Catálogo de publicaciones - libros

A multi-modal method to improve the performance of the anchorperson shot detection for news story segmentation is proposed in this paper. The anchorperson voice information is used for the verification of anchorperson shot candidates extracted by visual information. The algorithm starts with the anchorperson voice shot candidate extraction using time and silence condition. The anchorperson templates are generated from the anchorperson face and cloth information from the anchorperson voice shots extracted. The anchorperson voice models are then created after segregating anchorperson voice shots containing 2 or more voices. The anchorperson voice model verifies the anchorperson shot candidates obtained from visual information. 720 minutes of news programs are tested and experimental results are demonstrated.

Dialogue sequences constitute an important part of any movie or television program and their successful detection is an essential step in any movie summarisation/indexing system. The focus of this paper is to detect of dialogue, rather than complete scenes. We argue that these shorter sequences are more desirable as retrieval units than temporally long scenes. This paper combines various audiovisual features that reflect accepted and well know film making conventions using a selection of machine learning techniques in order to detect such sequences. Three systems for detecting dialogue sequences are proposed: one based primarily on audio analysis, one based primarily on visual analysis and one that combines the results of both. The performance of the three systems are compared using a manually marked-up test corpus drawn from a variety of movies of different genres. Results show that high precision and recall can be obtained using low-level features that are automatically extracted.

A system for video retrieval from a ubiquitous environment is presented. Data from pressure-based floor sensors are used as a supplementary input for retrieving video from a large number of cameras. An algorithm based on agglomerative hierarchical clustering is used to segment footpaths of individual persons. Video handover is proposed and two methods are implemented to retrieve video clips and key frame sequences showing a person moving inside the house. The video clips are further segmented according to the actions performed. We evaluate the performance of each stage of retrieval and present the results. The paper concludes with suggestions for improvements, and future directions.

We define a style similarity measure for video documents based on the localization of common elements and on the temporal order in which they appear in each document. Common elements for a couple of compared videos are segments presenting similar behaviors on a subset of low or mid level features extracted for the comparison process. We propose a method to compare two video documents and to extract those similar elements using dynamic programming and one-dimensional morphological operations. The similarity measure is applied on TV-news broadcast to illustrate its behavior.

We introduce a view–point invariant representation of moving object trajectories that can be used in video database applications. It is assumed that trajectories lie on a surface that can be locally approximated with a plane. Raw trajectory data is first locally–approximated with a cubic spline via least squares fitting. For each sampled point of the obtained curve, a projective invariant feature is computed using a small number of points in its neighborhood. The resulting sequence of invariant features computed along the entire trajectory forms the view–invariant descriptor of the trajectory itself. Time parametrization has been exploited to compute cross ratios without ambiguity due to point ordering. Similarity between descriptors of different trajectories is measured with a distance that takes into account the statistical properties of the cross ratio, and its symmetry with respect to the point at infinity. In experiments, an overall correct classification rate of about 95% has been obtained on a dataset of 58 trajectories of players in soccer video, and an overall correct classification rate of about 80% has been obtained on matching partial segments of trajectories collected from two overlapping views of outdoor scenes with moving people and cars.

In this study, we propose a new dynamic warping algorithm for cyclic sequence comparison, which approximate the optimal solution efficiently. The comparison of two sequences, whose starting points are known, is performed by finding the optimal correspondence between their elements, which minimize the distance. If the sequences are cyclic and their starting points are not known, the alignment computation must determine the amount of cyclic shift for the optimal solution. However, this process increases the complexity of the algorithm and may be cumbersome especially for large databases. Instead of finding the optimal solution, the proposed algorithm finds the approximate distance at once and decreases the time complexity substantially. The algorithm is tested in boundary based shape similarity problem. The experiments performed on MPEG-7 Shape database, show that the proposed method performs better than the classical cyclic string comparison methods in the literature and gives very similar results with the optimal solution.

With the advent of Storage Area Networks (SAN) and Network Attached Disks (NASD), a new trend in storage systems design is to move disks from behind storage servers and attach them to a SAN for direct client access. Such designs remove storage servers from the path of data transfer leading to highly scalable file systems. While there exists many file systems built using such direct attached disks, they are all optimized for traditional text-based data and are not well-suited for streaming continuous media, i.e., audio and video data. In this paper we present the architectural details of a scalable distributed continuous media file system built using SAN-attached disks. We describe the implementation details of our Linux-based prototype and show by experimention that the performance of our file system scales linearly with the number of disks and the number of clients, and the file system provides strict bandwidth guarantees for continuous media streams.

Recently, a digital document takes place of a paper document. However, because the paper document has many advantages, people prefer the paper document to digital document as before. The Advanced Paper Document using Projection Display System has advantages of both off-line documents and on-line documents. In this paper, we propose the Advanced Document Authoring Tool (ADAT), which can insert, delete, and modify on-line information to the off-line document. Users can attach on-line information to the off-line document using ADAT like that the user writes on the paper document with a pencil. Therefore this system provides a natural and intuitive environment to the user. As shown by experimental results, the proposed the ADAT is applicable to provide an interactive computing environment using PDS.

Keypics are graphical metadata intended for indexing of images on the Internet. They are conceived as hand-drawn sketches, not restricted to a definite set. An obvious difficulty when dealing with keypics is that they elude rigid geometric treatment.

A proposal of solution comes from Size Functions. This paper is the report of a complete experiment on 494 keypics with Size Functions based on three measuring functions (distances, projections and jumps) and their combination.

3D reconstruction over long sequences has been to the main problem of computer vision. Projective reconstruction is known to be an important process for 3D reconstruction in Euclidean space. In this paper, we present a new projective reconstruction algorithm using invariant properties of the line segments in projective space: collinearity, order of contact, and intersection. Points on each line segment in the image are reconstructed in projective space, and we calculate the best-fit 3D line from them by Least-Median-Squares (LMedS). Our method regards the points unsatisfying collinearity as outliers, which are caused by false feature detection and tracking. In addition, both order of contact and intersection in projective space are considered. By using the points that are the orthogonal projection of outliers onto the 3D line, we iteratively obtain more precise projective matrix than the previous method. The experimental results showed that the proposed algorithm can estimate camera parameters and reconstruct 3D model exactly.