Geometry, GIS} }

Um Gebäude in Bildern detektieren detektieren zu können, greifen wir als ein Merkmal auf Bildkanten zurück. Kanten liegen zum einen aus der Bildsegmentierung vor, können aber auch gezielt mit einer Kantenextraktion aus dem quadratisch Gradientenbild gewonnen werden. Durch die Auswahl von Kanten, die in beiden Fällen auftreten, wollen wir eine Einschränkung auf möglichst relevante Kanten vornehmen. Diese Arbeit beschäftigt sich mit dem Auffinden (und der Gewichtung) von Kantenzügen eines segmentierten Bildes, die einer Kante im quadratidschen Gradientenbild zugeordnet werden können.

@TECHREPORT{siemes07:extraktion, author = {Siemes, K.}, title = {{Extraktion relevanter Bildkanten f{\"u}r die Geb{\"a}udedetektion (umgesetzt in Matlab)}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2007}, number = {TR-IGG-P-2007-05} }

This report describes the details for mirroring the descriptors of the SIFT-features. We show how the mirrored versions are derived by simply resorting the descriptor elements. Furthermore, we describe the matching of features within an image. The peculiarity of this task is the search for more than one - the best - match within an single image. The presented methods are based on the work of (Wenzel2006,Detektion). After the introduction the functionallity of the SIFT-feature detector is drafted and the development of the descriptors is described in detail. The following sections describe the details of mirroring and matching the features. Dieser Bericht geht auf die Details zur Spiegelung von SIFT-Feature Deskritoren ein. Es wird gezeigt, wie durch einfaches Umsortieren der Elemente des Feature Deskriptors gespiegelte Versionen der Deskriptoren erlangt werden können. Des Weiteren wird erläutert, wie Features innerhalb eines Bildes zugeordnet werden können. Die Besonderheit dieser Aufgabenstellung liegt in der gesuchten Zuordnung nicht eines - des besten - Matches, sondern in der Zuordnung aller Matches in einem Bild. Die vorgestellten Methoden basieren auf (Wenzel2006,Detektion).

@TECHREPORT{wenzel07:spiegelung, author = {Susanne Wenzel}, title = {Spiegelung und Zuordnung der SIFT-Feature Deskriptoren für die Detektion von Symmetrien und wiederholten Strukturen in Bildern}, institution = {Department of Photogrammetry, University of Bonn}, year = {2007}, number = {TR-IGG-P-2007-04}, month = {August} }

Das Problem stellt sich wie folgt dar: Die Luftbilder wurden mit einem Vexcel-Scanner digitalisiert und als 16-Bit-Bilder abgespeichert. Die Bilder sollen automatisch nachbereitet werden.

@TECHREPORT{drauschke06:automatisch, author = {Drauschke, M.}, title = {{Automatisches Dodging von Luftbildern}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2006}, number = {TR-IGG-P-2006-01} }

This paper presents a method for detecting complex man-made-objects in images. The detection model is a bayesian net that aggregates cliques of image regions which may cover a complex object. Observable attributes of the regions are derived from a rich symbolic image description containing points, lines and regions as basic features including their relations. The model captures the dependency of the region aggregates on the features and their relations with respect to observability due to occlusions and to perspective deformations. Cliques are classified using MAP estimation. Up to now, the model captures cliques with one, two and three regions which is sufficient for detecting polyhedral objects. The model allows to detect and locate multiple appearances of object classes. The joint distribution of the Bayesian net is determined in a supervised learning step based on images with annotated regions. The method is realized and demonstrated for the detection of building roofs in aerial images.

@TECHREPORT{schuster05:detection, author = {Schuster, H.-F.}, title = {{Detection of man-made-objects based on spatial aggregations}}, institution = {Department of Photogrammetry, University of Bonn}, year = {2006}, number = {TR-IGG-P-2006-01} }

This report describes research on the extraction of urban structures from aerial images or hight resolution satellite scenes. We aim at a separation of neighboring textured regions important for different urban structures. It has been shown, that grey level segmentation alone s not sufficient to solve this problem. If we want to use texture additionally, the development of a suitable representation of texture images is required. We use the scale characteristics of the local autocovariance function, called SCAF, of the possibly multiband image function. The final result of the process are textrure edges. Zusammenfassung Dieser Bericht beschreibt den aktuellen Stand der Forschung auf dem Gebiet der Extraktion urbaner Strukturen aus Luft- bzw. hochauflösenden Satellitendaten am Institut für Photogrammetrie. Die Trennung (Segmentierung) benachbarter texturierter, für die Beschreibung solcher urbaner Strukturen bedeutender, Regionen ist das Ziel der Arbeit. Es hat sich gezeigt, dass Segmentierungsverfahren auf Intensitätsbasis allein oft unzureichende Ergebnisse liefern. Deshalb wurde eine Repräsentationsform für Bilddaten entwickelt, die es erlaubt, zusätzlich Texturinformationen für die Problemlösung zu nutzen. Dies erfolg unter Verwendung der Skalencharakteristik der Lokalen Autokovarianzfunktion, kurz SCAF, der Bildfunktion. Das Ergebnis des vorgestellten Prozesses sind Texturkanten.

