Search

This paper presents a new method for viewpoint invariant pedestrian recognition problem. We use a metric learning framework to obtain a robust metric for large margin nearest neighbor classification with rejection (i.e., classifier will return no matches if all neighbors are beyond a certain distance). The rejection condition necessitates the use of a uniform threshold for a maximum allowed distance for deeming a pair of images a match. In order to handle the rejection case, we propose a novel cost similar to the Large Margin Nearest Neighbor (LMNN) method and call our approach Large Margin Nearest Neighbor with Rejection (LMNN-R). Our method is able to achieve significant improvement over previously reported results on the standard Viewpoint Invariant Pedestrian Recognition (VIPeR [1]) dataset. [ABSTRACT FROM AUTHOR]

Effective object localization relies on efficient and effective searching method, and robust image representation and learning method. Recently, the Gaussianized vector representation has been shown effective in several computer vision applications, such as facial age estimation, image scene categorization and video event recognition. However, all these tasks are classification and regression problems based on the whole images. It is not yet explored how this representation can be efficiently applied in the object localization, which reveals the locations and sizes of the objects. In this work, we present an efficient object localization approach for the Gaussianized vector representation, following a branch-and-bound search scheme introduced by Lampert et al. [5]. In particular, we design a quality bound for rectangle sets characterized by the Gaussianized vector representation for fast hierarchical search. This bound can be obtained for any rectangle set in the image, with little extra computational cost, in addition to calculating the Gaussianized vector representation for the whole image. Further, we propose incorporating a normalization approach that suppresses the variation within the object class and the background class. Experiments on a multi-scale car dataset show that the proposed object localization approach based on the Gaussianized vector representation outperforms previous work using the histogram-of-keywords representation. The within-class variation normalization approach further boosts the performance. This chapter is an extended version of our paper at the 1st International Workshop on Interactive Multimedia for Consumer Electronics at ACM Multimedia 2009 [16]. [ABSTRACT FROM AUTHOR]

Emotion recognition from facial images is a very active research topic in human computer interaction (HCI). However, most of the previous approaches only focus on the frontal or nearly frontal view facial images. In contrast to the frontal/nearly-frontal view images, emotion recognition from non-frontal view or even arbitrary view facial images is much more difficult yet of more practical utility. To handle the emotion recognition problem from arbitrary view facial images, in this paper we propose a novel method based on the regional covariance matrix (RCM) representation of facial images. We also develop a new discriminant analysis theory, aiming at reducing the dimensionality of the facial feature vectors while preserving the most discriminative information, by minimizing an estimated multiclass Bayes error derived under the Gaussian mixture model (GMM). We further propose an efficient algorithm to solve the optimal discriminant vectors of the proposed discriminant analysis method. We render thousands of multi-view 2D facial images from the BU-3DFE database and conduct extensive experiments on the generated database to demonstrate the effectiveness of the proposed method. It is worth noting that our method does not require face alignment or facial landmark points localization, making it very attractive. [ABSTRACT FROM AUTHOR]

This paper presents a non-local kernel regression (NL-KR) method for image and video restoration tasks, which exploits both the non-local self-similarity and local structural regularity in natural images. The non-local self-similarity is based on the observation that image patches tend to repeat themselves in natural images and videos; and the local structural regularity reveals that image patches have regular structures where accurate estimation of pixel values via regression is possible. Explicitly unifying both properties, the proposed non-local kernel regression framework is robust and applicable to various image and video restoration tasks. In this work, we are specifically interested in applying the NL-KR model to image and video super-resolution (SR) reconstruction. Extensive experimental results on both single images and realistic video sequences demonstrate the superiority of the proposed framework for SR tasks over previous works both qualitatively and quantitatively. [ABSTRACT FROM AUTHOR]

This paper introduces a new framework for image classification using local visual descriptors. The pipeline first performs a nonlinear feature transformation on descriptors, then aggregates the results together to form image-level representations, and finally applies a classification model. For all the three steps we suggest novel solutions which make our approach appealing in theory, more scalable in computation, and transparent in classification. Our experiments demonstrate that the proposed classification method achieves state-of-the-art accuracy on the well-known PASCAL benchmarks. [ABSTRACT FROM AUTHOR]

