Advances in Visual Computing

Traditional iris recognition is based on computing efficiently coded representations of discriminative features of the human iris and employing Hamming Distance (HD) as fast and simple metric for biometric comparison in feature space. However, the International Organization for Standardization (ISO) specifies iris biometric data to be recorded and stored in (raw) image form (ISO/IEC FDIS 19794-6), rather than in extracted templates (e.g. iris-codes) achieving more interoperability as well as vendor neutrality. In this paper we propose the application of quality-metric based comparators operating directly on iris textures, i.e. without transformation into feature space. For this task, the Structural Similarity Index measure (SSIM), Local Edge Gradients metric (LEG), Natural Image Contour Evaluation (NICE), Edge Similarity Score (ESS) and Peak Signal to Noise ratio (PSNR) is evaluated. Obtained results on the CASIA-v3 iris database confirm the applicability of this type of iris comparison technique.

We focus on recognizing persons according to the way they walk. Our approach considers a human movement as a set of trajectories formed by specific anatomical landmarks, such as hips, feet, shoulders, or hands. The trajectories are used for the extraction of distance-time dependency signals that express how a distance between a pair of specific landmarks on the human body changes in time as the person walks. The collection of such signals characterizes a gait pattern of person’s walk. To determine the similarity of gait patterns, we propose several functions that compare various combinations of extracted signals. The gait patterns are compared on the level of individual walk cycles in order to increase the recognition effectiveness. The results evaluated on a 3D database of walking humans achieved the recognition rate up to 96 %.

Recently, the biometric information such as faces, fingerprints, and irises has been used widely in a security system for biometric authentication. Among these biometric features which are unique to each individual, the blood vessel pattern in fingers is superior for identifying individuals and verifying their identities: We may obtain easily the information on blood vessels which is almost impossible to counterfeit because the pattern exists inside the body unlike the others. In this work, we propose a finger vein recognition method using an illumination normalization and a SIFT (Scale-Invariant Feature Transform) matching identification. To verify individual identification, the proposed methodology is composed of two steps: (i) we first normalize the illumination of finger vein images, and (ii) extract SIFT descriptors from the image and match them to the given data. Experimental results indicate that the proposed method is shown to be successful for authentication system.

This work addresses the two class classification problem within the tensor-based large margin classification paradigm. To this end, we formulate the higher rank Support Tensor Machines (STMs), in which the parameters defining the separating hyperplane form a tensor (tensorplane) that is constrained to be the sum of rank one tensors. The corresponding optimization problem is solved in an iterative manner utilizing the CANDECOMP/PARAFAC (CP) decomposition, where at each iteration the parameters corresponding to the projections along a single tensor mode are estimated by solving a typical Support Vector Machine (SVM)-type optimization problem. The efficiency of the proposed method is illustrated on the problems of gait and action recognition where we report results that improve, in some cases considerably, the state of the art.

In this paper, we propose a multi-scale integral modified census transform (MsiMCT) for eye detection. Modified census transform shows a good classification performance. However, it does not represent block properties. We therefore consider a method for overcoming this limitation. First, we propose the integral modified census transform (iMCT) using integral images, which can compute the mean intensity of the rectangular region rapidly. Using iMCT, we propose the multi-scale integral modified census transform (MsiMCT), which is structured as a concatenation of various iMCTs. The proposed MsiMCT can describe from pixel features to block features, and can therefore be implemented in many applications, such as face detection and human body detection, without modification. Our experimental results using various images show that the proposed method provides good detection accuracy, in terms of the detection rate of the number of selected weak classifiers.

Facial expressions are formed through complicated muscular actions and can be taxonomized using the Facial Action Coding System (FACS). FACS breaks down human facial expressions into discreet action units (AUs) and often combines them together to form more elaborate expressions. In this paper, we present a comparative analysis of performance of automated facial expression recognition from thermal facial videos, visual facial videos, and their fusion. The feature extraction process consists of first placing regions of interest (ROIs) at 13 fiducial regions on the face that are critical for evaluating all action units, then extracting mean value in each of the ROIs, and finally applying principal component analysis (PCA) to extract the deviation from neutral expression at each of the corresponding ROIs. To classify facial expressions, we train a feed-forward multilayer perceptron with the standard deviation expression profiles obtained from the feature extraction stage. Our experimental results depicts that the thermal imaging modality outperforms visual modality, and hence overcomes some of the shortcomings usually noticed in the visual domain due to illumination and skin complexion variations. We have also shown that the decision level fusion of thermal and visual expression classification algorithms gives better results than either of the individual modalities.