@TECHREPORT{faber99:technischer, author = "Faber, Anette", title = "Technischer Bericht - Skalencharakteristik der Lokalen Autokovarianzfunktion für die Textursegmentierung", year = "1999", city = "Bonn",}

@TECHREPORT{heuel*99:technical, author = "Heuel, Stephan and Faber, Anette", title = "Technical Report - Segmentation and Classification of Landcover Areas using a Polymorphic Feature Extraction", year = "1999", city = "Bonn",}

@TECHREPORT{weidner*97:discriminating, author = "Weidner, Uwe and Brunn, Ansgar", title = "Discriminating Building and Vegetation Areas within Digital Surface Models", year = "1997", city = "Bonn",}

@TECHREPORT{faber97:segmentierung, author = "Faber, Anette", title = "Segmentierung und Klassifikation orthogonaler Straßennetze mittels Richtungscodierung - Technischer Bericht", year = "1997", city = "Bonn",}

In this report we deal with roof extraction from Digital Surface Models (DSM) as a first step to integrate polyhedral building models into our approach to building extraction from DSM (Weidner and Förstner 1995, Weidner 1997a). Results are also presented in Weidner 1997b.

@TECHREPORT{weidner97:roof, author = "Weidner, Uwe", title = "Roof Extraction from Digital Surface Models", year = "1997", city = "Bonn",}

The technical report deals with an approach for extracting the 3D-shape of buildings from high resolution Digital Elevation Models (DEMs), having a grid resolution between 0.5 and 5m. The steps of the proposed procedure increasingly use explicit domain knowledge, specifically geometric constraints in the form of parametric and prismatic building models. A new MDL-based approach generating a polygonal ground plan from segment boundaries is given. The used knowledge is object related making adaption to data of different density and resolution simple and transparent.

@TECHREPORT{weidner95:building, author = "Weidner, Uwe", title = "Building Extraction from Digital Elevation Models", year = "1995", city = "Bonn",}

The technical report describes the basics of our approach for building extraction using automatically derived Digital Elevation Models (DEMs), and the results for the ISPRS test data sets.

@TECHREPORT{weidner95:evaluation, author = "Weidner, Uwe", title = "Evaluation of Building Extraction from Digital Elevation Models", year = "1995", city = "Bonn",}

Kurzdokumentation über die Handhabung des Programms bex.

@TECHREPORT{weidner95:building, author = "Weidner, Uwe", title = "Building Extraction from Digital Elevation Models", year = "1995", city = "Bonn",}

The note proposes a metric for comparing symmetric positive definite matrices. Symmetric positive definite matrices can be interpreted as covariance matrices. The metric is shown to be invariant with respect to common affine transformations of the reference coordinate system of the covariance matrices and to inversions of the matrices, thus also measures the distance between the corresponding weight matrices.

@TECHREPORT{weidner95:metric, author = "Weidner, Uwe", title = "A Metric for Comparing Symmetric Positive Definite Matrices", year = "1995", city = "Bonn",}

Kurzdokumentation über die Handhabung des Programms geo. Weitergehende Informationen können aus Best (1990) und Bevacqua/Floris (1987) entnommen werden.

@TECHREPORT{weidner94:geo, author = "Weidner, Uwe", title = "GEO - Algorithmen der Digitalen Bildverarbeitung", year = "1994", city = "Bonn",}

A domain-independent objective mechanism is developed for generalization of linear patterns. It is based on the Minimum-Description-Length principle, seeking the simplest description of a given polyline. The hypotheses are generated by the farthest point algorithm. The whole mechanism is objactive in the sense of without using any control parameter. This mechanism has been tested on segmented images and polygon maps.

@TECHREPORT{förstner*94:generalization, author = "Förstner, Wolfgang and Pan, He-Ping", title = "Generalization of Linear Patterns Based on MDL Criterion", year = "1994", city = "Bonn",}

Kurzdokumentation über die Handhabung des Programms grinv. Weitergehende Informationen können aus Wang et al. (1981) entnommen werden.

@TECHREPORT{weidner94:grinv, author = "Weidner, Uwe", title = "GRINV - Gradient Inverse Weighted Smoothing", year = "1994", city = "Bonn",}

Kurzdokumentation über die Handhabung des Programms reko. Weitergehende Informationen können aus Weidner (1991) entnommen werden.

@TECHREPORT{weidner94:reko, author = "Weidner, Uwe", title = "REKO - Bildrekonstruktion mittels informationserhaltender Filterung", year = "1994", city = "Bonn",}

@TECHREPORT{winter94:bericht, author = "Winter, Stephan", title = "Bericht über die AGDM 94", year = "1994", city = "Bonn",}

Kurzdokumentation über die Handhabung des Programms vistri.

@TECHREPORT{weidner94:vistri, author = "Weidner, Uwe", title = "VISTRI - Visualisierung von Flächen", year = "1994", city = "Bonn",}

Dieses Arbeitspapier stellt differentialgeometrische Grundlagen in Tensorschreibweise dar und zeigt Möglichkeiten zur Berechnung von Krümmungsmaßen für durch diskrete Punkte gegebene Flächen auf. Weiterhin werden Möglichkeiten zur Segmentierung von Flächen auf der Basis dieser Krümmungsmaße diskutiert.

@TECHREPORT{weidner93:krümmungsmaße, author = "Weidner, Uwe", title = "Krümmungsmaße", year = "1993", city = "Bonn",}