Facial image computing has been an extensively studied topic since its wide applications in human-centered computing, human-avatar interaction, virtual reality, and multimedia communication. Successful systems have been equipped with realistic face models, efficient compression algorithms, reliable animation techniques and user friendly interaction schemes. In this chapter, we will mainly focus on techniques, algorithms, models, applications and real-world systems. Comprehensive summarization of the state-of-the-art works will be presented as well as our experiences and contributions in the field, especially several real prototype systems developed in our group, such as the online interactive gaming system hMouse, humanoid emotive audio-visual avatar, and 3D face/head tracking based video compression. Performances of these three systems are also illustrated based on standard evaluations. [ABSTRACT FROM AUTHOR]

The estimation of human age from face images has great potential in real-world applications. However, how to discover the intrinsic aging trend is still a challenging problem. In this work, we proposed a general distance metric learning scheme for regression problems, which utilizes not only data themselves, but also their corresponding labels to strengthen the credibility of distances. This metric could be learned by solving an optimization problem. Through the learned metric, it is easy to find the intrinsic variation trend of data by a relative small amount of samples without any prior knowledge of the structure or distribution of data. Furthermore, the test data could be projected to this metric by a simple linear transformation and it is easy to be combined with manifold learning algorithms to improve the performance. Experiments are conducted on the public available FG-NET database by gaussian process regression in the learned metric to validate our methods and the age estimation performance is improved over the traditional regression methods. [ABSTRACT FROM AUTHOR]

We describe a visual communication application for a dark, theater-like interactive virtual simulation training environment. Our system visually estimates and tracks the body position, orientation and the arm-pointing direction of the trainee. This system uses a near-IR camera array to capture images of the trainee from different angles in the dim-lighted theater. Image features like silhouettes and intermediate silhouette body axis points are then segmented and extracted from image backgrounds. 3D body shape information such as 3D body skeleton points and visual hulls can be reconstructed from these 2D features in multiple calibrated images. We proposed a particle-filtering based method that fits an articulated body model to the observed image features. Currently we focus on the arm-pointing gesture of either limb. From the fitted articulated model we can derive the position on the screen the user is pointing to. We use current graphic hardware to accelerate the processing speed so the system is able to work in real-time. The system serves as part of multi-modal user-input device in the interactive simulation. [ABSTRACT FROM AUTHOR]

This paper describes a simple yet effective method for achieving accurate, believable eye contact between humans and computer-generated characters, which to the author's knowledge is demonstrated here for the first time. A prototype system provides a high-fidelity stereoscopic head-tracked virtual environment, within which the user can engage in eye contact with a near-photorealistic virtual human model. The system does not require eye tracking. The paper describes design and implementation details, and reports qualitative positive feedback from initial testers. [ABSTRACT FROM AUTHOR]

We present an action recognition scheme that integrates multiple modality of cues that include shape, motion and depth to recognize human gesture in the video sequences. In the proposed approach we extend classification framework that is commonly used in 2D object recognition to 3D spatio-temporal space for recognizing actions. Specifically, a boosting-based classifier is used that learns spatio-temporal features specific to target actions where features are obtained from temporal patterns of shape contour, optical flow and depth changes occuring at local body parts. The individual features exhibit different strength and sensitivity depending on many factors that include action, underlying body parts and background. In the current method, the multiple cues of different modalities are combined optimally by fisher linear discriminant to form a strong feature that preserve strength of individual cues. In the experiment, we apply the integrated action classifier on a set of target actions and evaluate its performance by comparing with single cue-based cases and present qualitative analysis of performance gain. [ABSTRACT FROM AUTHOR]

Multi-camera networks offer potentials for a variety of novel human-centric applications through provisioning of rich visual information. In this paper, face orientation analysis and posture analysis are combined as components of a human-centered interface system that allows the user's intentions and region of interest to be estimated without requiring carried or wearable sensors. In pose estimation, image observations at the cameras are first locally reduced to parametrical descriptions, and Particle Swarm Optimization (PSO) is then used for optimization of the kinematics chain of the 3D human model. In face analysis, a discrete-time linear dynamical system (LDS), based on kinematics of the head, combines the local estimates of the user's gaze angle produced by the cameras and employs spatiotemporal filters to correct any inconsistencies. Knowing the intention and the region of interest of the user facilitates further interpretation of human behavior, which is the key to non-restrictive and intuitive human-centered interfaces. Applications in assisted living, speaker tracking, and gaming can benefit from such unobtrusive interfaces. [ABSTRACT FROM AUTHOR]

