Advances in Visual Computing

A new algorithm able to efficiently detect a large number of overlapping ellipses with a reduced number of false positives is described. The algorithm estimates the number of candidate ellipse centers in an image with the help of a 2-dimensional accumulator and determines the five ellipse parameters with an ellipse fitting algorithm. The proposed ellipse detection algorithm uses a heuristic to select, among all image points, those with greater probabilities of belonging to an ellipse. This leads to an increase in classification efficiency, even in the presence of noise. Testing has shown that the proposed algorithm detected 97.4% of the ellipses in 100 images. Each image contained ten overlapping ellipses surrounded by noise. The ellipse parameters were determined with great accuracy.

Numismatics deals with various historical aspects of the phenomenon money. Fundamental part of a numismatists work is the identification and classification of coins according to standard reference books. The recognition of ancient coins is a highly complex task that requires years of experience in the entire field of numismatics. To date, no optical recognition system for ancient coins has been investigated successfully. In this paper, we present an extension and combination of local image descriptors relevant for ancient coin recognition. Interest points are detected and their appearance is described by local descriptors. Coin recognition is based on the selection of similar images based on feature matching. Experiments are presented for a database containing ancient coin images demonstrating the feasibility of our approach.

Training discriminative classifiers for a large number of classes is a challenging problem due to increased ambiguities between classes. In order to better handle the ambiguities and to improve the scalability of classifiers to larger number of categories, we learn pairwise dissimilarity profiles (functions of spatial location) between categories and adapt them into nearest neighbor classification. We introduce a dissimilarity distance measure and linearly or nonlinearly combine it with direct distances. We illustrate and demonstrate the approach mainly in the context of appearance-based person recognition.

This paper presents a template matching approach to high accuracy detection and tracking of perspectively distorted objects. To this end we propose a robust match metric that allows significant perspective shape changes. Using a coarse-to-fine representation for the detection of the template further increases efficiency. Once an template is detected at interactive frame-rate, we immediately switch to tracking with the same algorithm, enabling detection times of only 20ms. We show in a number of experiments that the presented approach is not only fast, but also very robust and highly accurate in detecting the 3D pose of planar objects or planar subparts of non-planar objects. The approach is used in augmented reality applications that could up to now not be sufficiently solved, because existing approaches either needed extensive training data, like machine learning methods, or relied on interest point extraction, like descriptors-based methods.

A method is proposed to enhance boundary primitives of multi-part objects of unknown specific shape and appearance in natural images. Its input is a strictly over-segmented constant-curvature contour primitive (CCP) map. Each circular arc or straight-line segment primitive from the map has an unknown origin which may be the closed boundary of a multi-part object, the textured or marked region enclosed by that boundary, or the external background region. Five simple criteria are applied in order to weight each contour primitive and eliminate the weakest ones. The criteria are defined on the basis of the superposition of the CCP map on a multiscale quadtree segmentation of the original intensity image. A subjective ground-truth binary map is used to assess the degree to which the final weighted map corresponds to a selective enhancement of the primitives on the object boundary. Experimental results confirm the potential of the method to selectively enhance, in images of variable complexity, actual boundary primitives of natural and man-made multi-part objects of diverse shapes and appearances.

3D object modeling is a crucial issue for environment recognition. A difficult problem is how to separate objects from the background clutter. This paper presents a method of 3D object modeling and segmentation from images for specific object recognition. An object model is composed of edge points which are reconstructed using a structure-from-motion technique. A SIFT descriptor is attached to each edge point for object recognition. The object of interest is segmented by finding the edge points which co-occur in images with different backgrounds. Experimental results show that the proposed method creates detailed 3D object models successfully.

A realistic yet simple 3D cloud synthesis method is examined in this work. Our synthesis framework consists of two main components. First, by introducing a similarity approach for physics-based cumulus cloud simulation to reduce the computational complexity, we introduce a particle system for cloud modeling. Second, we adopt a voxel-based scheme to model the water phase transition effect and the cloud fractal structure. It is demonstrated by computer simulation that the proposed synthesis algorithm generates realistic 3D cumulus clouds with high computational efficiency and enables flexible cloud shape control and incorporation of the wind effect …

We propose a novel framework for compressing articulated human motions using multiresolution wavelet techniques. Given a global error tolerance, the number of wavelet coefficients required to represent the motion is minimized. An adaptive error approximation metric is designed so that the optimization process can be accelerated by dynamic programming. The performance is then further improved by choosing wavelet coefficients non-linearly. To handle the footskate artifacts on the contacts, a contact stabilization algorithm which incorporates an

Inverse Kinematics

solver is adopted. Our framework requires far fewer computations, and achieves better performance in compression ratio compared to the best existing methods.

Determining the least

m

such that one

m

×

m

bi-cubic macro-patch per quadrilateral offers enough degrees of freedom to construct a smooth surface by local operations regardless of the vertex valences is of fundamental interest; and it is of interest for computer graphics due to the impending ability of GPUs to adaptively evaluate polynomial patches at animation speeds.

We constructively show that

m

 = 3 suffices, show that

m

 = 2 is unlikely to always allow for a localized construction if each macro-patch is internally parametrically

C

1

and that a single patch per quad is incompatible with a localized construction. We do not specify the GPU implementation.

We introduce a general method for animating controlled stepping motion for use in combining motion capture sequences. Our stepping algorithm is characterized by two simple models which idealize the movement of the stepping foot and the projected center of mass based on observations from a database of step motions. We draw a parallel between stepping and point-to-point reaching to motivate our foot model and employ an inverted pendulum model common in robotics for the center of mass. Our system computes path and speed profiles from each model and then adapts an interpolation to follow the synthesized trajectories in the final motion. We show that our animations can be enriched through the use of step examples, but also that we can synthesize stepping to create transitions between existing segments without the need for a motion example. We demonstrate that our system can generate precise, realistic stepping for a number of scenarios.

In this paper, we propose a new algorithm for fairing discrete surfaces resulting from stereo-based 3D reconstruction task. Such results are typically too dense, uneven and noisy, which is inconvenient for further processing. Our approach jointly optimises mesh smoothness and regularity. The definition is given on a discrete surface and the solution is found by discrete diffusion of a scalar function. Experiments on synthetic and real data demonstrate that the proposed approach is robust, stable, preserves qualitative shape and is applicable to even moderate-size real surfaces with boundary (0.8M vertices and 1.7M triangles).

A fast greedy algorithm for automatic decimation of polygonal meshes is proposed. Two important components of an automatic decimation algorithm are: the measure of fidelity and the iterative framework for incrementally and locally simplifying a polyhedra. The proposed algorithm employs vertex-based greedy framework for incrementally simplifying a polygonal model. Exploiting the normal field of one-ring neighborhood of a vertex, a new measure of fidelity is proposed that reflects the impotence of the vertices and is used to guide the vertex-based greedy procedure. A vertex causing minimum distortion is selected for removal and it is eliminated by collapsing one of its half-edges that causes minimum geometric distortion in the mesh. The proposed algorithm is about two times faster than QSlim algorithm, which is considered to be the fastest state-of-the-art greedy algorithm that produces reliable approximations; it competes well with QSlim in terms of Hausdorff distance, and preserves visually important features in a better way.

Constructing arguments and understanding them is not easy. Visualization of argument structure has been shown to help understanding and improve critical thinking. We describe a visualization tool for understanding arguments. It utilizes a novel hi-tree based representation of the argument’s structure and provides focus based interaction techniques for visualization. We give efficient algorithms for computing these layouts.

We present a method to interactively visualize large industrial models by replacing most triangles with implicit GPU primitives: cylinders, cone and torus slices. After a reverse-engineering process that recovers these primitives from triangle meshes, we encode their implicit parameters in a texture that is sent to the GPU. In rendering time, the implicit primitives are visualized seamlessly with other triangles in the scene. The method was tested on two massive industrial models, achieving better performance and image quality while reducing memory use.

Many recently developed information visualization techniques are radial variants of originally Cartesian visualizations. Almost none of these radial variants have been evaluated with respect to their benefits over their original visualizations. In this work we compare a radial and a Cartesian variant of a visualization tool for sequences of transactions in information hierarchies. The Timeline Trees (TLT) approach uses a Cartesian coordinate system to represent both the hierarchy and the sequence of transactions whereas the TimeRadarTrees (TRT) technique is the radial counterpart which makes use of a radial tree, as well as circle slices and sectors to show the sequence of transactions. For the evaluation we use both quantitative as well as qualitative evaluation methods including eye tracking.

