Catálogo de publicaciones - libros

We present a markerless real-time, model-based 3D face tracking methodology. The system combines two robust and complimentary op-timization-based strategies, namely active contours and mutual information template matching, in order to obtain real-time performances for full 6dof tracking. First, robust head contour estimation is realized by means of the Contracting Curve Density algorithm, effectively employing local color statistics separation for contour shape optimization. Afterwards, the 3D face template is robustly matched to the underlying image, through fast mutual information optimization. Off-line model building is done using a fast modeling procedure, providing a unique appearance model for each user. Re-initialization criteria are employed in order to obtain a complete and autonomous tracking system.

In this paper a new extension of the CONDENSATION algorithm, with application to infants face tracking, will be introduced. In this work we address the problem of tracking a face and its features in baby video sequences. A mixed state particle filtering scheme is proposed, where the distribution of observations is derived from an active appearance model. The mixed state approach combines several dynamic models in order to account for different occlusion situations. Experiments on real video show that the proposed approach augments the tracker robustness to occlusions while maintaining the computational time competitive.

Optical motion capture can be classified as an inference problem: given the data produced by a set of cameras, the aim is to extract the hidden state, which in this case encodes the posture of the subject’s body. Problems with motion capture arise due to the multi-modal nature of the likelihood distribution, the extremely large dimensionality of its state-space, and the narrow region of support of local modes. There are also problems with the size of the data and the difficulty with which useful visual cues can be extracted from it, as well as how informative these cues might be. Several algorithms exist that use stochastic methods to extract the hidden state, but although highly parallelisable in theory, such methods produce a heavy computational overhead even with the power of today’s computers. In this paper we assume a set of pre-calibrated cameras and only extract the subject’s silhouette as a visual cue. In order to describe the 2D silhouette data we define a 2D model consisting of conic fields. The resulting likelihood distribution is differentiable w.r.t. the state, meaning that its global maximum can be located fast using gradient ascent search, given manual initialisation at the first frame. In this paper we explain the construction of the model for tracking a human hand; we describe the formulation of the derivatives needed, and present initial results on both real and simulated data.

We present an approach for tracking human body parts and classification of human actions. We introduce Gaussian Processing Annealed Particle Filter Tracker (GPAPF), which is an extension of the annealed particle filter tracker and uses Gaussian Process Dynamical Model (GPDM) in order to reduce the dimensionality of the problem, increase the tracker’s stability and learn the motion models. Motion of human body is described by concatenation of low dimensional manifolds which characterize different motion types. The trajectories in the latent space provide low dimensional representations of sequences of body poses performed during motion. Our approach uses these trajectories in order to classify human actions. The approach was checked on HumanEva data set as well as on our own one. The results and the comparison to other methods are presented.

We introduce a framework for object detection and tracking in video of natural outdoor scenes based on fast per-frame segmentations using Felzenszwalb’s graph-based algorithm. Region boundaries obtained at multiple scales are first temporally filtered to detect stable structures to be considered as object hypotheses. Depending on object type, these are then classified using appearance characteristics such as color and texture and geometric attributes derived from the Hough transform. We describe preliminary results on image sequences taken from low-flying aircraft in which object categories are relevant to UAVs, consisting of sky, ground, and navigationally-useful ground features such as roads and pipelines.

This paper presents a lossy compression method for motion capture data. Each degree of freedom of a motion clip is smoothed by an anisotropic diffusion process and then divided into segments at feature discontinuities. Feature discontinuities are identified by the zero crossings of the second derivative in the smoothed data. Finally, each segment of each degree of freedom is approximated by a cubic Bézier curve. The anisotropic diffusion process retains perceptually important high-frequency parts of the data, including the exact location of discontinuities, while smoothing low-frequency parts of the data. We propose a hierarchical coding method to further compress the cubic control points. We compare our method with wavelet compression methods, which have the best compression rates to date. Experiments show that our method, relative to this work, can achieve about a 65% higher compression rate at the same approximation level.

This paper presents a new image-space algorithm for real-time collision detection, where the GPU computes the potentially colliding sets, and the CPU performs the standard triangle/triangle intersection test. The major strengths of the proposed algorithm can be listed as follows: it can handle dynamic models including deforming and fracturing objects, it can take both closed and open objects, it does not require any preprocessing but is quite efficient, and its accuracy is proportional to the visual sensitivity or can be controlled on demand. The proposed algorithm would fit well to real-time applications such as 3D games.

The ability to recognize the strokes drawn by the user, is central to most sketch-based interfaces. However, very few solutions that rely on recognition are robust enough to make sketching a definitive alternative to traditional WIMP user interfaces. In this paper, we propose an approach based on classification that given an unconstrained sketch, can robustly assign a label to each stroke that comprises the sketch. A key contribution of our approach is a technique for grouping strokes that eliminates outliers and enhances the robustness of the classification. We also propose a set of features that capture important attributes of the shape and mutual relationship of strokes. These features are statistically well-behaved and enable robust classification with Support Vector Machines (SVM). We conclude by presenting a concrete implementation of these techniques in an interface for driving facial expressions.

A new method for constructing a smooth surface that interpolates the vertices of an arbitrary mesh is presented. The mesh can be open or closed. Normals specified at vertices of the mesh can also be interpolated. The interpolating surface is obtained by locally adjusting the limit surface of the given mesh (viewed as the control mesh of a Catmull-Clark subdivision surface) so that the modified surface would interpolate all the vertices of the given mesh. The local adjustment process is achieved through locally blending the limit surface with a surface defined by non-uniform transformations of the limit surface. This local blending process can also be used to smooth out the shape of the interpolating surface. Hence, a process is not needed in the new method. Because the interpolation process does not require solving a system of linear equations, the method can handle meshes with large number of vertices. Test results show that the new method leads to good interpolation results even for complicated data sets. The new method is demonstrated with the Catmull-Clark subdivision scheme. But with some minor modification, one should be albe to apply this method to other parametrizable subdivision schemes as well.

In industrial pattern reproduction, clustered-dot screens are usually created to transform continuous tone image into halftone image for batch printing. But the algorithms generating clustered-dot screens are usually difficult to process large image because they are very slowly and need lot of memory. In addition, the generated halftone image often have periodic patterns, leading to poor tone reproduction. In this paper, a GPU-based algorithm for building stochastic clustered-dot screens is proposed. In the algorithm, after stochastically laying screen dot centers within a large dither matrix, Voronoi diagram is constructed to obtain the region of each screen dot, which is implemented with GPU. Then, each screen dot’s region is filled to get the stochastic clustered-dot screens, where a better gray density filling method that can be implemented easily on GPU is used. Experiments show the method can generate screens faster and with less memory than traditional algorithms. Moreover, in a halftone image generated by our method, the details and highlight part can be better expressed.