Catálogo de publicaciones - libros

In this paper, we consider the problem of finding the most similar objects given a query object, in large multimedia datasets. We focus on scenarios where the similarity measure itself is not fixed, but is continuously being refined with user feedback. Conventional database techniques for efficient similarity search are not effective in this environment as they take a specific similarity/distance measure as input and build index structures tuned for that measure. Our approach works effectively in this environment as validated by the experimental study where we evaluate it over a wide range of datasets. The experiments show it to be efficient and scalable. In fact, on all our datasets, the response times were within a few seconds, making our approach suitable for interactive applications.

A music structured query language, called MuSQL, is presented in this paper. MuSQL consists of a schema definition sub-language and a data manipulation sub-language. The former is composed of schema-setup statements, schema-alter statements, and schema-drop statements. The latter is composed of selection, retrieval, extraction, insertion, update, deletion, commission, rollback, and other statements. MuSQL can be used to cut, delete and merge content of music, insert, delete and extract features of music, and exactly or approximately search music pieces, especially in the processing of music based on content. Also, it makes some music processing operations easier due to its built-in semantics. MuSQL has been implemented in a music data management system.

Real time rendering of animated crowds has many practical multimedia applications. The Graphics Processor Unit (GPU) is being increasingly employed to accelerate associated rendering and deformation calculations. This paper explores skeletal deformation calculations on the GPU for crowds of articulated figures. It compares a few strategies for efficient reuse of such calculations on clones. We further propose ideas that will reduce chances of detecting such duplication. The system has been implemented for modern PCs with Graphics Accelerator cards that support GPU Shader Model 3.0, and come with accelerated bi-directional PCI express bus communication. We have achieved a realistic crowd population of 1000 animated humans at interactive rates.

Over the years, multimedia applications are getting increasingly more complex and large in scale. Multimedia Web Service is identified as one of the possible solutions to meet the challenges. The advantages of using Web Services are ease of application development, adaptive to changes, fault tolerance and etc. In the paper, a sample tracking application will be discussed and developed using multimedia Web Services (multimediaWS) approach. Throughout the paper, we will suggest some general rules on designing the multimediaWS as well as evaluate the pros and cons of using multimediaWS for multimedia application.

The problem of automatic melody line identification in a MIDI file plays an important role towards taking QBH systems to the next level. We present here, a novel algorithm to identify the melody line in a polyphonic MIDI file. A note pruning and track / channel ranking method is used to identify the melody line. We use results from musicology to derive certain simple heuristics for the note pruning stage. This helps in the robustness of the algorithm, by way of discarding “spurious” notes. A ranking based on the melodic information in each track / channel enables us to choose the melody line accurately. Our algorithm makes no assumption about MIDI performer specific parameters, is simple and achieves an accuracy of 97% in identifying the melody line correctly. This algorithm is currently being used by us in a QBH system built in our lab.

This paper proposes a robust point transferring method for markerless AR applications. Using this method, any points specified at the initialization stage can be stably transferred during the augmentation process. These transferred points can be used for registration, annotation and video augmentation in markerless AR applications. This proposed point transferring method is based on a simple nonlinear optimization model. The proposed method has several advantages. Firstly, it is robust and stable as it remains effective when the camera is moved about quickly or when the scenes are largely occluded or filled with moving objects. Second, it is simple as the points that will be used for registration, annotation and video augmentation are only required to be specified in one image. Lastly, it is fast as the proposed simple optimization model can be solved quickly. Several experiments have been conducted to validate the performance of this proposed method.

3D facial synthesis has been an intensive research topic in both image processing and computer graphics. So far common facial synthesizing methods were either statistic model based or laser range scanner (LRS) based. However, these methods could only provide approximate models. In this paper, we present a Magnetic Resonance Image (MRI) based 3D facial synthesizing approach, which can build accurate virtual faces for plastic surgery. This approach provides an database free modeling process so that the precision will not be limited by the face database. During modeling, we combine Sobel filter with snake algorithm to extract the 3D facial surface from noised MRI and build Bezier facial model, which facilitates the local geometric control. Furthermore, the proposed photo mapping algorithm help add realistic texture to the model. It avoids the complex and time-consuming texture fitting process while achieves satisfying result.

In this paper, we present a 3-D modeling for analyzing court-net sports videos, which enables to map points in the real-world coordinates to the image coordinates. To this end, we propose a two-step algorithm to extract the feature lines and points from two perpendicular planes (ground and net plane) for determining the camera calibration parameters. In the first step, we bridge the gap between the 2-D standard court model and the image coordinate system that is described by a plane-to-plane mapping. With this mapping, it is possible to distinguish the feature lines like court lines in the ground plane. The second step is to detect the net line located in the net plane, where the line is classified as net line if it passes several tests. The feature points at well-known positions within these two planes (e.g. intersection of two lines) are utilized to calculate the camera calibration parameters. We demonstrate the performance of the proposed algorithm by evaluating it for a variety of court-net sports videos including badminton, tennis and volleyball. Results show that the algorithm is robust to partial court views or bad lighting conditions, and can be applied to various applications.

Global common illumination between real and virtual objects is the process of illuminating scenes and objects with images of light from the real world. After including the virtual objects, the resulting scene should have consistent shadow configuration. This paper presents a novel algorithm that integrates synthetic objects in the real photographs by using the radiosity on graphics processing unit (GPU) and high dynamic range (HDR) radiance map. In order to reconstruct 3D illumination environment of the scene, we estimate the camera model and the extrinsic parameters from omni-directional images. The simulation results showed that our method can generate photo-realistic images.

In the paper, the detection of IR weak and small targets is investigated in natural background based on fractal features. One feature of multi-scale variance ratio of fractal surface is proposed according to the fact that the fractal feature of man made objects changes shaper than the natural background. The new feature stands out the artificial objects much better from natural background than what can be done by fractal dimension feature or fractal model fit error feature, thus inhibiting background clutters well. Local gray histogram statistics is applied to object detection in the images with feature of multi-scale variance ratio of fractal surface. Experimental results shows that the detecting algorithm based on such a feature can localize weak and small objects stably in a single-frame image, and is a effective algorithm.