In this paper, we present

VoxelBars

as an informative interface for volume visualization. VoxelBars arrange voxels into a 2D space and visually encode multiple attributes of voxels into one display. VoxelBars allow users to easily find out clusters of interesting voxels, set opacities and colors of a specific group of voxels, and achieve various sophisticated visualization tasks at voxel level. We provide results on real volume data to demonstrate the advantages of the VoxelBars over scatterplots and traditional transfer function specification methods. Some novel visualization techniques including visibility-aware transfer function design and selective clipping based on VoxelBars are also introduced.

Doppler radars are useful facilities for weather data gathering. In this paper, we propose a visualization pipeline to extract and display horizontal wind fields for Doppler radar data. At first, the input radar data are filtered with adaptive filters to reduce noise and enhance features. Then the horizontal wind field is computed by using a hierarchical

optical flow

method. In the visualization stage, a multi-level streamline construction method is employed to generate evenly-spaced streamlines to reveal the wind field structures.

We discuss the dual marching tetrahedra (DMT) method. The DMT can be viewed as a generalization of the classical cuberille method of Chen et al. to a tetrahedronal. The cuberille method produces a rendering of quadrilaterals comprising a surface that separates voxels deemed to be contained in an object of interest from those voxels not in the object. A cuberille is a region of 3D space partitioned into cubes. A tetrahedronal is a region of 3D space decomposed into tetrahedra. The DMT method generalizes the cubille method from cubes to tetrahedra and corrects a fundamental problem of the original cuberille method where separating surfaces are not necessarily manifolds. For binary segmented data, we propose a method for computing the location of vertices this is based upon the use of a minimal discrete norm curvature criterion. For applications where dependent function values are given at grid points, two alternative methods for computing vertex positions are discussed and compared. Examples are drawn from a variety of applications, including the Yes/No/Don’t_Know data sets resulting from inconclusive segmentation processes and Well-Log data sets.

A robot localization scheme is presented in which a mobile robot finds its position within a known environment through image comparison. The images being compared are those taken by the robot throughout its reconnaissance trip and those stored in an image database that contains views taken from strategic positions within the environment, and that also contain position and orientation information. Image comparison is carried out using a scale-dependent keypoint-matching technique based on SIFT features, followed by a graph-based outlier elimination technique known as Graph Transformation Matching. Two techniques for position and orientation estimation are tested (epipolar geometry and clustering), followed by a probabilistic approach to position tracking (based on Monte Carlo localization).

Biological visual systems are becoming an interesting source for the improvement of artificial visual systems. A biologically inspired read-out and pixel processing strategy is presented. This read-out mechanism is based on Selective pixel Change-Driven (SCD) processing. Pixels are individually processed and read-out instead of the classical approach where the read-out and processing is based on complete frames. Changing pixels are read-out and processed at short time intervals. The simulated experiments show that the response delay using this strategy is several orders of magnitude lower than current cameras while still keeping the same, or even tighter, bandwidth requirements.

This paper presents

S

3

E

, a software implementation of a high-quality dense stereo matching algorithm. The algorithm is based on a Census transform with a large mask size. The strength of the system lies in the flexibility in terms of image dimensions, disparity levels, and frame rates. The program runs on standard PC hardware utilizing various SSE instructions. We describe the performance optimization techniques that had a considerably high impact on the run-time performance. Compared to a generic version of the source code, a speedup factor of 112 could be achieved. On input images of 320×240 and a disparity range of 30,

S

3

E

achieves 42fps on an Intel Core 2 Duo CPU running at 2GHz.

Real-time stereo matching is necessary for many practical applications, including robotics. There are already many real-time stereo systems, but they typically use local approaches that cause object boundaries to be blurred and small objects to be removed. We have selected the Semi-Global Matching (SGM) method for implementation on graphics hardware, because it can compete with the currently best global stereo methods. At the same time, it is much more efficient than most other methods that produce a similar quality. In contrast to previous work, we have fully implemented SGM including matching with mutual information, which is partly responsible for the high quality of disparity images. Our implementation reaches 4.2 fps on a GeForce 8800 ULTRA with images of 640 ×480 pixel size and 128 pixel disparity range and 13 fps on images of 320 ×240 pixel size and 64 pixel disparity range.

In this paper, an accurate optical flow sensor based on our previous design is proposed. Improvements are made to make the optical flow sensor more suitable for obstacle avoidance tasks on a standalone FPGA platform. Firstly, because optical flow algorithms are sensitive to the noise, more smoothing units are added into the hardware pipeline to suppress the noise in real video source. These units are hardware efficient to accommodate limited hardware resources. Secondly, a cost function is used to evaluate the estimated motion vector at each pixel for higher level analysis. Experiment results show that the proposed design can substantially improve the optical flow sensor performance for obstacle avoidance applications.

In this paper we present a novel application which uses 2D barcode for object tracking and event detection. We analyze the relationship between the spatial efficiency of a marker and its robustness against defocusing. Based on our analysis we design a spatially efficient and robust 2D barcode, M-Code (Marker-Code) which can be attached to the surface of static or moving objects. Compared with traditional object detection and tracking methods, M-Code not only identifies and tracks an object but also reflects its position and the orientation of the surface where it is attached. We implemented an efficient algorithm that tracks multiple M-Codes in real time in the presence of rotation and perspective distortion. In experiments we compare the spatial efficiency of M-Code with existing 2D barcodes, and quantitatively measure its robustness, including its scaling capability and tolerance to perspective distortion. As an application we use the system to detect door movements and track multiple moving objects in real scenes.

We propose a new technique for unsupervised segmentation of the lung region from low dose computed tomography (LDCT) images. We follow the most conventional approaches such that initial images and desired maps of regions are described by a joint Markov-Gibbs random field (MGRF) model of independent image signals and interdependent region labels. But our focus is on more accurate model identification for the MGRF model and the gray level distribution model. To better specify region borders between lung and chest, each empirical distribution of volume signals is precisely approximated by a linear combination of Gaussians (LCG) with positive and negative components. LCG models parameters are estimated by the modified EM algorithm. Initial segmentation (labeled volume) based on the LCG models is then iteratively refined by using the MGRF with analytically estimated potentials. In this framework the graph cuts is used as a global optimization algorithm to find the segmented data (labeled data) that minimize a certain energy function, which integrates the LCG model and the MGRF model. To validate the accuracy of our algorithm, a special 3D geometrical phantom motivated by statistical analysis of the LDCT data is designed. Experiments on both phantom and 3D LDCT data sets show that the proposed segmentation approach is more accurate than other known alternatives.

This study investigates a variational multiphase image segmentation method which combines the advantages of graph cut discrete optimization and multiphase piecewise constant image representation. The continuous region parameters serve both image representation and graph cut labeling. The algorithm iterates two consecutive steps: an original closed-form update of the region parameters and partition update by graph cut labeling using the region parameters. The number of regions/labels can decrease from an initial value, thereby relaxing the assumption that the number of regions is known beforehand. The advantages of the method over others are shown in several comparative experiments using synthetic and real images of intensity and motion.

We present a novel graph-based approach to image segmentation. The objective is to partition images such that nearby pixels with similar colors or greyscale intensities belong to the same segment. A graph representing an image is derived from the similarity between the pixels and partitioned by a computationally efficient graph clustering method, which identifies representative nodes for each cluster and then expands them to obtain complete clusters of the graph. Experiments with synthetic and natural images are presented. A comparison with the well known graph clustering method of normalized cuts shows that our approach is faster and produces segmentations that are in better agreement with visual assessment on original images.

In this paper, we propose a new active contour model for supervised texture segmentation driven by a binary classifier instead of a standard motion equation. A recent level set implementation developed by Shi

et al

in [1] is employed in an original way to introduce the classifier in the active contour. Carried out on a learning image, an expert segmentation is used to build the learning dataset composed of samples defined by their Haralick texture features. Then, the pre-learned classifier is used to drive the active contour among several test images. Results of three active contours driven by binary classifiers are presented: a k-nearest-neighbors model, a support vector machine model and a neural network model. Results are presented on medical echographic images and remote sensing images and compared to the Chan-Vese region-based active contour in terms of accuracy, bringing out the high performances of the proposed models.

