Catálogo de publicaciones - libros

There is biological evidence that human actions are composed out of action primitives, like words and sentences being composed out of phonemes. Similarly to language processing, one possibility to model and recognize complex actions is to use with action primitives as the alphabet. A major challenge here is that the action primitives need to be recovered first from the noisy input signal before further processing with the action grammar can be done. In this paper we combine a Hidden Markov Model-based approach with a simplified version of a condensation algorithm which allows to recover the action primitives in an observed action. In our approach, the primitives may have different lengths, no clear “divider” between the primitives is necessary. The primitive detection is done online, no storing of past data is required. We verify our approach on a large database. Recognition rates are slightly lower than the rate when recognizing the singular action primitives.

This paper deals with the segmentation of multiple moving objects in image sequences. A method for estimating the motion of objects without the use of features is presented. This is used to predict the position and orientation in future frames of the sequence. Experiments on real data show that this estimation can be used to improve segmentation.

The detection of basic events such as turning points in object trajectories is an important low-level task of image sequence analysis. We propose extending the SUSAN algorithm to the spatio-temporal domain for a context-free detection of salient events, which can be used as a starting point for further motion analysis. While in the static 2D-case SUSAN returns a map indicating edges and corners, we obtain in a straight forward extension of SUSAN a 2D+1D saliency map indicating edges and corners in both space and time. Since the mixture of spatial and temporal structures is still unsatisfying, we propose a modification better suited for event analysis.

Tracking of a camera pose in all 6 degrees of freedom is a task with many applications in 3D-imaging as i.e. augmentation or robot navigation. Structure from motion is a well known approach for this task, with several well known restrictions. These are namely the scale ambiguity of the calculated relative pose and the need of a certain camera movement (preferably lateral) to initiate the tracking.

In the last few years time-of-flight imaging sensors were developed that allow the measuring of metric depth over a whole region with a frame rate similar to a standard CCD-camera.

In this work a camera rig consisting of a standard 2D CCD camera and a time-of-flight 3D camera is used. Structure from motion is calculated on the 2D image, aided by the depth measurement from the time-of-flight camera to overcome the restrictions named above. It is shown how the additional 3D-information can be used to improve the accuracy of the camera pose estimation.

In this paper, we propose novel blur invariant features for the recognition of objects in images. The features are computed either using the phase-only spectrum or bispectrum of the images and are invariant to centrally symmetric blur, such as linear motion or defocus blur as well as linear illumination changes. The features based on the bispectrum are also invariant to translation, and according to our knowledge they are the only combined blur-translation invariants in the frequency domain. We have compared our features to the blur invariants based on image moments in simulated and real experiments. The results show that our features can recognize blurred images better and, in a practical situation, they are faster to compute using FFT.

Discriminative feature extraction is one of the fundamental problems in pattern recognition and signal processing. It was recently proposed that maximizing the class prediction by neighboring samples in the transformed space is an effective objective for learning a low-dimensional linear embedding of labeled data. The associated methods, Neighborhood Component Analysis (NCA) and Relevant Component Analysis (RCA), have been proven to be useful preprocessing techniques for discriminative information visualization and classification. We point out here that NCA and RCA are prone to overfitting and therefore regularization is required. NCA and RCA’s failure for high-dimensional data is demonstrated in this paper by experiments in facial image processing. We also propose to incorporate a Gaussian prior into the NCA objective and obtain the Regularized Neighborhood Component Analysis (RNCA). The empirical results show that the generalization can be significantly enhanced by using the proposed regularization method.

Applying a minimum-cost path algorithm to find the path through the bottom of a curvilinear valley yields a biased path through the inside of a corner. DNA molecules, blood vessels, and neurite tracks are examples of string-like (network) structures, whose minimum-cost path is cutting through corners and is less flexible than the underlying centerline. Hence, the path is too short and its shape too stiff, which hampers quantitative analysis. We developed a method which solves this problem and results in a path whose distance to the true centerline is more than an order of magnitude smaller in areas of high curvature. We first compute an initial path. The principle behind our iterative algorithm is to deform the image space, using the current path in such a way that curved string-like objects are straightened before calculating a new path. A damping term in the deformation is needed to guarantee convergence of the method.

Several novel methods based on locally extracted image features and spatial constellation models have recently been introduced for invariant object class detection and recognition. The accuracy and reliability of the methods depend on the success of both tasks: image feature extraction and spatial constellation model search. In this study a novel method for object class detection is introduced. It combines supervised Gabor-based confidence-ranked image features and affine invariant point pattern matching. The method is able to deal with occlusions and its potential is demonstrated on a standard face database.

We present a technique for simultaneous segmentation and classification of image partitions using combinatorial optimization techniques. By combining existing image segmentation approaches with simple learning techniques we show how prior knowledge can be incorporated into the visual grouping process through the formulation of a quadratic binary optimization problem. We further show how such to efficiently solve such problems through relaxation techniques and trust region methods. This has resulted in an method that partitions images into a number of disjoint regions based on previously learned example segmentations. Preliminary experimental results are also presented in support of our suggested approach.

Hyperspectral data is usually high dimensional, and there is often a scarcity of available ground truth pixels . Thus the task of applying even a simple classifier such as the Gaussian Maximum Likelihood (GML) classifier usually forces the analyst to reduce the complexity of the implicit parameter estimation task. For decades, the common perception in the literature has been that the solution to this has been to reduce data dimensionality. However, as can be seen from a result by Cover [1], reducing dimensionality increases the risk of making the classification problem more complex.Using the simple GML classifier we compare state of the art dimensionality reduction strategies with a recently proposed strategy for sparsing of parameter estimates in full dimension [2]. Results show that reducing parameter estimation complexity by fitting sparse models in full dimension have a slight edge on the common approaches.