Publikationen

@PHDTHESIS{korc12:tractable, author={Kor{\v c}, Filip}, title={Tractable Learning for a Class of Global Discriminative Models for Context Sensitive Image Interpretation}, School={Department of Photogrammetry, University of Bonn}, year={2012} }

Abstract

We propose an innovative machine learning algorithm called incremental import vector machines that is used for classification purposes. The classifier is specifically designed for the task of sequential learning, in which the data samples are successively presented to the classifier. The motivation for our work comes from the effort to formulate a classifier that can manage the major challenges of sequential learning problems, while being a powerful classifier in terms of classification accuracy, efficiency and meaningful output. One challenge of sequential learning is that data samples are not completely available to the learner at a given point of time and generally, waiting for a representative number of data is undesirable and impractical. Thus, in order to allow for a classification of given data samples at any time, the learning phase of the classifier model needs to start immediately, even if not all training samples are available. Another challenge is that the number of sequential arriving data samples can be very large or even infinite and thus, not all samples can be stored. Furthermore, the distribution of the sample can vary over time and the classifier model needs to remain stable and unchanged to irrelevant samples while being plastic to new, important samples. Therefore our key contribution is to develop, analyze and evaluate a powerful incremental learner for sequential learning which we call incremental import vector machines (I²VMs). The classifier is based on the batch machine learning algorithm import vector machines, which was developed by Zhu and Hastie (2005). I²VM is a kernel-based, discriminative classifier and thus, is able to deal with complex data distributions. Additionally ,the learner is sparse for an efficient training and testing and has a probabilistic output. A key achievement of this thesis is the verification and analysis of the discriminative and reconstructive model components of IVM and I²VM. While discriminative classifiers try to separate the classes as well as possible, classifiers with a reconstructive component aspire to have a high information content in order to approximate the distribution of the data samples. Both properties are necessary for a powerful incremental classifier. A further key achievement is the formulation of the incremental learning strategy of I²VM. The strategy deals with adding and removing data samples and the update of the current set of model parameters. Furthermore, also new classes and features can be incorporated. The learning strategy adapts the model continuously, while keeping it stable and efficient. In our experiments we use I²VM for the semantic segmentation of images from an image database, for large area land cover classification of overlapping remote sensing images and for object tracking in image sequences. We show that I²VM results in superior or competitive classification accuracies to comparable classifiers. A substantial achievement of the thesis is that I²VM's performance is independent of the ordering of the data samples and a reconsidering of already encountered samples for learning is not necessary. A further achievement is that I²VM is able to deal with very long data streams without a loss in the efficiency. Furthermore, as another achievement, we show that I²VM provide reliable posterior probabilities since samples with high class probabilities are accurately classified, whereas relatively low class probabilities are more likely referred to misclassified samples.

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.

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 present a novel approach for a rigorous bundle adjustment for omnidirectional and multi-view cameras, which enables an efficient maximum-likelihood estimation with image and scene points at infinity. Multi-camera systems are used to increase the resolution, to combine cameras with different spectral sensitivities (Z/I DMC, Vexcel Ultracam) or - like omnidirectional cameras - to augment the effective aperture angle (Blom Pictometry, Rollei Panoscan Mark III). Additionally multi-camera systems gain in importance for the acquisition of complex 3D structures. For stabilizing camera orientations - especially rotations - one should generally use points at the horizon over long periods of time within the bundle adjustment that classical bundle adjustment programs are not capable of. We use a minimal representation of homogeneous coordinates for image and scene points. Instead of eliminating the scale factor of the homogeneous vectors by Euclidean normalization, we normalize the homogeneous coordinates spherically. This way we can use images of omnidirectional cameras with single-view point like fisheye cameras and scene points, which are far away or at infinity. We demonstrate the feasibility and the potential of our approach on real data taken with a single camera, the stereo camera FinePix Real 3D W3 from Fujifilm and the multi-camera system Ladybug3 from Point Grey.

@INPROCEEDINGS{schneider*12:bundle, author = {Schneider, Johannes and Schindler, Falko and L{\"a}be, Thomas and F{\"o}rstner, Wolfgang}, title = {{Bundle Adjustment for Multi-camera Systems with Points at Infinity}}, booktitle = {ISPRS Annals of Photogrammetry, Remote Sensing and the Spatial Information Sciences; Proc. of 22nd Congress of the International Society for Photogrammetry and Remote Sensing (ISPRS)}, year = {2012}, note = {to appear} }

Our objective is the categorization of the most dominant objects in facade images, like windows, entrances and balconies. In order to execute an image interpretation of complex scenes we need an interaction between low level bottom-up feature detection and highlevel inference from top-down. A top-down approach would use results of a bottom-up detection step as evidence for some high-level inference of scene interpretation. We present a statistically founded object categorization procedure that is suited for bottom-up object detection. Instead of choosing a bag of features in advance and learning models based on these features, it is more natural to learn which features best describe the target object classes. Therefore we learn increasingly complex aggregates of line junctions in image sections from man-made scenes. We present a method for the classification of image sections by using the histogram of diverse types of line aggregates.

@INPROCEEDINGS{wenzel*12:learning, author = {Wenzel, Susanne and F{\" o}rstner, Wolfgang}, title = {Learning a compositional representation for facade object categorization}, booktitle = {ISPRS Annals of Photogrammetry, Remote Sensing and the Spatial Information Sciences; Proc. of 22nd Congress of the International Society for Photogrammetry and Remote Sensing (ISPRS)}, year = {2012}, volume = {I-3}, number = { 2012}, pages = {197-202}, doi = {10.5194/isprsannals-I-3-197-2012} }