We present a novel multi-scale image segmentation approach based on irregular triangular and polygonal tessellations produced by proximity graphs. Our approach consists of two separate stages: polygonal seeds generation followed by an iterative bottom-up polygon agglomeration. We employ constrained Delaunay triangulation combined with the principles known from visual perception to extract an initial irregular polygonal tessellation of the image. These initial polygons are built upon a triangular mesh composed of irregular sized triangles, whose spatial arrangement is adapted to the image content. We represent the image as a graph with vertices corresponding to the built polygons and edges reflecting polygon relations. The segmentation problem is then formulated as Minimum Spanning Tree (MST) construction. We build a successive fine-to-coarse hierarchy of irregular polygonal partitions by an iterative graph contraction. It uses local information and merges the polygons bottom-up based on local region- and edge- based characteristics.

In this paper, we propose a different framework for incorporating spatial information with the aim of achieving robust and accurate segmentation in case of mixed noise without using experimentally set parameters, called improved adaptive spatial information clustering (IASIC) algorithm. The proposed objective function has a new dissimilarity measure, and the weighting factor for neighborhood effect is fully adaptive to the image content. It enhances the smoothness towards piecewise-homogeneous segmentation and reduces the edge-blurring effect. Furthermore, a unique characteristic of the new information segmentation algorithm is that it has the capabilities to eliminate outliers at different stages of the IASIC algorithm. These result in improved segmentation result by identifying and relabeling the outliers in a relatively stronger noisy environment. The experimental results with both synthetic and real images demonstrate that the proposed method is effective and robust to mixed noise and the algorithm outperforms other popular spatial clustering variants.

This paper addresses the problem of grouping image primitives; its principal contribution is an explicit definition of the Gestalt principle of

Prägnanz

, which organizes primitives into descriptions of images that are both simple and stable. Our definition of Prägnanz assumes just two things: that a vector of free variables controls some general grouping algorithm, and a scalar function measures the information in a grouping. Stable descriptions exist where the gradient of the function is zero, and these can be ordered by information content (simplicity) to create a “grouping” or “Gestalt” scale description. We provide a simple measure for information in a grouping based on its structure alone, leaving our grouper free to exploit other Gestalt principles as we see fit. We demonstrate the value of our definition of Prägnanz on several real-world images.

Multi-view correspondence of wide-baseline image matching is still a challenge task in computer vision. There are two main steps in dealing with correspondence issue: feature description and similarity search. The well-known SIFT descriptor is shown to be a-state-of-art descriptor which could keep distinctive invariant under transformation, large scale changes, noises and even small view point changes. This paper uses the SIFT as feature descriptor, and proposes a new search algorithm for similarity search. The proposed dichotomy based hash (DBH) method performs better than the widely used BBF (Best Bin First) algorithm, and also better than LSH (Local Sensitive Hash). DBH algorithm can obtain much higher (1-precision)-recall ratio in different kinds of image pairs with rotation, scale, noises and weak affine changes. Experimental results show that DBH can obviously improve the search accuracy in a shorter time, and achieve a better coarse match result.

We present and evaluate an approach for finding local interest points in images based on the non-minima suppression of Gradient Vector Flow (GVF) magnitude. Based on the GVF’s properties it provides the approximate centers of blob-like structures or homogeneous structures confined by gradients of similar magnitude. It results in a scale and orientation invariant interest point detector, which is highly stable against noise and blur. These interest points outperform the state of the art detectors in various respects. We show that our approach gives a dense and repeatable distribution of locations that are robust against affine transformations while they outperform state of the art techniques in robustness against lighting changes, noise, rotation and scale changes. Extensive evaluation is carried out using the Mikolajcyzk framework for interest point detector evaluation.

This paper presents a new method for 3D shape retrieval based on the bags-of-words model along with a weak spatial constraint. First, a two-pass sampling procedure is performed to extract the local shape descriptors, based on spin images, which are used to construct a shape dictionary. Second, the model is partitioned into different regions based on the positions of the words. Then each region is denoted as a histogram of words (also known as

bag-of-words

) as found in it along with its position. After that, the 3D model is represented as the collection of histograms, denoted as bags-of-words, along with their relative positions, which is an extension of an orderless bag-of-words 3D shape representation. We call it as

Spatial Enhanced Bags-of-Words

(SEBW). The spatial constraint shows improved performance on 3D shape retrieval tasks.

Obtaining a top match for a given query image from a set of images forms an important part of the scene identification process. The query image typically is not identical to the images in the data set, with possible variations of changes in scale, viewing angle and lighting conditions. Therefore, features which are used to describe each image should be invariant to these changes. Standard image capturing devices lose much of the color and lighting information due to encoding during image capture. This paper uses high dynamic range images to utilize all the details obtained at the time of capture for image matching. Once the high dynamic range images are obtained through the fusion of low dynamic range images, feature detection is performed on the query images as well as on the images in the database. A junction detector algorithm is used for detecting the features in the image. The features are described using the wedge descriptor which is modified to adapt to high dynamic range images. Once the features are described, a voting algorithm is used to identify a set of top matches for the query image.

Following recent studies, the automatic analysis of intracranial pressure pulses (ICP) seems to be a promising tool for forecasting critical intracranial and cerebrovascular pathophysiological variations during the management of many neurological disorders. MOCAIP algorithm has recently been developed to automatically extract ICP morphological features. The algorithm is able to enhance the quality of ICP signals, to segment ICP pulses, and to recognize the three peaks occurring in a ICP pulse. This paper extends MOCAIP by introducing a generic framework to perform robust peak recognition. The method is local in the sense that it exploits subwindows that are randomly extracted from ICP pulses. The recognition process combines recently developed machine learning algorithms. The experimental evaluations are performed on a database built from several hundreds of hours of ICP recordings. They indicate that the proposed extension increases the robustness of the peak recognition.

Recently, content based 3D shape retrieval has been an active area of research. Benchmarking allows researchers to evaluate the quality of results of different 3D shape retrieval approaches. Here, we propose a new publicly available 3D shape benchmark to advance the state of art in 3D shape retrieval. We provide a review of previous and recent benchmarking efforts and then discuss some of the issues and problems involved in developing a benchmark. A detailed description of the new shape benchmark is provided including some of the salient features of this benchmark. In this benchmark, the 3D models are classified mainly according to visual shape similarity but in contrast to other benchmarks, the geometric structure of each model is modified and normalized, with each class in the benchmark sharing the equal number of models to reduce the possible bias in evaluation results. In the end we evaluate several representative algorithms for 3D shape searching on the new benchmark, and a comparison experiment between different shape benchmarks is also conducted to show the reliability of the new benchmark.

We present a graph-based contour extraction algorithm for images with low contrast regions and faint contours. Our innovation consists of a new graph setup that exploits complementary information given by region segmentation and contour grouping. The information of the most salient region segments is combined together with the edge map obtained from the responses of an oriented filter bank. This enables us to define a new contour flow on the graph nodes, which captures region membership and enhances the flow in the low contrast or cluttered regions. The graph setup and our proposed region based normalization give rise to a random walk that allows bifurcations at junctions arising between region boundaries and favors long closed contours. Junctions become key routing points and the resulting contours enclose globally significant regions.

A 24/7 traffic surveillance system needs to perform robustly in dynamic traffic conditions. Despite the amount of work that has been done in creating suitable background models, we observe limitations with the state-of-the-art methods when there is minimal color information and the background processes have a high variance due to lighting changes or adverse weather conditions. To improve the performance, we propose in this paper a Difference of Gaussian (DoG) edge-texture based modeling for learning the background and detecting vehicles in such conditions. Background DoG images are obtained at different scales and summed to obtain an Added DoG image. The edge-texture information contained in the Added DoG image is modeled using the Local Binary Pattern (LBP) texture measure. For each pixel in the Added DoG image, a group of weighted adaptive LBP histograms are obtained. Foreground vehicles are detected by matching an existing histogram obtained from the current Added DoG image to the background histograms. The novelty of this technique is that it provides a higher level of learning by establishing a relationship an edge pixel has with its neighboring edge and non-edge pixels which in turn provides us with better performance in foreground detection and classification.

We present a method for detecting common human actions in video, common to athletics and surveillance, using intuitive sketches and motion cues. The framework presented in this paper is an automated end-to-end system which (1) interprets the sketch input, (2) generates a query video based on motion cues, and (3) incorporates a new content-based action descriptor for matching. We apply our method to a publicly-available video repository of many common human actions and show that a video matching the concept of the sketch is generally returned in one of the top three query results.

