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

Three methods of automatic classification of leaf diseases are described based on high-resolution multispectral stereo images. Leaf diseases are economically important as they can cause a loss of yield. Early and reliable detection of leaf diseases has important practical relevance, especially in the context of precision agriculture for localized treatment with fungicides. We took stereo images of single sugar beet leaves with two cameras (RGB and multispectral) in a laboratory under well controlled illumination conditions. The leaves were either healthy or infected with the leaf spot pathogen Cercospora beticola or the rust fungus Uromyces betae. To fuse information from the two sensors, we generated 3-D models of the leaves. We discuss the potential of two pixelwise methods of classification: k-nearest neighbour and an adaptive Bayes classification with minimum risk assuming a Gaussian mixture model. The medians of pixelwise classification rates achieved in our experiments are 91% for Cercospora beticola and 86% for Uromyces betae. In addition, we investigated the potential of contextual classification with the so called conditional random field method, which seemed to eliminate the typical errors of pixelwise classification.

@article{bauer*11:potential, author={Bauer, Sabine and Kor{\v c}, Filip and F\"{o}rstner, Wolfgang}, title={The potential of automatic methods of classification to identify leaf diseases from multispectral images}, journal={Precision Agriculture}, year={2011}, number={3}, pages={361-377}, }

We present a fast automatic reproducible method for 3d semantic segmentation of magnetic resonance images of the knee. We formulate a single global model that allows to jointly segment all classes. The model estimation was performed automatically without manual interaction and parameter tuning. The segmentation of a magnetic resonance image with 11 Mio voxels took approximately one minute. Our labeling results by far do not reach the performance of complex state of the art approaches designed to produce clinically relevant results. Our results could potentially be useful for rough visualization or initialization of computationally demanding methods. Our main contribution is to provide insights in possible strategies when employing global statistical models.

@INPROCEEDINGS{korc*10:nonparametric, author = {Kor{\v c}, F. and Schneider, D. and F{\" o}rstner, W.}, title = {{On Nonparametric Markov Random Field Estimation for Fast Automatic Segmentation of MRI Knee Data}}, booktitle = {Proceedings of the 4th Medical Image Analysis for the Clinic - A Grand Challenge workshop, MICCAI}, year = {2010} }

This paper reports on methods for the automatic detection and classification of leaf diseases based on high resolution multispectral images. Leaf diseases are economically important as they could cause a yield loss. Early and reliable detection of leaf diseases therefore is of utmost practical relevance - especially in the context of precision agriculture for localized treatment with fungicides. Our interest is the analysis of sugar beet due to their economical impact. Leaves of sugar beet may be infected by several diseases, such as rust (Uromyces betae), powdery mildew (Erysiphe betae) and other leaf spot diseases (Cercospora beticola and Ramularia beticola). In order to obtain best classification results we apply conditional random fields. In contrast to pixel based classifiers we are able to model the local context and contrary to object centred classifiers we simultaneously segment and classify the image. In a first investigation we analyse multispectral images of single leaves taken in a lab under well controlled illumination conditions. The photographed sugar beet leaves are healthy or either infected with the leaf spot pathogen Cercospora beticola or with the rust fungus Uromyces betae. We compare the classification methods pixelwise maximum posterior classification (MAP), objectwise MAP as soon as global MAP and global maximum posterior marginal classification using the spatial context within a conditional random field model.

@INPROCEEDINGS{bauer*09:investigation, author = {Bauer, Sabine Daniela and Kor{\v c}, Filip and F\"{o}rstner, Wolfgang}, title = {{Investigation into the classification of diseases of sugar beet leaves using multispectral images}}, booktitle={Precision Agriculture 2009}, year = {2009}, pages = {229-238}, city = {Bonn}, proceeding = {Precision Agriculture} }

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/" }

We investigate maximum likelihood parameter learning in Conditional Random Fields (CRF) and present an empirical study of pseudo-likelihood (PL) based approximations of the parameter likelihood gradient. We show that these parameter learning methods can be improved and evaluate the resulting performance employing different inference techniques. We show that the approximation based on penalized pseudo-likelihood (PPL) in combination with the Maximum A Posteriori (MAP) inference yields results comparable to other state of the art approaches, while providing low complexity and advantages to formulating parameter learning as a convex optimization problem. Eventually, we demonstrate applicability on the task of detecting man-made structures in natural images.