The analysis of facial expression temporal dynamics is of great importance for many real-world applications. Being able to automatically analyse facial muscle actions (Action Units, AUs) in terms of recognising their neutral, onset, apex and offset phases would greatly benefit application areas as diverse as medicine, gaming and security. The base system in this paper uses Support Vector Machines (SVMs) and a set of simple geometrical features derived from automatically detected and tracked facial feature point data to segment a facial action into its temporal phases. We propose here two methods to improve on this base system in terms of classification accuracy. The first technique describes the original time-independent set of features over a period of time using polynomial parametrisation. The second technique replaces the SVM with a hybrid SVM/Hidden Markov Model (HMM) classifier to model time in the classifier. Our results show that both techniques contribute to an improved classification accuracy. Modeling the temporal dynamics by the hybrid SVM-HMM classifier attained a statistically significant increase of recall and precision by 4.5% and 7.0%, respectively. [ABSTRACT FROM AUTHOR]

In this paper we address the correspondence problem, with its application to nonrigid tracking and unsupervised modelling, as a nonparametric, active-linking topology learning problem. Unlike existing soft competitive learning methods, Active Growing Neural Gas (A-GNG) has both global and local properties which allows part of the network to reconfigure while tracking. In addition, A-GNG uses a number of features (e.g. topographic product, local grey-level and map transformation) so that the topological relations are preserved and nodes correspondences are retained between tracked configurations. Experimental results in a sequence of hand gestures and artificial data have shown the superiority of our proposed method over the original GNG. [ABSTRACT FROM AUTHOR]

In this article, we present an object tracking system which allows interactive user feedback to improve the accuracy of the tracking process in real-time video. In addition, we describe the hybrid maximum likelihood similarity, which integrates traditional metrics with the maximum likelihood estimated metric. The hybrid similarity measure is used to improve the dynamic relevance feedback process between the human user and the objects detected by our system. [ABSTRACT FROM AUTHOR]

When interacting with wall-sized, high-resolution tiled displays, users typically stand or move in front of it rather than sit at fixed locations. Using a mouse to interact can be inconvenient in this context, as it must be carried around and often requires a surface to be used. Even for devices that work in mid-air, accuracy when trying to hit small or distal targets becomes an issue. Ideally, the user should not need devices to interact with applications on the display wall. We have developed a hybrid vision- and sound-based system for device-free interaction with software running on a 7x4 tile 220-inch display wall. The system comprises three components that together enable interaction with both distal and proximal targets: (i) A camera determines the direction in which a user is pointing, allowing distal targets to be selected. The direction is determined using edge detection followed by applying the Hough transform. (ii) Using four microphones, a user double-snapping his fingers is detected and located, before the selected target is moved to the location of the snap. This is implemented using correlation and multilateration. (iii) 16 cameras detect objects (fingers, hands) in front of the display wall. The 1D positions of detected objects are then used to triangulate object positions, enabling touch-free multi-point interaction with proximal content. The system is used on the display wall in three contexts to (i) move and interact with windows from a traditional desktop interface, (ii) interact with a whiteboard-style application, and (iii) play two games. [ABSTRACT FROM AUTHOR]

This paper presents an approach to large lexicon sign recognition that does not require tracking. This overcomes the issues of how to accurately track the hands through self occlusion in unconstrained video, instead opting to take a detection strategy, where patterns of motion are identified. It is demonstrated that detection can be achieved with only minor loss of accuracy compared to a perfectly tracked sequence using coloured gloves. The approach uses two levels of classification. In the first, a set of viseme classifiers detects the presence of sub-Sign units of activity. The second level then assembles visemes into word level Sign using Markov chains. The system is able to cope with a large lexicon and is more expandable than traditional word level approaches. Using as few as 5 training examples the proposed system has classification rates as high as 74.3% on a randomly selected 164 sign vocabulary performing at a comparable level to other tracking based systems. [ABSTRACT FROM AUTHOR]

We present a completely autonomous algorithm for the real-time creation of a moving subject's kinematic model from optical motion capture data and with no a priori information. Our approach solves marker tracking, the building of the kinematic model, and the tracking of the body simultaneously. The novelty lies in doing so through a unifying Markov random field framework, which allows the kinematic model to be built incrementally and in real-time. We validate the potential of this method through experiments in which the system is able to accurately track the movement of the human body without an a priori model, as well as through experiments on synthetic data. [ABSTRACT FROM AUTHOR]