This paper presents an automatic visual analysis system for simultaneously tracking humans and vehicles using multiple cameras in an unconstrained outdoor environment. The system establishes correspondence between views using a principal axis approach for humans and a footage region approach for vehicles. Novel methods for locating humans in groups and solving ambiguity when matching vehicles across views are presented. Foreground segmentation for each view is performed using the codebook method and HSV shadow suppression. The tracking of objects is performed in each view, and occlusion situations are resolved by probabilistic appearance models. The system is tested on hours of video and on three different datasets.

This paper introduces the use of self-organizing maps for the visualization of crowd dynamics and to learn models of the dominant motions of crowds in complex scenes. The self-organizing map (SOM) model is a well known dimensionality reduction method proved to bear resemblance with characteristics of the human brain, representing sensory input by topologically ordered computational maps. This paper proposes algorithms to learn and compare crowd dynamics with the SOM model. Different information is employed as input to the used SOM. Qualitative and quantitative results are presented in the paper.

To function in the real world, a robot must be able to understand human intentions. This capability depends on accurate and reliable detection and tracking of trajectories of agents in the scene. We propose a visual tracking framework to generate and maintain trajectory information for all agents of interest in a complex scene. We employ this framework in an intent recognition system that uses spatio-temporal contextual information to recognize the intentions of agents acting in different scenes, comparing our system with the state of the art.

This paper presents an effective algorithm to segment color video into shots for video indexing or retrieval applications. This work adds global texture information to our previous work, which extended the scale-invariant feature transform (SIFT) to color global texture SIFT (CGSIFT). Fibonacci lattice-quantization is used to quantize the image and extract five color features for each region of the image using a symmetrical template. Then, in each region of the image partitioned by the template, the entropy and energy of a co-occurrence matrix are calculated as the texture features. With these global color and texture features, we adopt clustering ensembles to segment video shots. Experimental results show that the additional texture features allow the proposed CGTSIFT algorithm to outperform our previous work, fuzzy-c means, and SOM-based shot detection methods.

The principle of a vision system based on a progressive focus of attention principle is presented. This approach considers the visual recognition strategy as an estimation problem for which the purpose is to estimate both precisely and reliably the parameters of the object to be recognized. The object is constituted of parts statistically dependent one each other thanks to a statistical model. The reliability is calculated within a bayesian framework. The case of lane sides detection for driving assistance is given as an illustration.

In this paper we present a quantitative assessment of user performance in an assembly modeling application. The purpose of the evaluation is to identify the influence of immersion and collaboration on the performance in assembly and manipulation tasks in a virtual environment (VE). The environment used in this study is a projection-based system that supports 6DOF head and hand tracking for two independent users. In four experiments we compare the performance in stereoscopic and monoscopic viewing modes and in collaborative and single-user interaction modes. The user study is based on a realistic, application-oriented modeling scenario from the manufacturing industry. The task completion time is used as the performance metric. The results of this study show that collaborative interaction and stereoscopic viewing is not only preferred by users, but also have a significant effect on their performance.

Combining physically based simulation and motion capture data for animation is becoming a popular alternative to large motion databases for rich character motion. In this paper, our focus is on adapting motion-captured sequences for character response to external perturbations. Our technique is similar to approaches presented in the literature, but we propose a novel, straightforward way of computing feedforward control. While alternatives such as inverse dynamics and feedback error learning (FEL) exist, they are more complicated and require offline processing in contrast to our method which uses an auxiliary dynamic simulation to compute feedforward torques. Our method is simple, general, efficient, and can be performed at runtime. These claims are demonstrated through various experimental results of simulated impacts.

This paper presents a new robust and reliable marker less camera tracking system for outdoor augmented reality using only a mobile handheld camera. The proposed method is particularly efficient for partially known 3D scenes where only an incomplete 3D model of the outdoor environment is available. Indeed, the system combines an edge-based tracker with a sparse 3D reconstruction of the real-world environment to continually perform the camera tracking even if the model-based tracker fails. Experiments on real data were carried out and demonstrate the robustness of our approach to occlusions and scene changes.

In this paper, we present a technique that allows multiple participants within a large-scale immersive, virtual environment to interact with it using pointing gestures. Multiple cameras observing the environment are used along with various computer vision techniques to obtain a 3-D reconstruction of each participant’s finger position, allowing the participants to point at and interact with virtual objects. This use of the pointing gesture provides an intuitive method of interaction, and has many applications in the field of human computer interaction. No markers or special clothing are required to be worn. Furthermore, we show that the system is able to provide robust results in real time.

This paper presents an algorithm for using projective invariant patterns in augmented reality applications. It is actually an adaptation of a previous algorithm for an optical tracking device, that works with infrared illumination and filtering. The present algorithm removes the necessity of working in a controlled environment, which would be inadequate for augmented reality applications. In order to compensate the excess of image noise caused by the absence of the infrared system, the proposed algorithm includes a fast binary decision tree in the process flow. We show that the algorithm achieves real time rates.

Camera-based tracking systems which reconstruct a feature map with structure from motion or SLAM techniques highly depend on the ability to track a single feature in different scales, different lighting conditions and a wide range of viewing angles. The acquisition of high quality features is therefore indispensable for a continuous tracking of a feature with a maximum possible range of valid appearances.

We present a tracking system where not only the position of a feature but also its surface normal is reconstructed and used for precise prediction and tracking recovery of lost features. The appearance of a reference patch is also estimated sequentially and refined during the tracking, which leads to a more stable feature tracking step. Such reconstructed reference templates can be used for tracking a camera pose with a great variety of viewing positions.

This feature reconstruction process is combined with a feature management system, where a statistical analysis of the ability to track a feature is performed, and only the most stable features for a given camera viewing position are used for the 2D feature tracking step. This approach results in a map of high quality features, where the the real time capabilities can be preserved by only tracking the most necessary 2D feature points.

Segmentation of cellular images presents a challenging task for computer vision, especially when the cells of irregular shapes clump together. Level set methods can segment cells with irregular shapes when signal-to-noise ratio is low, however they could not effectively segment cells that are clumping together. We perform topological analysis on the zero level sets to enable effective segmentation of clumped cells. Geometrical shapes and intensities are important information for segmentation of cells. We assimilated them in our approach and hence we are able to gain from the advantages of level sets while circumventing its shortcoming. Validation on a data set of 4916 neural cells shows that our method is 93.3 ±0.6% accurate.

This paper proposes an automatic segmentation algorithm that combines clustering and deformable models. First, a k-means clustering is performed based on the image intensity. A hierarchical recognition scheme is then used to recognize the structure to be segmented, and an initial seed is constructed from the recognized region. The seed is then evolved under certain deformable model mechanism. The automatic recognition is based on fuzzy logic techniques. We apply our algorithm for the segmentation of the corpus callosum and the thalamus from brain MRI images. Depending on the specific features of the segmented structures, the most suitable recognition schemes and deformable models are employed. The whole procedure is automatic and the results show that this framework is fast and robust.

Segmentations of brain lesions from Magnetic Resonance (MR) images is crucial for quantitative analysis of lesion populations in neuroimaging of neurological disorders. We propose a new method for segmenting lesions in brain MRI by inferring the underlying physical models for pathology. We use the reaction-diffusion model as our physical model, where the diffusion process is guided by real diffusion tensor fields that are obtained from Diffusion Tensor Imaging (DTI). The method performs segmentation by solving the inverse problem, where it determines the optimal parameters for the physical model that generates the observed image. We show that the proposed method can infer reasonable models for multiple sclerosis (MS) lesions and healthy MRI data. The method has potential for further extensions with different physical models or even non-physical models based on existing segmentation schemes.

In this paper we propose a method for geometrical calibration of bi-planar radiography that aims at minimising the impact of calibration objects on the content of radiographs, while remaining affordable. For accomplishing this goal, we propose a small extension to conventional imaging systems: a low-cost rangefinder that enables to estimate some of the geometrical parameters. Even with this extension, a small calibration object is needed for correcting scale.

The proposed method was tested on 17 pairs of radiographs of a phantom object of known dimensions. For calculating scale, only a reference distance of 40mm was used. Results show a RMS error of 0.36mm with 99% of the errors inferior to 0.85mm.

We conclude that the method presented here performs robustly and achieves sub-millimetric accuracy, while remaining affordable and requiring only two radiopaque marks visible in radiographs.