@INPROCEEDINGS{korc*08:approximate, author = {Kor{\v c}, Filip and F{\" o}rstner, Wolfgang}, title = {Approximate Parameter Learning in {C}onditional {R}andom {F}ields: An Empirical Investigation}, booktitle = {Pattern Recognition}, year = {2008}, editor = {G. Rigoll}, number = {5096}, series = {LNCS}, pages = {11-20}, publisher = {Springer} }

We investigate Discriminative Random Fields (DRF) which provide a principled approach for combining local discriminative classifiers that allow the use of arbitrary overlapping features, with adaptive data-dependent smoothing over the label field. We discuss the differences between a traditional Markov Random Field (MRF) formulation and the DRF model, and compare the performance of the two models and an independent sitewise classifier. Further, we present results suggesting the potential for performance enhancement by improving state of the art parameter learning methods. Eventually, we demonstrate the application feasibility on both synthetic and natural images.

@INPROCEEDINGS{korc*08:interpreting, author = {Kor{\v c}, Filip and F{\" o}rstner, Wolfgang}, title = {Interpreting Terrestrial Images of Urban Scenes Using Discriminative Random Fields}, booktitle = {Proc. of the 21st Congress of the International Society for Photogrammetry and Remote Sensing (ISPRS)}, year = {2008} }

We present a solution to the following discrete optimization problem. Given a set of independent, possibly overlapping image regions and a non-negative likeliness of the individual regions, we select a non-overlapping subset that is optimal with respect to the following requirements: First, every region is either part of the solution or has an overlap with it. Second, the degree of overlap of the solution with the rest of the regions is maximized together with the likeliness of the solution. Third, the likeliness of the individual regions influences the overall solution proportionally to the degree of overlap with neighboring regions. We represent the problem as a graph and solve the task by reduction to a constrained binary integer programming problem. The problem involves minimizing a linear objective function subject to linear inequality constraints. Both the objective function and the constraints exploit the structure of the graph. We illustrate the validity and the relevance of the proposed formulation by applying the method to the problem of facade window extraction. We generalize our formulation to the case where a set of hypotheses is given together with a binary similarity relation and similarity measure. Our formulation then exploits combination of degree and structure of hypothesis similarity and likeliness of individual hypotheses. In this case, we present a solution with non-similar hypotheses which can be viewed as a non-redundant representation.

@INPROCEEDINGS{korc*08:optimal, author = {Kor{\v c}, Filip and F{\" o}rstner, Wolfgang}, title = {Finding Optimal Non-Overlapping Subset of Extracted Image Objects}, booktitle = {Proc. of the 12th International Workshop on Combinatorial Image Analysis (IWCIA)}, year = {2008}, }

This work contributes to detection and tracking of walking or running humans in surveillance video sequences. We propose a 2D model-based approach to the whole body tracking in a video sequence captured from a single camera view. An extended six-link biped human model is employed. We assume that a static camera observes the scene horizontally or obliquely. Persons can be seen from a continuum of views ranging from a lateral to a frontal one. We do not expect humans to be the only moving objects in the scene and to appear at the same scale at different image locations.

@INBOOK{KorcHlavacDetectTrackHumans2007, author = {Kor{\v c}, Filip and Hlav{\'a}{\v c}, V{\'a}clav}, year = {2007}, chapter = {Detection and Tracking of Humans in Single View Sequences Using 2D Articulated Model}, title = {Human Motion - Understanding, Modeling, Capture and Animation}, editor = {Rosenhahn, Bodo and Klette, Reinhard and Metaxas, Dimitris}, publisher = {Springer}, series = {Computational Imaging and Vision}, volume = {36}, edition = {1}, isbn = {978-1-4020-6692-4}, pages = {105--130}, keywords = {human detection in video, model-based human detection}, url = {ftp://cmp.felk.cvut.cz/pub/cmp/articles/hlavac/2007KorcHlavacDetectTrackHumans.pdf} }