Catálogo de publicaciones - libros

We present an efficient occlusion culling algorithm for interactive rendering of large complex virtual scene with high depth complexity. Our method exploits both spatial and temporal coherence of visibility. A space hierarchy of scene is constructed and its nodes are rendered in an approximate front-to-back order. Nodes in view frustum are inserted into one of layered node lists, called layered buffers(), according to its distance to the view point. Each buffer in the LBs is rendered with hardware occlusion queries. Using a visibility predictor() for each node and interleaving occlusion queries with rendering, we reduce the occlusion queries count and graphics pipeline stalls greatly. This occlusion culling algorithm can work in a conservative way for high image quality rendering or in an approximate way for time critical rendering. Experimental results of different types of virtual scene are provided to demonstrate its efficiency and generality.

A Free Form Projection Display using a distortion-eliminating algorithm was constructed. In the system, an object in the curved surface is used and the virtual object is represented by image projection. The virtual object is shown as if it fixed in the curved screen surface and gives the user a sense of motion parallax. Using this construction, the virtual object can be handled directly by the user’s hand and examined.

This system seems to be effective in manual operation and shape recognition of virtual object.

In order to evaluate the effectivity, we performed the experiments to compare FFPD and conventional flat display monitor with mouse from the viewpoint of depth perception and manual operation.

Content adaptation techniques for progressive delivery of 3D models to clients with different rendering capabilities and network connections has been extensively studied in recent years. In this paper, a client-server based novel service-oriented architecture for 3D content delivery and adaptive rendering is presented. The architecture is integrated into the service-oriented architecture (SOA) framework, and it is designed to enhance the client user experience by progressively delivering 3D content stream quickly, reliably, and with high quality based on the adaptive requirements. At the client end the stream is progressively decoded up to the “best” level-of-details as defined by the client computational resources and the user inclination.

An adaptive watermarking algorithm operating in the ridgelet domain is proposed. Since the most significant coefficients of the ridgelet transform (RT) can represent the most energetic direction of an image with straight edge, the image is first partitioned into small blocks and RT is applied for each block. Followed by the multiplicative rule, the watermark sequence is casting into local linear singularity coefficients within the highest energy direction of each block. Through analyzing the distribution of the texture in ridgelet coefficients of each block, the feature of luminance masking and texture masking is incorporated to adjust the watermark’s embedding strength. Then the embedded watermark can be blindly detected by correlation detector. Experiments show that the proposed algorithm can achieve a better tradeoff between the robustness and transparency.

This paper extends the upright face detection framework proposed by Viola et al. 2001 to handle in-plane rotated faces. These haar-like features work inefficiently on rotated faces, so this paper proposes a new set of ±26.565 ° haar-like features which can be calculated quickly to represent the features of rotated faces. Unlike previous face detection techniques in training quantities of samples to build different rotated detectors, with these new features, we address to build different rotated detectors by rotating an upright face detector directly so as to achieve in-plane rotated face detection. This approach is selected because of its computational efficiency, simplicity and training time saving. This proposed method is tested on CMU-MIT rotated test data and yields good results in accuracy and maintains speed advantage.

Contour grouping is an important issue in computer vision. However, traditional ways tackling the problem usually fail to provide as satisfying results as human vision can do. One important feature of human vision mechanism is that human vision tends to group together edges that are not only geometrically and topologically related, but also similar in their appearances – the appearances of image patches around them including their brightness, color, texture cues, etc. But in traditional grouping approaches, after edges or lines have been detected, the appearances of image patches around them are seldom considered again, leading to the results that edges belonging to boundaries of different objects are sometimes falsely grouped together. In this paper, we introduce an to describe the appearance of an image patch around a line segment, and incorporate this appearance feature into a to evaluate contours on an image. The most salient contour is found by optimizing this saliency measure using a genetic algorithm. Experimental results prove the effectiveness of our approach.

Presented here is a discussion of the techniques required to create stereoscopic panoramic images. Such images allow interactive exploration of 3D environments with stereoscopic depth cues. If projected in a surround display environment they can engage the two characteristics of the human visual system responsible for immersion, namely stereopsis and peripheral vision.

It is difficult to deal with large datasets by kernel based methods since the number of basis functions required for an optimal solution equals the number of samples. We present an approach to build a sparse kernel classifier by adding constraints to the number of support vectors and to the classifier function. The classifier is considered on Riemannian manifold. And the sparse greedy learning algorithm is used to solve the formulated problem. Experimental results over several classification benchmarks show that the proposed approach can reduce the training and runtime complexities of kernel classifier applied to large datasets without scarifying high classification accuracy.

Chaotic optimization is a new optimization technique. For image segmentation, conventional chaotic sequence is not very effective to three-dimension gray histogram. In order to solve this problem, a three-dimension chaotic sequence generating method is presented. Simulation results show that the generated sequence is pseudorandom and its distribution is approximately inside a sphere whose centre is (0.5 , 0.5 , 0.5). Based on this work, we use the proposed chaotic sequence to optimize three-dimension maximum between-variance image segmentation method. Experiments results show that our method has better segmentation effect and lower computation time than that of the original three-dimension maximum between-variance image segmentation method for mixed noise disturbed image.

In this paper, a registration method based on Harris corners is proposed. It is composed of three steps. First, corner extraction and matching. We use the gray level information around the corner to setup the correspondences, then use the affine invariant of Mahalannobis distance to remove the mis-matched corner points. From this correspondence of the corner points, the affine matrix between two different images can be determined. Finally, map all points in the sensed image to the reference using the estimated transformation matrix and assign the corresponding gray level by re-sampling the image in the sensed image. Experiments with different types of multi-sensor images demonstrated the feasibility of our method.