Detection of cell nucleus is critical in microscopy image analysis and ellipse detection plays an important role because most nuclei are elliptical in shapes. We developed an ellipse detection algorithm based on the Mumford-Shah model that inherits its superior properties. In our ellipse detector, the active contours in the Mumford-Shah model are constrained to be non-overlapping ellipses. A quantitative comparison with the randomized Hough transform shows that the Mumford-Shah based approach detects nucleus significantly better on our data sets.

We present a novel automated method for assessment of image alignment, applied to non-rigid registration of brain Magnetic Resonance Imaging data (MRI) for image-guided neurosurgery. We propose a number of robust modifications to the Hausdorff distance (HD) metric, and apply it to the edges recovered from the brain MRI to evaluate the accuracy of image alignment. The evaluation results on synthetic images, simulated tumor growth MRI and real neurosurgery data with expert-identified anatomical landmarks, confirm that the accuracy of alignment error estimation is improved compared to the conventional HD. The proposed approach can be used to increase confidence in the registration results, assist in registration parameter selection, and provide local estimates and visual assessment of the registration error.

Hand drawn ornamentation, such as floral or geometric patterns, is a tedious and time consuming task that requires skill and training in ornamental design principles and aesthetics. Ornamental drawings both historically and presently play critical roles in all things from art to architecture, and when computers handle the repetition and overall structure of ornament, considerable savings in time and money can result. Due to the importance of keeping an artist in the loop, we present an application, designed and implemented utilizing a user-driven global planning strategy, to help guide the generation of two-dimensional ornament. The system allows for the creation of beautiful, organic ornamental 2D art which follows a user-defined curve. We present the application and the algorithmic approaches used.

Current MCMC algorithms are limited from achieving high rendering efficiency due to possibly high failure rates in caustics perturbations and stratified exploration of the image plane. In this paper we improve the MCMC approach significantly by introducing new lens perturbation and new path-generation methods. The new lens perturbation method simplifies the computation and control of caustics perturbation and can increase the perturbation success rate. The new path-generation methods aim to concentrate more computation on “high perceptual variance” regions and “hard-to-find-but-important” paths. We implement these schemes in the Population Monte Carlo Energy Redistribution framework to demonstrate the effectiveness of these improvements. In addition., we discuss how to add these new schemes into the Energy Redistribution Path Tracing and Metropolis Light Transport algorithms. Our results show that rendering efficiency is improved with these new schemes.

Graphics rendering on mobile devices is severely restricted by available battery energy. The frame rate of real-time graphics applications fluctuates due to continual changes in the LoD, visibility and distance of scene objects, user interactivity, complexity of lighting and animation, and many other factors. Such frame rate spikes waste precious battery energy. We introduce an adaptive CPU scheduler that predicts the applications workload from frame to frame and allocates just enough CPU cycles to render the scene at a target rate of 25 FPS. Since the applications workload needs to be re-estimated whenever the scenes LoD changes, we integrate our CPU scheduler with LoD management. To further save energy, we try to render scenes at the lowest LoD at which the user does not see visual artifacts on a given screen. Our integrated Energy-efficient Adaptive Real-time Rendering (EARR) heuristic reduces energy consumption by up to 60% while maintaining acceptable image quality at interactive frame rates.

This paper describes a quick 3D-to-2D point matching algorithm. Our major contribution is to substitute a new

O

(2

n

) algorithm for the traditional

N

! method by introducing a convex hull based enumerator. Projecting a 3D point set into a 2D plane yields a corresponding 2D point set. In some cases, matching information is lost. Therefore, we wish to recover the 3D-to-2D correspondence in order to compute projection parameters. Traditionally, an exhaustive enumerator permutes all the potential matching sets, which is

N

! for

N

points, and a projection parameter computation is used to choose the correct one. We define ”correct” as the points match whose computed parameters result in the lowest residual error. After computing the convex hull for both 2D and 3D points set, we show that the 2D convex hull must match a circuit of the 3D convex hull having the same length. Additionally a novel validation method is proposed to further reduce the number of potential matching cases. Finally, our matching algorithm is applied recursively to further reduce the search space.

A simple but very efficient method for snake locomotion generation is presented in this paper. Instead of relying on conventional physically based simulation or tedious key-framing, a novel deformation-based approach is utilized to create realistic looking snake motion patterns. Simple sinusoidal, winding, and bending functions constitute the deformation. The combination of various types of deformation becomes a powerful tool for describing the characteristic motions of a typical snake. As an example, three basic deformations and their combinations are utilized and various locomotive animations are generated with a high degree of realism. The proposed method provides an easy-to-use, fast, and interactive mechanism for an animator with little experience. The method is versatile in that it also works in conjunction with the creative input of an experienced animator for the improvement of the overall quality of the animation.

In this paper we introduce a novel GPU-supported JPEG image compression technique with a focus on its application for remote visualization purposes. Fast and high quality compression techniques are very important for the remote visualization of interactive simulations and Virtual reality applications (IS/VR) on hybrid clusters. Thus the main goals of the design and implementation of this compression technique were low compression times and nearly no visible quality loss, while achieving compression rates that allow for 30+ Frames per second over 10 MBit/s networks. To analyze the potential of the technique and further development needs and to compare it to existing methods, several benchmarks are conducted and described in this paper. Additionally a quality assessment is performed to allow statements about the achievable quality of the lossy image compression. The results show that using the GPU not only for rendering but also for image compression is a promising approach for interactive remote rendering.

In recent years the design of space-efficient algorithms that work within a limited amount of memory is becoming a hot topic of research. This is particularly crucial for intelligent peripherals used in image analysis and processing, such as digital cameras, scanners, or printers, that are equipped with considerably lower memory than the usual computers. In the present paper we propose a constant-working space algorithm for determining the genus of a binary digital object. More precisely, given an

m

×

n

binary array representing the image, we show how one can count the number of holes of the array with an optimal number of

O

(

mn

) integer arithmetic operations and optimal

O

(1) working space. Our consideration covers the two basic possibilities for object and hole types determined by the adjacency relation adopted for the object and for the background. The algorithm is particularly based on certain combinatorial relation between some characteristics of a digital picture.

In this paper we investigate an approach of constructing a 3D digital line by taking the integer points within an offset of a certain radius of the line. Alternatively, we also investigate digital lines obtained through a “pseudo-offset” defined by a parallelepiped enclosing the integer points around the line. We show that if the offset radius (resp. side of the parallelepiped section) is greater than

$\sqrt{3}$

(resp. 2

$\sqrt{3}$

), then the digital line is at least 1-connected. Extensive experiments show that the lines obtained feature satisfactory appearance.

We propose a new torsion estimator for spatial curves based on results of discrete geometry that works in

O

(

n

log

2

n

) time. We also present a curvature estimator for spatial curves. Our methods use the 3D extension of the 2D blurred segment notion [1]. These estimators can naturally work with disconnected curves.

Tomographic reconstructions are often segmented to extract valuable quantitative information. In this paper, we consider the problem of segmenting a dense object of constant density within a continuous tomogram, by means of global thresholding. Selecting the proper threshold is a nontrivial problem, for which hardly any automatic procedures exists. We propose a new method that exploits the available projection data to accurately determine the optimal global threshold. Results from simulation experiments show that our algorithm is capable of finding a threshold that is close to the optimal threshold value.

Several curvature estimators along digital contours were proposed in recent works [1,2,3]. These estimators are adapted to non perfect digitization process and can process noisy contours. In this paper, we compare and analyse the performances of these estimators on several types of contours and we measure execution time on both perfect and noisy shapes. In a second part, we evaluate these estimators in the context of corner detection. Finally to evaluate the performance of a non curvature based approach, we compare the results with a morphological corner detector [4].

Almost all three dimensional manifolds admit canonical metrics with constant sectional curvature. In this paper we proposed a new algorithm pipeline to compute such canonical metrics for hyperbolic 3-manifolds with high genus boundary surfaces. The computation is based on the discrete curvature flow for 3-manifolds, where the metric is deformed in an angle-preserving fashion until the curvature becomes uniform inside the volume and vanishes on the boundary. We also proposed algorithms to visualize the canonical metric by realizing the volume in the hyperbolic space ℍ

3

, both in single period and in multiple periods. The proposed methods could not only facilitate the theoretical study of 3-manifold topology and geometry using computers, but also have great potentials in volumetric parameterizations, 3D shape comparison, volumetric biomedical image analysis and etc.

