Catálogo de publicaciones - libros

Interest point detection in still images is a well-studied topic in computer vision. In the spatiotemporal domain, however, it is still unclear which features indicate useful interest points. In this paper we approach the problem by a detector from examples: we record eye movements of human subjects watching video sequences and train a neural network to predict which locations are likely to become eye movement targets. We show that our detector outperforms current spatiotemporal interest point architectures on a standard classification dataset.

This paper presents a reliable coin recognition system that is based on a registration approach. To optimally align two coins we search for a rotation in order to reach a maximal number of colinear gradient vectors. The gradient magnitude is completely neglected. After a quantization of the gradient directions the computation of the induced similarity measure can be done efficiently in the Fourier domain. The classification is realized with a simple nearest neighbor classification scheme followed by several rejection criteria to meet the demand of a low false positive rate.

We present a new technique for the extraction of features from 3D volumetric data sets based on group integration. The features are invariant to translation, rotation and global radial deformations. They are robust to local arbitrary deformations and nonlinear gray value changes, but are still sensitive to fine structures. On a data set of 389 confocally scanned pollen from 26 species we get a precision/recall of 99.2% with a simple 1NN classifier. On volumetric transmitted light data sets of about 180,000 airborne particles, containing about 22,700 pollen grains from 33 species, recorded with a low-cost optic in a fully automated online pollen monitor the mean precision for allergenic pollen is 98.5% (recall: 86.5%) and for the other pollen 97.5% (recall: 83.4%).

We present a model for learning convex kernel combinations in classification problems with structured output domains. The main ingredient is a hidden Markov model which forms a layered directed graph. Each individual layer represents a multilabel version of nonlinear kernel discriminant analysis for estimating the emission probabilities. These kernel learning machines are equipped with a mechanism for finding convex combinations of kernel matrices. The resulting can handle multiple partial paths through the label hierarchy in a consistent way. Efficient approximation algorithms allow us to train the model to large-scale learning problems. Applied to the problem of document categorization, the method exhibits excellent predictive performance.

We address the problem of describing the mean object for a set of planar shapes in the case that the considered dissimilarity measures are semi-metrics, i.e. in the case that the triangle inequality is generally not fulfilled. To this end, a matching of two planar shapes is computed by cutting an appropriately defined graph the edge weights of which encode the local similarity of respective contour parts on either shape. The cost of the minimum cut can be interpreted as a semi-metric on the space of planar shapes. Subsequently, we introduce the notion of a mean shape for the case of semi-metrics and show that this allows to perform a shape retrieval which mimics human notions of shape similarity.

This paper presents a real-time system for vision-based pedestrian recognition from a moving vehicle-mounted catadioptric camera. For efficiency, a rectification of the catadioptric image using a virtual cylindrical camera is employed. We propose a novel hybrid combination of a boosted cascade of wavelet-based classifiers with a subsequent texture-based neural network involving adaptive local features as final cascade stage. Within this framework, both fast object detection and powerful object classification are combined to increase the robustness of the recognition system. Further, we compare the hybrid cascade framework to a state-of-the-art multi-cue pedestrian recognition system utilizing shape and texture cues. Image distortions of the objects of interest due to the virtual cylindrical camera transformation are both explicitly and implicitly addressed by shape transformations and machine learning techniques. In extensive experiments, both systems under consideration are evaluated on a real-world urban traffic dataset. Results show the contributions of the various components in isolation and document superior performance of the proposed hybrid cascade system.

In this article we present EANT, a method that creates neural networks (NNs) by evolutionary reinforcement learning. The structure of NNs is developed using mutation operators, starting from a minimal structure. Their parameters are optimised using CMA-ES. EANT can create NNs that are very specialised; they achieve a very good performance while being relatively small. This can be seen in experiments where our method competes with a different one, called NEAT, to create networks that control a robot in a visual servoing scenario.

Although lasers are widely used for welding in precision engineering industry, it is still a challenge to achieve high accuracy in creating and positioning welding spots at extremely high processing speed.

Towards this end, we propose a system for monitoring the welding process in order to ensure good quality of the welding spots. Our technology enables high speed image acquisition confocally to the laser beam with a direct view onto the melt. This innovative system permits accurate estimation of the melt pool’s position and radius, which, however, must be performed at frame rates above 200 fps. We therefore employ fast correlation based approaches for sampling the melt pool’s contour and robustly fitting a circle to it. In addition, the approaches enable sputter detection via outlier classification.

To assess the performance of each presented method, extensive experiments are conducted. The proposed paradigms can furthermore be conveniently adapted to a variety of problems dealing with rapid shape estimation in noisy environments.

Model-based image interpretation extracts high-level information from images using a priori knowledge about the object of interest. The computational challenge is to determine the model parameters that best match a given image by searching for the global optimum of the involved objective function. Unfortunately, this function is usually designed manually, based on implicit and domain-dependent knowledge, which prevents the fitting task from yielding accurate results.

In this paper, we demonstrate how to improve model fitting by learning objective functions from annotated training images. Our approach automates many critical decisions and the remaining manual steps hardly require domain-dependent knowledge. This yields more robust objective functions that are able to achieve the accurate model fit. Our evaluation uses a publicly available image database and compares the obtained results to a recent state-of-the-art approach.

The relationship between geometric folding of the chromatin fiber and genome function is a key issue in cell biology. We propose different approaches based on statistical shape theory to investigate the geometric variability of chromatin folding in nuclei of interphase human fibroblasts. Our main purpose is to assess the degree of variability of folding of the chromatin fiber, measured by fluorescent in situ hybridization, using BAC probes in combination with 3D confocal microscopy. We employ point-based registration, the complex Bingham distribution, generalized Procrustes analysis, and the Kendall spherical coordinate system. The approaches have been applied using 337 3D multi-channel microscopy images. We have analyzed the geometric structure formed by gene-rich highly expressed genomic regions and areas that are gene-poor and have a low transcriptional activity. It turned out that the structure formed by these genomic regions exhibit high shape variation, however, most of them can be characterized by a non-uniform shape distribution.