Catálogo de publicaciones - libros

The rendering process of graphics rendering pipeline is usually completed by both the CPU and the GPU, and a bottleneck can be located either in the CPU or the GPU. This paper focuses on reducing the bottleneck between the CPU and the GPU. We are proposing a method for improving the performance of parallel processing for realtime graphics rendering by separating the CPU operations into two parts: pure CPU operations and operations related to the GPU, and let them operate in parallel. This allows for maximizing the parallelism in processing the communication between the CPU and the GPU. Some experiments lead us to confirm that our method proposed in this paper can allow for faster graphics rendering. In addition to our method of using a dedicated thread for GPU related operations, we are also proposing an algorithm for balancing the graphics pipeline using the idle time due to the bottleneck. We have implemented the two methods proposed in this paper in our networked 3D game engine and verified that our methods are effective in real systems.

A user oriented grid platform for multidisciplinary design optimization (MDO) of aero-crafts based on web service highly requires to construct a tool for visualization of the computing data and visual steering of the whole process for MDO. In this paper, a distributed data visualization tool for MDO of Aero-crafts is described. And the visual steering scheme for MDO is described in detail, which is constructed under web service environment and is performed as a series of web pages. Visualization Toolkit (VTK) and Java are adopted in visualization service to process the results of MDO of Geometry and the distributed computational data.

In a video database, the similarity between video sequences is usually measured by the percentages of similar frames shared by both video sequences, where each frame is represented as a high-dimensional feature vector. The direct computation of the similarity measure involves time-consuming sequential scans over the whole dataset. On the other hand, adopting existing indexing technique to high-dimensional datasets suffers from the “Dimensionality Curse”. Thus, an efficient and effective indexing method is needed to reduce the computation cost for the similarity search. In this paper, we propose a Multi-level Hierarchical Divisive Dimensionality Reduction technique to discover correlated clusters, and develop a corresponding indexing structure to efficiently index the clusters in order to support efficient similarity search over video data. By using dimensionality reduction techniques as Principal Component Analysis, we can restore the critical information between the data points in the dataset using a reduced dimension space. Experiments show the efficiency and usefulness of this approach.

This paper presents the concept of progressive filtering (PF) and its efficient design based on dynamic programming. The proposed PF is scalable for large music retrieval systems and is data-driven for performance optimization. Moreover, its concept and design are general in nature and can be applied to any multimedia retrieval systems. The application of the proposed PF to a 5-stage query-by-singing/humming (QBSH) system is reported, and the experimental results demonstrate the feasibility of the proposed approach.

Cross-media retrieval is an interesting research problem, which seeks to breakthrough the limitation of modality so that users can query multimedia objects by examples of different modalities. In order to cross-media retrieve, the problem of similarity measure between media objects with heterogeneous low-level features needs to be solved. This paper proposes a novel approach to learn both intra- and inter-media correlations among multi-modality feature spaces, and construct MLE semantic subspace containing multimedia objects of different modalities. Meanwhile, relevance feedback strategies are developed to enhance the efficiency of cross-media retrieval from both short- and long-term perspectives. Experiments show that the result of our approach is encouraging and the performance is effective.

Quality fluctuation has a major negative effect on perceptive video quality. Many recent video quality smoothing works target on constant distortion (i.e., constant PSNR) throughout the whole coded video sequence. In [1], a target distortion was set up for each frame based on a hypothesis that maintaining constant distortion over frames would boast video quality smoothing and extensive experiments showed the constant-distortion bit allocation (CDBA) scheme significantly outperforms the popular constant bit allocation (CBA) scheme and Xie et al’s recent work [2, 3] in terms of delivered video quality. But during the scene changes, it has been observed that the picture energy often dramatically changes. Maintaining constant PSNR would result in dramatically different SNR performance and translate into dramatically different perceptive effects. Although computationally more complex, SNR represents a more objective measure than PSNR in assessing video quality. In this paper, a single-pass frame-level constant-SNR bit allocation scheme (CSNRBA) is developed for video quality smoothing throughout the video sequence. To achieve constant-SNR, a power series weighted actual SNR average of previous coded frames is adopted as the target SNR for the current frame. From the target SNR, the target distortion for the current frame is calculated. Then according to the analytic close-form model and the linear rate control algorithm, the bit budget for the current frame can be estimated. Experimental results show that the proposed CSNRBA scheme provides much smoother video quality and achieve much better subjective video quality in terms of natural color, sharp objects and silhouette significantly on all testing video sequences than both CBA and CDBA schemes.

A method of shadow removal from sole uncalibrated outdoor image is proposed. Existing approaches usually decompose the image into albedo and illumination images, in this paper, based on the mechanism of shadow generation, the occlusion factor is introduced, and the illumination image is further decomposed as the linear combination of solar irradiance and ambient irradiance images. The involved irradiance are achieved from the user-supplied hints. The shadow matte are evaluated by the anisotropic diffusion of posterior probability. Experiments show that our method could simultaneously extract the detailed shadow matte and recover the texture beneath the shadow.

In this paper, the 3D head model classification problem is addressed by use of a newly developed subspace analysis method: kernel clustering-based discriminant analysis or KCDA as an abbreviation. This method works by first mapping the original data into another high-dimensional space, and then performing clustering-based discriminant analysis in the feature space. The main idea of clustering-based discriminant analysis is to overcome the Gaussian assumption limitation of the traditional linear discriminant analysis by using a new criterion that takes into account the multiple cluster structure possibly embedded within some classes. As a result, Kernel CDA tries to get through the limitations of both Gaussian assumption and linearity facing the traditional linear discriminant analysis simultaneously. A novel application of this method in 3D head model classification is presented in this paper. A group of tests of our method on 3D head model dataset have been carried out, reporting very promising experimental results.

It is generally agreed that traditional transcoding involves complex computations, which may introduce substantial additional delay to content delivery. Inspired by new multimedia data formats, like JPEG 2000, a new adaptation called modulation is devised. Unlike transcoding, modulation is fast since it basically generates an object’s representation by selecting fragments of the object without decoding/encoding it. In this paper, a framework for pervasive Web content delivery is proposed to exploit the modulation’s benefits.

In order to reduce the gap between low-level image features and high-level image semantics, various long term learning strategies were integrated into content-based image retrieval system. The strategies always use the semantic relationships among images to improve the effectiveness of the retrieval system. This paper proposes a semantic similarity propagation method to mine the hidden semantic relationships among images. The semantic relationships are propagated between the similar images and regions. Experimental results verify the improvement on similarity propagation and image retrieval.