The dramatic growth in practical applications for iris biometrics has been accompanied by many important developments in the underlying algorithms and techniques. Among others, one of the most active research areas concerns about the development of iris recognition systems less constrained to users, either increasing the image acquisition distances or the required lighting conditions. The main point of this paper is to give a process suitable for the automatic segmentation of iris images captured at the visible wavelength, on-the-move and within a large range of image acquisition distances (between 4 and 8 meters). Our experiments were performed on images of the UBIRIS.v2 database and show the robustness of the proposed method to handle the types of non-ideal images resultant of the aforementioned less constrained image acquisition conditions.

This paper presents a solution to a problem existing in the cork industry: cork stopper/disk classification according to their quality. Cork is a natural and heterogeneous material; therefore, its automatic classification (seven quality classes exist) is very difficult. The solution proposed in this paper combines the extraction of 3D cork features and soft-computing. In order to evaluate the performance of the neuro-fuzzy network designed, we compare its results with other 4 basic classifiers working with the same feature space. In conclusion, our experiments showed that the best results in case of cork quality classification were obtained with the proposed system that works with the following features: depth+intensity combined feature, weighted depth, second depth level feature, root mean square roughness and other three textural features (wavelets). The obtained classification results have highly improved other results reported in similar studies.

Breast cancer is the most commonly diagnosed form of cancer in women accounting for about 30% of all cases. Medical thermography has been shown to be well suited for the task of detecting breast cancer, in particular when the tumour is in its early stages or in dense tissue. In this paper we perform breast cancer analysis based on thermography. We employ a series of statistical features extracted from the thermograms which describe bilateral differences between left and right breast areas. These features then form the basis of a hybrid fuzzy rule-based classification system for diagnosis. The rule base of the classifier is optimised through the application of a genetic algorithm which ensures a small set of rules coupled with high classification performance. Experimental results on a large dataset of nearly 150 cases confirm the efficacy of our approach.

In this paper, an efficient algorithm for facial feature detection is presented. Complex regions in a face image, such as the eye, exhibit unpredictable local intensity and hence high entropy. We use this characteristic to obtain eye candidates, and then these candidates are sent to a SVM classifier to get real eyes. According to the geometry relationship of human face, mouth search region is specified by the coordinates of the left eye and the right eye. And then precise mouth detection is done. Experimental results demonstrate the effectiveness of the proposed method.

Background modeling is a key step of background subtraction methods used in the context of static camera. The goal is to obtain a clean background and then detect moving objects by comparing it with the current frame. Mixture of Gaussians Model [1] is the most popular technique and presents some limitations when dynamic changes occur in the scene like camera jitter, illumination changes and movement in the background. Furthermore, the MGM is initialized using a training sequence which may be noisy and/or insufficient to model correctly the background. All these critical situations generate false classification in the foreground detection mask due to the related uncertainty. To take into account this uncertainty, we propose to use a Type-2 Fuzzy Mixture of Gaussians Model. Results show the relevance of the proposed approach in presence of camera jitter, waving trees and water rippling.

A new unsupervised clustering algorithm, Spectral-division Unsupervised Shot-clustering Algorithm (SUSC), is proposed in this paper. Key-fames are picked out to represent the shots, and color feature of key-frames are extracted to describe video shots. Spherical Gaussian Model (SGM) is constructed for every shot category to form effective descriptions of them. Then Spectral Division (SD) method is employed to divide a category into two categories, and the method is iteratively used for further divisions. After each iterative shot-division, Bayesian information Criterion (BIC) is utilized to automatically judge whether to stop further division. During this processes, one category may be dissevered by mistake. In order to correct these mistakes, similar categories will be merged by calculating the similarities of every two categories. This approach is applied to three kinds of sports videos, and the experimental results show that the proposed approach is reliable and effective.

Many different shape from shading (SFS) algorithms have emerged during the last three decades. Recently, we proposed [1] a unified framework that is capable of solving the SFS problem under various settings of imaging conditions representing the image irradiance equation of each setting as an explicit Partial Differential Equation (PDE). However, the result of any SFS algorithm is mainly affected by errors in the given image brightness, either due to image noise or modeling errors. In this paper, we are concerned with quantitatively assessing the degree of robustness of our unified approach with respect to these errors. Experimental results have revealed promising performance against noisy images but has also lacked in reconstructing the correct shape due to error in the modeling process. This result emphasizes the need for robust algorithms for surface reflectance estimation to aid SFS algorithms producing more realistic shapes.

Many models and algorithms have been proposed since the shape from texture problem was tackled by the pioneering work of Gibson in 1950. In the present work, a general assumption of stochastic homogeneity is chosen so as to include a wide range of natural textures. Under this assumption, the Fourier transform of the image and a minimal set of Gabor filters are used to efficiently estimate all the main local spatial frequencies of the texture, i.e. so as to compute distortion measures. Then a known method which uses singular value decomposition to process the frequencies under orthographic projection is considered. The method is extended to general perspective cases and used to reconstruct the 3D shape of real pictures and video sequences. The robustness of the algorithm is proven on general shapes, and results are compared with the literature when possible.

We present a novel algorithm that applies conics to realize reliable camera calibration. In particular, we show that a single view of two coplanar circles is sufficiently powerful to give a fully automatic calibration framework that estimates both intrinsic and extrinsic parameters. This method stems from the previous work of conic based calibration and calibration-free scene analysis. It eliminates many

a priori

constraints such as known principal point, restrictive calibration patterns, or multiple views. Calibration is achieved statistically through identifying multiple orthogonal directions and optimizing a probability function by maximum likelihood estimate. Orthogonal vanishing points, which build the basic geometric primitives used in calibration, are identified based on the fact that they represent conjugate directions with respect to an arbitrary circle under perspective transformation. Experimental results from synthetic and real scenes demonstrate the effectiveness, accuracy, and popularity of the approach.

In this paper we present a new approach for automatic external calibration for two camera views under general motion based on both line and point features. Detected lines are classified into two classes: either vertical or horizontal. We make use of these lines extensively to determine the camera pose. First, the rotation is estimated directly from line features using a novel algorithm. Then normalized point features are used to compute the translation based on epipolar constraint. Compared with point-feature-based approaches, the proposed method can handle well images with little texture. Also, our method bypasses sophisticated post-processing stage that is typically employed by other line-feature-based approaches. Experiments show that, although our approach is simple to implement, the performance is reliable in practice.

A method for stabilizing the computation of stereo correspondences is presented in this paper. Delaunay triangulation is employed to partition the input images into small, localized regions. Instead of simply assuming that the surface patches viewed from these small triangles are locally planar, we explicitly examine the planarity hypothesis in the 3D space. To perform the planarity test robustly, adjacent triangles are merged into larger polygonal patches first and then the planarity assumption is verified. Once piece-wise planar patches are identified, point correspondences within these patches are readily computed through planar homographies. These point correspondences established by planar homographies serve as the ground control points (GCPs) in the final dynamic programming (DP)-based correspondence matching process. Our experimental results show that the proposed method works well on real indoor, outdoor, and medical image data and is also more efficient than the traditional DP method.

We describe an immersive visualization application for point cloud data gathered by LiDAR (Light Detection And Ranging) scanners. LiDAR is used by geophysicists and engineers to make highly accurate measurements of the landscape for study of natural hazards such as floods and earthquakes. The large point cloud data sets provided by LiDAR scans create a significant technical challenge for visualizing, assessing, and interpreting these data. Our system uses an out-of-core view-dependent multiresolution rendering scheme that supports rendering of data sets containing billions of 3D points at the frame rates required for immersion (48–60 fps). The visualization system is the foundation for several interactive analysis tools for quality control, extraction of survey measurements, and the extraction of isolated point cloud features. The software is used extensively by researchers at the UC Davis Department of Geology and the U.S. Geological Survey, who report that it offers several significant advantages over other analysis methods for the same type of data, especially when used in an immersive visualization environment such as a CAVE.

Scientific research has become increasingly interdisciplinary, and clear communication is fundamental when bringing together specialists from different areas of knowledge. This work aims at discussing the role of fully immersive virtual reality experience to facilitate interdisciplinary communication by utilising the Duke Immersive Virtual Environment (DiVE), a CAVE-like system, to explore the complex and high-resolution results from the Regional Atmospheric Modelling System-based Forest Large-Eddy Simulation (RAFLES) model coupled with the Ecosystem Demography model (ED2). VR exploration provided an intuitive environment to simultaneously analyse canopy structure and atmospheric turbulence and fluxes, attracting and engaging specialists from various backgrounds during the early stages of the data analysis. The VR environment facilitated exploration of large multivariate data with complex and not fully understood non-linear interactions in an intuitive and interactive way. This proved fundamental to formulate hypotheses about tree-scale atmosphere-canopy-structure interactions and define the most meaningful ways to display the results.