Affective Computing, one of the frontiers of Human-Computer Interaction studies, seeks to provide computers with the capability to react appropriately to a user's affective states. In order to achieve the required on-line assessment of those affective states, we propose to extract features from physiological signals from the user (Blood Volume Pulse, Galvanic Skin Response, Skin Temperature and Pupil Diameter), which can be processed by learning pattern recognition systems to classify the user's affective state. An initial implementation of our proposed system was set up to address the detection of "stress" states in a computer user. A computer-based "Paced Stroop Test" was designed to act as a stimulus to elicit emotional stress in the subject. Signal processing techniques were applied to the physiological signals monitored to extract features used by three learning algorithms: Naïve Bayes, Decision Tree and Support Vector Machine to classify relaxed vs. stressed states. [ABSTRACT FROM AUTHOR]

Human-computer interaction (HCI) is one of the foremost challenges of our society. New paradigms for interacting with computers are being developed which will define the 21st century and enable the world to communicate and interact effortlessly and intuitively. In this short survey, we explain the major research clusters comprising the state of the art, and indicate promising future research directions. [ABSTRACT FROM AUTHOR]

We designed and implemented a vision-based projected tabletop interface for finger interaction. The system offers a simple and quick setup and economic design. The projection onto the tabletop provides more comfortable and direct viewing for users, and more natural, intuitive yet flexible interaction than classical or tangible interfaces. Homography calibration techniques are used to provide geometrically compensated projections on the tabletop. A robust finger tracking algorithm is proposed to enable accurate and efficient interactions using this interface. Two applications have been implemented based on this interface. [ABSTRACT FROM AUTHOR]

We present the architecture and algorithm design of a visual motion based perceptual interface for mobile devices with cameras. In addition to motion vector, we use the term "visual motion" to be any dynamic changes on consecutive image frames. In the lower architectural hierarchy, visual motion events are defined by identifying distinctive motion patterns. In the higher hierarchy, these visual events are used for interacting with user applications. We present an approach to context aware motion vector estimation to better tradeoff between speed and accuracy. It switches among a set of motion estimation algorithms of different speeds and precisions based on system context such as computation load and battery level. For example, when the CPU is heavily loaded or the battery level is low, we switch to a fast but less accurate algorithm, and vice versa. Moreover, to obtain more accurate motion vectors, we propose to adapt the search center of fast block matching methods based on previous motion vectors. Both quantitative evaluation of algorithms and subjective usability study are conducted. It is demonstrated that the proposed approach is very robust yet efficient. [ABSTRACT FROM AUTHOR]

In this paper we take a look at the predominant form of human computer interaction as used in image retrieval, called interactive search, and discuss a new approach called artificial imagination. This approach addresses two of the grand challenges in this field as identified by the research community: reducing the amount of iterations before the user is satisfied and the small sample problem. Artificial imagination will deepen the level of interaction with the user by giving the computer the ability to think along by synthesizing (‘imagining') example images that ideally match all or parts of the picture the user has in mind. We discuss two methods of how to synthesize new images, of which the evolutionary synthesis approach receives our main focus. [ABSTRACT FROM AUTHOR]

The advance of computing technology has provided the means for building intelligent vehicle systems. Drowsy driver detection system is one of the potential applications of intelligent vehicle systems. Previous approaches to drowsiness detection primarily make pre-assumptions about the relevant behavior, focusing on blink rate, eye closure, and yawning. Here we employ machine learning to datamine actual human behavior during drowsiness episodes. Automatic classifiers for 30 facial actions from the Facial Action Coding system were developed using machine learning on a separate database of spontaneous expressions. These facial actions include blinking and yawn motions, as well as a number of other facial movements. In addition, head motion was collected through automatic eye tracking and an accelerometer. These measures were passed to learning-based classifiers such as Adaboost and multinomial ridge regression. The system was able to predict sleep and crash episodes during a driving computer game with 96% accuracy within subjects and above 90% accuracy across subjects. This is the highest prediction rate reported to date for detecting real drowsiness. Moreover, the analysis revealed new information about human behavior during drowsy driving. [ABSTRACT FROM AUTHOR]

Parameter-embedded watermarking is effected through slight perturbations of parametric models of some deeply-integrated dynamics of a signal. This paper is concerned with particular model form, linear prediction (LP), which is naturally suited to the application of interest, speech watermarking. The focus of this paper is on the robustness performance of LP-embedded speech watermarking. It is shown that the technique is quite robust to a wide array of attacks including noise addition, cropping, compression, filtering, and others. In the LP formulation, a set-theoretic adjunct to the parameter embedding can be used to identify a watermark that is optimally robust against certain attacks, within a quantified fidelity constraint. [ABSTRACT FROM AUTHOR]

