Catálogo de publicaciones - libros

One problem in appearance-based pose estimation is the need for many training examples, i.e. images of the object in a large number of known poses. Some invariance can be obtained by considering translations, rotations and scale changes in the image plane, but the remaining degrees of freedom are often handled simply by sampling the pose space densely enough. This work presents a method for accurate interpolation between training views using local linear models. As a view representation local soft orientation histograms are used. The derivative of this representation with respect to the image plane transformations is computed, and a Gauss-Newton optimization is used to optimize all pose parameters simultaneously, resulting in an accurate estimate.

We present results from a study where we segment fish in images captured within fish cages. The ultimate goal is to use this information to extract the weight distribution of the fish within the cages. Statistical shape knowledge is added to a Mumford-Shah functional defining the image energy. The fish shape is represented explicitly by a polygonal curve, and the energy minimization is done by gradient descent. The images represent many challenges with a highly cluttered background, inhomogeneous lighting and several overlapping objects. We obtain good segmentation results for silhouette-like images containing relatively few fish. In this case, the fish appear dark on a light background and the image energy is well behaved. In cases with more difficult lighting conditions the contours evolve slowly and often get trapped in local minima

Automatic construction of Shape Models from examples has recently been the focus of intense research. These methods have proved to be useful for shape segmentation, tracking, recognition and shape understanding. In this paper we discuss automatic landmark selection and correspondence determination from a discrete set of landmarks, typically obtained by feature extraction. The set of landmarks may include both outliers and missing data. Our framework has a solid theoretical basis using principles of Minimal Description Length (MDL). In order to exploit these ideas, new non-heuristic methods for (i) principal component analysis and (ii) Procrustes mean are derived - as a consequence of the modelling principle. The resulting MDL criterion is optimised over both discrete and continuous decision variables. The algorithms have been implemented and tested on the problem of automatic shape extraction from feature points in image sequences.

The segmentation of objects in image sequences is an important and difficult problem in computer vision with applications to e.g. video surveillance. In this paper we propose a new method for variational segmentation of image sequences containing nonrigid, moving objects. The method is based on the classical Chan-Vese model augmented with a novel frame-to-frame interaction term, which allow us to update the segmentation result from one image frame to the next using the previous segmentation result as a shape prior. The interaction term is constructed to be pose-invariant and to allow moderate deformations in shape. It is expected to handle the appearance of occlusions which otherwise can make segmentation fail. The performance of the model is illustrated with experiments on real image sequences.

A robust object/face detection technique processing every frame in real-time (video-rate) is presented. A methodological novelty are the suggested quantized angle features (“quangles”), being designed for illumination invariance without the need for pre-processing, e.g. histogram equalization. This is achieved by using both the gradient direction and the double angle direction (the structure tensor angle), and by ignoring the magnitude of the gradient. Boosting techniques are applied in a quantized feature space. Separable filtering and the use of lookup tables favor the detection speed. Furthermore, the gradient may then be reused for other tasks as well. A side effect is that the training of effective cascaded classifiers is feasible in very short time, less than 1 hour for data sets of order 10. We present favorable results on face detection, for several public databases (e.g. 93% Detection Rate at 1×10 False Positive Rate on the CMU-MIT frontal face test set).

Many potential applications exist where a fast and robust detection of human faces is required. Different cues can be used for this purpose. Since each cue has its own pros and cons we, in this paper, suggest to combine several complimentary cues in order to gain more robustness in face detection. Concretely, we apply skin-color, shape, and texture to build a robust detector. We define the face detection problem in a state-space spanned by position, scale, and rotation. The state-space is searched using a Particle Filter where 80% of the particles are predicted from the past frame, 10% are chosen randomly and 10% are from a texture-based detector. The likelihood of each selected particle is evaluated using the skin-color and shape cues. We evaluate the different cues separately as well as in combination. An improvement in both detection rates and false positives is obtained when combining them.

The accuracy of data classification methods depends considerably on the data representation and on the selected features. In this work, the elastic net model selection is used to identify meaningful and important features in face recognition. Modelling the characteristics which distinguish one person from another using only subsets of features will both decrease the computational cost and increase the generalization capacity of the face recognition algorithm. Moreover, identifying which are the features that better discriminate between persons will also provide a deeper understanding of the face recognition problem. The elastic net model is able to select a subset of features with low computational effort compared to other state-of-the-art feature selection methods. Furthermore, the fact that the number of features usually is larger than the number of images in the data base makes feature selection techniques such as forward selection or lasso regression become inadequate. In the experimental section, the performance of the elastic net model is compared with geometrical and color based algorithms widely used in face recognition such as Procrustes nearest neighbor, Eigenfaces, or Fisherfaces. Results show that the elastic net is capable of selecting a set of discriminative features and hereby obtain higher classification rates.

The work presented here takes place in the field of computer aided analysis of facial expressions displayed in sign language videos. We use Active Appearance Models to model a face and its variations of shape and texture caused by expressions. The algorithm is used to accurately fit an AAM to the face seen on each video frame. In the context of sign language communication, the signer’s face is frequently occluded, mainly by hands. A facial expression tracker has then to be robust to occlusions. We propose to rely on a robust variant of the AAM fitting algorithm to explicitly model the noise introduced by occlusions. Our main contribution is the automatic detection of hand occlusions. The idea is to model the behavior of the fitting algorithm on unoccluded faces, by means of residual image statistics, and to detect occlusions as being what is not explained by this model. We use residual parameters with respect to the fitting iteration , the AAM distance to the solution, which greatly improves occlusion detection compared to the use of fixed parameters. We also propose a robust tracking strategy used when occlusions are too important on a video frame, to ensure a good initialization for the next frame.

This paper deals with model based regularization of velocity encoded cardiac magnetic resonance images (MRI). We extend upon an existing spatiotemporal model of cardiac kinematics by considering data certainty and regularity of the model in order to improve its performance. The method was evaluated using a computer simulated phantom and using in vivo gridtag MRI as gold standard. We show, both quantitatively and qualitatively, that our modified model performs better than the original one.

In this paper it is shown that there is a difference in local fractal dimension between imaged glandular tissue, supporting tissue and muscle tissue based on an assessment from a mammogram. By estimating the density difference at four different local dimensions (2.06, 2.33, 2.48, 2.70) from 142 mammograms we can define a measure and by using this measure we are able to distinguish between the tissue types. A ROC-analysis gives us an area under the curve-value of 0.9998 for separating glandular tissue from muscular tissue and 0.9405 for separating glandular tissue from supporting tissue. To some extent we can say that the measured difference in fractal properties is due to different fractal properties of the unprojected tissue. For example, to a large extent muscle tissue seems to have different fractal properties than glandular or supportive tissue. However, a large variance in the local dimension densities makes it difficult to make proper use of the proposed measure for segmentation purposes.