Numerical models such as the Mesoscale Model 5 (MM5) or the Weather Research and Forecasting Model (WRF) are used by meteorologists in the prediction and the study of hurricanes. The outputs from such models vary greatly depending on the model, the initialization conditions, the simulation resolution and the computational resources available. The overwhelming amount of data that is generated can become very difficult to comprehend using traditional 2D visualization techniques. We studied the presentation of such data as well as methods to compare multiple model run outputs using 3D visualization techniques in an immersive virtual environment. We also relate the experiences and opinions of two meteorologists using our system. The datasets used in our study are outputs from two separate MM5 simulation runs of Hurricane Lili (2002) and a WRF simulation run of Hurricane Isabel (2003).

In software design, physical CRC cards (Classes – Responsibilities - Collaborators) is a well-known method for rapid software-design prototyping, heavily relying on visualization and metaphors. The method is commonly applied with heuristics for encoding design semantics or denoting architectural relationships, such as card coloring, size variations and spatial grouping. Existing software-design tools are very weak in terms of interactivity, immersion and visualization, focusing primarily on detailed specification and documentation. We present a tool for visual prototyping of software designs based on CRC cards offering: 3d visualizations with zooming and panning, rotational inspection and 3d manipulators, with optional immersive navigation through stereoscopic views. The tool is accompanied with key encoding strategies to represent design semantics, exploiting spatial memory and visual pattern matching, emphasizing highly interactive software visualizations.

We present a novel custom-built 3D immersive environment, called the Immersive Cabin (IC). The IC is fully enclosed with an automatic door on the rear screen, and thus very different from existing CAVE environments. Our IC, the construction of the projection screens and stereo projectors as well as the calibration procedure are explained in details. The projectors are driven by our Visual Computing cluster for computation and rendering. Three applications that have been developed on the IC are described, 3D virtual colonoscopy, dispersion simulation for urban security, and 3D imagery and artistic creations.

Vrui (Virtual Reality User Interface) is a C++ development toolkit for highly interactive and high-performance VR applications, aimed at producing completely environment-independent software. Vrui not only hides differences between display systems and multi-pipe rendering approaches, but also separates applications from the input devices available at any environment. Instead of directly referencing input devices, e. g., by name, Vrui applications work with an intermediate

tool layer

that expresses interaction with input devices at a higher semantic level. This allows environment integrators to provide tools to map the available input devices to semantic events such as selection, location, dragging, navigation, menu selection, etc., in the most efficient and intuitive way possible. As a result, Vrui applications run effectively on widely different VR environments, ranging from desktop systems with only keyboard and mouse to fully-immersive multi-screen systems with multiple 6-DOF input devices. Vrui applications on a desktop are not run in a “simulator” mode mostly useful for debugging, but are fully usable and look and feel similar to native desktop applications.

In this work, we analyze three different registration algorithms: Chamfer distance matching, the well-known iterated closest points (ICP) and an optic flow based registration. Their pairwise combination is investigated in the context of silhouette based pose estimation. It turns out that Chamfer matching and ICP used in combination do not only perform fairly well with small offset, but also deal with large offset significantly better than the other combinations. We show that by applying different optimized search strategies, the computational cost can be reduced by a factor eight. We further demonstrate the robustness of our method against simultaneous translation and rotation.

This paper addresses the problem of the maximal recognition of hyperplanes for an invertible reconstruction of 3D discrete objects.

k

-cuspal cells are introduced as a three dimensional extension of discrete cusps defined by R.Breton. With

k

-cuspal cells local non planarity on discrete surfaces can be identified in a very straightforward way.

We have shown a supervised approach for pattern classification, which interprets the training samples as nodes of a complete arc-weighted graph and computes an optimum-path forest rooted at some of the closest samples between distinct classes. A new sample is classified by the label of the root which offers to it the optimum path. We propose a variant, in which the training samples are the nodes of a graph, whose the arcs are the

k

-nearest neighbors in the feature space. The graph is weighted on the nodes by their probability density values (pdf) and the optimum-path forest is rooted at the maxima of the pdf. The best value of

k

is computed by the maximum accuracy of classification in the training set. A test sample is assigned to the class of the maximum, which offers to it the optimum path. Preliminary results have shown that the proposed approach can outperform the previous one and the SVM classifier in some datasets.

Tile rewriting grammars are a new model for defining picture languages. In this paper we propose hexagonal tile rewriting grammars (HTRG) for generating hexagonal picture languages. Closure properties of HTRG are proved for some basic operations. We compare HTRG with hexagonal tiling systems.

The Centroidal Voronoi Diagram (CVD) is a very versatile structure, well studied in Computational Geometry. It is used as the basis for a number of applications. This paper presents a deterministic algorithm, entirely computed using graphics hardware resources, based on Lloyd’s Method for computing CVDs. While the computation of the ordinary Voronoi diagram on GPU is a well explored topic, its extension to CVDs presents some challenges that the present study intends to overcome.

Fundamental group is one of the most important topological invariants for general manifolds, which can be directly used as manifolds classification. In this work, we provide a series of practical and efficient algorithms to compute fundamental groups for general 3-manifolds based on CW cell decomposition. The input is a tetrahedral mesh, while the output is symbolic representation of its first fundamental group. We further simplify the fundamental group representation using computational algebraic method. We present the theoretical arguments of our algorithms, elaborate the algorithms with a number of examples, and give the analysis of their computational complexity.

Typical video projection systems display rectangular images on flat screens. Optical and perspective correction techniques must be employed to produce undistorted output on non-planar display surfaces. A two-pass algorithm, called projective perspective mapping, is a solution well suited for use with commodity graphics hardware. This algorithm is implemented in the OmniMap API providing an extensible, reusable C++ interface for porting 3D engines to wide field-of-view, non-planar displays. This API is shown to be easily integrated into a wide variety of 3D applications.

Two-handed control techniques for precisely and efficiently manipulating a virtual 3D object by hand in an immersive virtual reality environment are proposed. In addition, one-handed and two-handed techniques are described and comparatively evaluated. The techniques are used to precisely control and efficiently adjust an object with the speed of one hand or the distance between both hands. The controlled adjustments are actually position and viewpoint adjustments. The results from experimental evaluations show that two-handed control methods that are used to make the position and viewpoint adjustments are the best, but the simultaneous use of both one-handed and two-handed control techniques does not necessarily improve the usability.

A system is introduced for the simulation of spray painting. Head mounted display goggles are combined with a tracking system to allow users to paint a virtual surface with a spray gun. Ray tracing is used to simulate droplets landing on the surface of the object, allowing arbitrary shapes and spray gun patterns to be used. This system is combined with previous research on spray gun characteristics to provide a realistic simulation of the spray paint including the effects of viscosity, air pressure, and paint pressure. The simulation provides two different output modes: a non-photorealistic display that gives a visual representation of how much paint has landed on the surface, and a photorealistic simulation of how the paint would actually look on the object once it has dried. Useful feedback values such as overspray are given. Experiments were performed to validate the system.

Engineers and designers of various product development fields show an increasing interest in rapid prototyping techniques to help them optimize the design process of their products. In this work we present an Augmented Reality (AR) application with a model size water turbine in order to demonstrate how rapid prototyping with a hybrid prototype, simulation data of water flow characteristics and an optical AR tool can be realized. The application integrates interactive simulation, a tangible user interface and several interaction concepts for 3D CFD. Due to the intuitive and automated workflow as well as seamless process iterations, the application is easy to use by users without expert knowledge in the field of parallel simulations. Our approach points out the main benefit and problems of AR in rapid prototyping and thus provides an informative basis for future research and optimizations to offer a seamless and automated workflow.

This paper presents a simulator’s development. Its objective consists in training Colombian Navy soldiers. This device has three subsystems: A mobile platform, a graphical interface and a shooting device. The first was constructed by the connection of two linear actuators to a seat shaped, single user platform with two rotations over the horizontal axes. These actuators are activated by servomotors that are connected to a motion controller. Furthermore, the graphical interface permits the visualization of a realistic three dimensional world composed by a river, a firearm over a moving boat, targets and natural elements over the riversides. The software has the option of stereoscopic view and captures the shots provided by the third subsystem. The shooting device is the result of combining encoders with a cardan-type joint. It allows the aim movement in elevation and azimuth coordinates. Finally, preliminary tests of its potential use were conducted with satisfactory results.