Audio classification is very important in multimedia retrieval such as audio indexing, analysis and content-based video retrieval. In this paper, we have proposed a clip-based support vector machine (SVM) approach to classify audio signals into six classes, which are pure speech, music, silence, environmental sound, speech with music and speech with environmental sound. The classification results are then used to partition a video into homogeneous audio segments, which is used to analyze and retrieve its higher-level content. The experimental results show that the proposed system not only improves classification accuracy, but also performs better than the other classification systems using the decision tree (DT), K Nearest Neighbor (K-NN) and Neural Network (NN). [ABSTRACT FROM AUTHOR]

Discrete Cosine Transform (DCT), which is employed by block-based hybrid video coding to encode motion prediction errors, has dominated practical video coding standards for several decades. However, DCT is only a good approximation to Principle Component Analysis (PCA, also called KLT), which is optimal among all unitary transformations. PCA is rejected by coding standards due to its complexity. This paper tries to use a matrix form of PCA (which we call tensor-PCA) to encode prediction errors in video coding. This method retains the performance of traditional PCA, but can be computed with much less time and space complexity. We compared tensor-PCA with DCT and GPCA in motion prediction error coding, which shows that it is a good trade-off between compression efficiency and computational cost. [ABSTRACT FROM AUTHOR]

The H.264 video coding standard provides considerably higher coding efficiency than those of previous standards but its complexity is significantly increased. In this paper, we propose an efficient method of fast mode decision by exploiting spatio-temporal correlation in H.264. Firstly, we select skip mode or inter mode by considering the temporal correlation. Secondly, we select variable block size on inter mode by considering the spatial correlation. Simulations show that the proposed method reduces the encoding time by 71% on average without any significant PSNR losses. [ABSTRACT FROM AUTHOR]

Data broadcasting services are required to provide user interactivity through connecting additional contents such as object information to audio-visual contents. H.264/AVC-based metadata authoring tools include functions which identify and track position and motion of objects. In this work, we propose a method for tracking the target object by using partially decoded texture data and motion vectors extracted directly from H.264/AVC bitstream. This method achieves low computational complexity and high performance through the dissimilarity energy minimization algorithm which tracks feature points adaptively according to these characteristics. The experiment has shown that the proposed method had high performance with fast processing time. [ABSTRACT FROM AUTHOR]

This paper concentrates on a unified 3-D camera modeling technique, which can be applied to analyze several sports types. To this end, we have based our modeling on collecting feature points from two perpendicular planes: the ground plane and the net plane, as they exist in most court-net sports. A two-step algorithm is employed to extract and distinguish the feature lines and points from these two planes for determining the camera calibration parameters. Our proposed modeling enables a defined mapping from real 3-D scene coordinates to image coordinates. The use of this modeling helps in the improvement of many emerging applications, such as moving-player segmentation and 3-D scene adaptation. We evaluate the promising performance of the proposed modeling for a variety of court-net sports videos containing badminton, tennis and volleyball, and also demonstrate its capability on the case study of player segmentation and 3-D scene adaptation. [ABSTRACT FROM AUTHOR]

A novel blind watermarking scheme is developed based on discrete wavelet transform (DWT) in this paper. In order to make the watermarking imperceptibility and robustness, the watermarking is embedded in the average of wavelet blocks using the visual model based on the human visual system (HVS). Low-pass wavelet coefficients n least significant bits (LSBs) are adjusted in concert with the average. Simulation results show that the proposed scheme is imperceptibility and robustness against many attacks such as JPEG compression, adding noise, rescaling, cropping, rotation, filtering, etc. [ABSTRACT FROM AUTHOR]

Face recognition has been addressed both in 2D, using still images or video sequences, and in 3D using three-dimensional face models. In this paper, we propose an original framework which provides a description capable to support 3D-3D face recognition as well as to directly compare 2D face images against 3D face models. This representation is extracted by measuring geodesic distances in 3D and 2D. In 3D, the geodesic distance between two points on a surface is computed as the length of the shortest path connecting the two points on the model. In 2D, the geodesic distance between two pixels is computed based on the differences of gray level intensities along the segment connecting the two pixels in the image. Experimental results are reported for 3D-3D and 2D-3D face recognition, in order to demonstrate the viability of the proposed approach. [ABSTRACT FROM AUTHOR]