The analysis of morphometric data of the vasculature of any organ requires appropriate visualization methods to be applied due to the vast number of vessels that can be present in such data. In addition, the geometric properties of vessel segments, i.e. being rather long and thin, can make it difficult to judge on relative position, despite depth cues such as proper lighting and shading of the vessels. Virtual environments that provide true 3-D visualization of the data can help enhance the visual perception. Ideally, the system should be relatively cost-effective. Hence, this paper describes a Linux-based virtual environment that utilizes a 50 inch plasma screen as its main display. The overall cost of the entire system is less than $3,500 which is considerably less than other commercial systems. The system was successfully used for visualizing vascular data sets providing true three-dimensional perception of the morphometric data.

A visualization methodology for the analysis of dynamic synchronization in electroencephalographic signals is presented here. The proposed method is based on a seeded region-growing segmentation of the time-frequency space in terms of spatial connectivity patterns, a process that can be fully automated by cleverly choosing the seeds. A Bayesian regularization technique is applied to further improve the results. Finally, preliminary results from the analysis of a high electrode-density dataset with 120 channels are shown.

3D medial axis (skeleton) extracted by a skeletonization algorithm is a compact representation of a 3D model. Among all connectivity-preservation skeletonization methods, 3D thinning algorithms are generally faster than the others. However, most 3D thinning algorithms cannot guarantee generating a unit-width curve skeleton, which is desirable in many applications, e.g. 3D object similarity match and retrieval. This paper presents a novel valence driven spatial median (VDSM) algorithm, which eliminates crowded regions and ensures that the output skeleton is unit-width. The proposed technique can be used to refine skeletons generated from 3D skeletonization algorithms to achieve unit-width. We tested the VDSM algorithm on 3D models with very different topologies. Experimental results demonstrate the feasibility of our approach.

Current approaches to cell motion analysis rely on cell tracking. In certain cases, the trajectories of each cell is not as informative as a representation of the overall motion in the scene. In this paper, we extend a cell motion descriptor and provide methods for the intuitive visualization and querying of cell motion. Our approach allows for searches of scale- and rotation-invariant motion signatures, and we develop a desktop application that researchers can use to query biomedical video quickly and efficiently. We demonstrate this application on synthetic video sets and

in vivo

microscopy video of cells in a mouse liver.

To provide better insight in bone modeling and remodeling around implants, information is extracted using different imaging techniques. Two types of data used in this project are 2D histological images and 3D SR

μ

CT (synchrotron radiation-based computed microtomography) volumes. To enable a direct comparison between the two modalities and to bypass the time consuming and difficult task of manual annotation of the volumes, registration of these data types is desired.

In this paper, we present two 2D–3D intermodal rigid-body registration methods for the mentioned purpose. One approach is based on Simulated Annealing (SA) while the other uses Chamfer Matching (CM). Both methods use Normalized Mutual Information for measuring the correspondence between an extracted 2D-slice from the volume and the 2D histological image whereas the latter approach also takes the edge distance into account for matching the implant boundary. To speed up the process, part of the computations are done on the Graphic Processing Unit.

The results show that the CM-approach provides a more reliable registration than the SA-approach. The registered slices with the CM-approach correspond visually well to the histological sections, except for cases where the implant has been damaged.

Some studies on epilepsy have shown that seizures might change the body temperature of a patient. Furthermore, other works have shown that kainic acid, a drug used to study seizures, modify body temperature of a laboratory rat. Thus, thermographic cameras may have an important role in investigating seizures. In this paper, we present the methods we have developed to measure the temperature of a moving rat subject to seizure using a thermographic camera and image processing. To accurately measure the body temperature, a particle filter tracker has been developed and tested along with an experimental methodology. The obtained measures are compared with a ground truth. The methods are tested on a 2-hour video and it is shown that our method achieves the promising results.

Medical image segmentation plays an important role in medical image analysis and visualization. The Fuzzy c-Means (FCM) is one of the well-known methods in the practical applications of medical image segmentation. FCM, however, demands tremendous computational throughput and memory requirements due to a clustering process in which the pixels are classified into the attributed regions based on the global information of gray level distribution and spatial connectivity. In this paper, we present a parallel implementation of FCM using a representative data parallel architecture to overcome computational requirements as well as to create an intelligent system for medical image segmentation. Experimental results indicate that our parallel approach achieves a speedup of 1000x over the existing faster FCM method and provides reliable and efficient processing on CT and MRI image segmentation.

The generation of natural and photorealistic images in computer graphics, normally make use of a well known method called ray tracing. Ray tracing is being adopted as a primary image rendering method in the research community for the last few years. With the advent of todays high speed processors, the method has received much attention over the last decade. Modern power of GPUs/CPUs and the accelerated data structures are behind the success of ray tracing algorithms.

kd

-tree is one of the most widely used data structures based on surface area heuristics (SAH). The major bottleneck in

kd

-tree construction is the time consumed to find optimum split locations. In this paper, we propose a prediction algorithm for animated ray tracing based on Kalman and Wiener filters. Both the algorithms successfully predict the split locations for the next consecutive frame in the animation sequence. Thus, giving good initial starting points for one dimensional search algorithms to find optimum split locations – in our case parabolic interpolation combined with golden section search. With our technique implemented, we have reduced the “running

kd

-tree construction” time by between 78% and 87% for dynamic scenes with 16.8K and 252K polygons respectively.

Ambient occlusion models the appearance of objects under indirect illumination. This effect can be combined with local lighting models to improve the real-time rendering of surfaces. We present a hardware-accelerated approach to precomputing ambient occlusion maps which can be applied at runtime using conventional texture mapping. These maps represent mesh self-occlusion computed on a per-texel basis. Our approach is to transform the computation into an image histogram problem, and to use point primitives to achieve memory scatter when accumulating the histogram. Results are presented for multiple meshes and computation time is compared with a popular alternative GPU-based technique.

We propose a 2D method based on a comics art style analysis to generate stylized comics pictures from photographs. We adapt the line drawing and the colorization according to the depth of each object to avoid depthless problems. Our model extracts image structures from the photograph, generates automatically a depth map from them and performs a rendering to give a comics style to the image. Different comics styles have been realized and are presented. Results prove that this model is suitable to create comics pictures from photographs.

This paper proposes a novel method to simulate the flow of the fluids passing through the boundary, which has not been studied in the previous works. The proposed method requires no significant modification of the existing simulation techniques. Instead, it extends the common fluid simulation techniques by adding two post-steps, adjustment and projection. Therefore, the proposed method can be easily integrated with existing techniques. Specifically, the method extends the staggered Marker-and-Cell scheme, semi-Lagrangian advection, level set method, fast marching method, etc. With the extensions, the method can successfully produce the realistic behavior of the leaking fluids.

A fully automatic algorithm for substitution of missing visual information is presented. The missing parts of a picture may have been caused by damages to or transmission loss of the physical picture. In the former case, the picture is scanned and the damage is considered as holes in the picture while, in the latter case, the lost areas are identified. The task is to derive subjectively matching contents to be filled into the missing parts using the available picture information. The proposed method arises from the observation that dominant structures, such as object contours, are important for human perception. Hence, they are accounted for in the filling process by using tensor voting, which is an approach based on the Gestalt laws of proximity and good continuation. Missing textures surrounding dominant structures are determined to maximize a new segmentation-based plausibility criterion. An efficient post-processing step based on a cloning method minimizes the annoyance probability of the inpainted textures given a boundary condition. The experiments presented in this paper show that the proposed method yields better results than the state-of-the-art.

Characteristic curves like isophotes, reflection lines and reflection circles are well–established concepts which have been used for automatic fairing of both parametric and piecewise linear surfaces. However, the result of the fairing strongly depends on the choice of a particular family of characteristic curves: isophotes or reflection lines may look perfect for a certain orientation of viewing and projection direction, but still have imperfections for other directions. Therefore, fairing methods are necessary which consider multiple families of characteristic curves. To achieve this, we first introduce a new way of controlling characteristic curves directly on the surface. Based on this, we introduce a fairing scheme which incorporates several families of characteristic curves simultaneously. We confirm effectiveness of our method for a number of test data sets.