As the amount of visual information is rapidly increasing, users want to find the more semantic information easily. Most retrieval systems by low-level features(such as color, texture) could not satisfy user's demand. To interpret semantic of image, many researchers use keywords as textual annotation. However, it's the image retrieval without ranking by text matching which is the simplest way to retrieval according to keyword's existence or nonexistence. In this paper, we propose conceptualization by similarity measure using relations among keywords for efficient image retrieval. We experiment annotated image retrieval by lowering the unrelated keyword's weight value and raising important keyword's one. [ABSTRACT FROM AUTHOR]

This paper proposes a scheme to detect and locate the players and the ball on the grass playfield in soccer videos. We put forward a shape analysis-based approach to identify the players and the ball from the roughly extracted foreground, which is obtained by a trained, color histogram-based playfield detector and connected component analysis. We employ Euclidean distance transform to extract skeletons for every foreground blob, and then perform shape analysis to remove false alarms (non-player and non-ball blobs) and cut-off the artifacts (mostly due to playfield lines) based on skeleton pruning and reverse Euclidean distance transform. Results are given to demonstrate the proposed algorithm works well in soccer video clips. [ABSTRACT FROM AUTHOR]

This paper presents a novel data organizing and delivery framework for P2P media streaming which is aware of media content. In this framework, the media data is partitioned GOP by GOP, and reorganized according to the frame-priorities within a GOP. All GOPs cached in the media buffer of a peer node can be scheduled concurrently, while the frames contained in each GOP are transmitted by order strictly according to their priorities. Analysis and primary experimental results show that the proposed techniques can provide not only self-adaptive QoS for heterogeneous network conditions, but also quick channel switching with near fixed delay time. [ABSTRACT FROM AUTHOR]

The Freeband I-Share project aims to define the mechanisms for trust, willingness, resource discovery and sharing mechanisms in virtual communities. To improve the secure and performance of a 3D video streaming application, which is a research vehicle of the I-Share project, we propose a virtual community based access control approach for secure service discovery and access (VICSDA) which groups services in virtual communities and only grants authenticated community members to discover and access these community services. There are two main contributions associated with this approach. First, different from most of the other access control approaches it adopts a dual access control mechanism which allows community services to define their local access control policy besides following the community membership policy. Second, behavior of these community services is monitored in order to guarantee a better QoS provision. Using this approach, the 3D video streaming application can be guaranteed with authentication and message confidentiality through the dual secure service discovery and access mechanism. Better application performance can also be achieved through the community member behavior audit. [ABSTRACT FROM AUTHOR]

Error Concealment methods for Intra frames in H.264 reconstruct the missing macroblock by computing weighted average of the boundary pixels of the neighboring blocks. However, the simple averaging of pixel values leads to blurring and degrades the picture quality severely. To solve this problem, an adaptive interpolation considering directions of neighboring blocks is proposed. Directional interpolation is adaptively chosen by a threshold computed from the distribution of local directions in areas around the lost block. Experiments show improvement of picture quality of about 0.5~2.0 dB compared to existing methods. [ABSTRACT FROM AUTHOR]

Existing Web image search engines index images by textual descriptions including filename, image caption, surrounding text, etc. However, the textual description available on the Web could be ambiguous or inaccurate in describing the actual image content and some images irrelevant to user's query are also returned by text-based search engines. In this paper, we propose to integrate the existing text-based image search engine with visual features, in order to improve the performance of pure text-based Web image search. The proposed algorithm is named SIEVE. Practical fusion methods are proposed to integrate SIEVE with contemporary text-based search engines. In our approach, text-based image search results for a given query are obtained first. Then, SIEVE is used to filter out those images which are semantically irrelevant to the query. Experimental results show that the image retrieval performance using SIEVE improves over Google image search significantly. [ABSTRACT FROM AUTHOR]

Advanced imaging applications are producing large-scale image data for medical diagnosis, part visualization and inspection, content retrieval, analysis, e-reporting, and so on. In many cases, service and non-textual annotations are made on top of the raw images to provide essential information on regions of interest, diseases, defects, evaluations, comments, and etc. These applications pose several challenges: i) Large data volume and the latency of data transfer over the Internet leads to a performance bottleneck; ii) The applications need advanced query support such as similarity queries on the multidimensional annotation data; and iii) Local applications need synchronizing data with a remote central database. In our work, we develop a general distributed multimedia data management system that achieves these goals by providing: i) an intelligent multimedia content caching system to support smooth local applications; ii) a loosely coupled extensible multi-indexing server to support different types of multimedia queries including similarity queries; iii) unified multimedia data access interfaces for universal data access; and iv) an integrated architecture that brings these technologies together into a robust system. The system is now successfully used to support image-based inspection and diagnosis applications such as global part inspection and knowledge database guided medical diagnosis. [ABSTRACT FROM AUTHOR]

Over the past few years many studies have proposed reversible data hiding schemes, but few have been applied to binary images. Some studies have utilized spread spectrum, compression and binary operations methods to achieve the data hiding goal, but most of them suffered from poor visual quality, capacity and inability to extract the hidden data during recovery. The performance of existing methods also is unsatisfactory. Therefore, this paper proposes a reversible data hiding scheme for binary images: SHC (senary Huffman compression). SHC adopts the half-white and half-black pixels of 4×1 or 2×2 blocks of six types to increase visual quality. Moreover, SHC is senary instead of binary and becomes double senary as a compression unit to increase compression rate and secret hiding capacity. Experimental results show that recovered images are well within human visual perception and have PSNR(s) greater than 33, high secret hiding capacity for 1.1 secret bits / 2 bits, and an effective compression rate of over 17% on average. All results demonstrate that the scheme has advantages in reversible data hiding for binary images. [ABSTRACT FROM AUTHOR]

In this paper, a closed-form solution of 3D reconstruction from nonparallel stereo geometry is derived theoretically. From a pair of conjugate image points in an arbitrarily configured stereo system that has been calibrated, one can directly reconstruct the 3D scene point using the closed-form solution, bypassing the process of rectifying the images or iteratively optimizing. Experiment results from both simulated data and real images validate the closed-form solution. Practical application to Image Guided System for surgery shows that the closed-form solution has advantages to improve efficiency and accuracy of 3D reconstruction from nonparallel stereo system in comparison with the conventional method that employs algorithm for the standard parallel axis stereo geometry. [ABSTRACT FROM AUTHOR]

In this paper, we present an optimized topology maintenance framework used in AVStreamer P2P media streaming system. To help new incoming clients get an initial view of the whole network, a new weak hierarchy model is designed to share burdens among different servers. Based on this new model, several key techniques are adopted: node evaluation criteria which can help optimize the node-list cache, receiver-driven gossip which is used to update node-list cache from others more effectively, random walk algorithm which can make the network more random, and status polling which is used to ensure the aliveness of each node cached. Experimental results show that the proposed techniques are effective. [ABSTRACT FROM AUTHOR]

Traditional boosting method like adaboost, boosts a weak learning algorithm by updating the sample weights (the relative importance of the training samples) iteratively. In this paper, we propose to integrate feature re-weighting into boosting scheme, which not only weights the samples but also weights the feature elements iteratively. To avoid overfitting problem caused by feature re-weighting on a small training data set, we also incorporate relevance feedback into boosting and propose an interactive boosting called i.Boosting. It merges adaboost, feature re-weighting and relevance feedback into one framework and exploits the favorable attributes of these methods. In this paper, i.Boosting is implemented using Adaptive Discriminant Analysis (ADA) as base classifiers. It not only enhances but also combines a set of ADA classifiers into a more powerful one. A feature re-weighting method for ADA is also proposed and integrated in i.Boosting. Extensive experiments on UCI benchmark data sets, three facial image data sets and COREL color image data sets show the superior performance of i.Boosting over AdaBoost and other state-of-the-art projection-based classifiers. [ABSTRACT FROM AUTHOR]

Recently, there is great interest of using P2P network in multi-source media streaming. In this paper, we present a self-adjustable scheduling algorithm for multi-source P2P media streaming, called feedback based scheduling algorithm (FBSA). FBSA is simple to implement, fully distributed, asynchronous, and can afford better quality of service. Furthermore, it can adjust itself automatically to adapt network heterogeneity, load variety, bandwidth fluctuation, and overlay dynamics. We have implemented FBSA, and tested it in a P2P live media streaming system: AVStreamer. Experimental results show that FBSA outperforms other conventional P2P data scheduling algorithms, especially in heterogeneous networks. [ABSTRACT FROM AUTHOR]