Vertebrae tracking in lumbar spinal video-fluoroscopy is the first step in the analysis of vertebrae kinematic in patients with lower back pain. This paper presents a technique to track the vertebrae using particle filters with image gradient based likelihood measurement. In the first X-ray frame, the vertebrae are semi-automatically segmented and a bi-spline curve is fitted to the landmark points to construct the vertebrae outlines; then a particle filter is used to track the vertebrae through the sequence. The proposed technique is able to track the vertebrae in both lateral and frontal video-fluoroscopy sequences. The tracking results compare well with the ground truth data obtained by manually segmenting the vertebrae.

Multi-sensory data fusion and medical image analysis often pose the challenging task of aligning dense, non-rigid and multi-modal images. However, optical sequences may also present illumination variations and noise. The above problems can be addressed by an invariant similarity measure, such as mutual information. However, in a variational setting convex formulations are generally recommended for efficiency reasons, especially when discontinuities at the motion boundaries have to be preserved. In this paper we propose the TV-MI approach, addressing for the first time all of the above issues, through a primal-dual estimation framework, and a novel approximation of the pixel-wise Hessian matrix, decoupling pixel dependencies while being asymptotically correct. At the same time, we keep a high computational efficiency by means of pre-quantized kernel density estimation and differentiation. Our approach is demonstrated on ground-truth data from the Middlebury database, as well as medical and visible-infrared image pairs.

Breast cancer is the second major type of cancer that causes mortality among women. This can be reduced if the cancer is detected at its early stage but the existing methods result in a large number of false positives/negatives. Detection of masses is more challenging. A new method for mass detection is proposed that uses textural properties of masses. A Gabor filter bank is used for this purpose. The decision of how many Gabor filters must be there in the bank and the selection of the appropriate parameters of each individual Gabor filter is critical. Particle swarm optimization (PSO) and a clustering technique are used to design and select the optimal Gabor filter bank. Support vector machine (SVM) is used as an application oriented fitness criteria. The empirical evaluation of the method over 512 ROIs from DDSM database depicts that it yields better performance (99.41%) than the traditional Gabor filter bank and other state-of-the-art methods that exploit texture properties of masses.

This paper presents a comparison of local descriptors for a set of 26 craniofacial landmarks annotated on 144 scans acquired in the context of clinical research. We focus on the accuracy of the different descriptors on a per-landmark basis when constrained to a local search. For most descriptors, we find that the curves of expected error against the search radius have a plateau that can be used to characterize their performance, both in terms of accuracy and maximum usable range for the local search. Six histograms-based descriptors were evaluated: three describing distances and three describing orientations. No descriptor dominated over the rest and the best accuracy per landmark was strongly distributed among 3 of the 6 algorithms evaluated. Ordering the descriptors by average error (over all landmarks) did not coincide with the ordering by most frequently selected, indicating that a comparison of descriptors based on their global behavior might be misleading when targeting facial landmarks.

This paper proposes a method for vision-based automated tracking of schistosomula, the etiological agent of schistosomiasis, a disease which affects over 200 million people worldwide. The proposed tracking system is intended to facilitate high-throughput and high-content drug screening against the schistosomula by taking into account their complex phenotypic response to different candidate drug molecules. Our method addresses the unique challenges in tracking schistosomula, which include temporal changes in morphology, appearance, and motion characteristics due to the effect of drugs, as well as behavioral specificities of the parasites such as their tendency to remain stagnant in dense clusters followed by sudden rapid fluctuations in size and shape of individuals. Such issues are difficult to address using current bio-image tracking systems that have predominantly been developed for tracking simpler cell movements. We also propose a novel method for utilizing the results of tracking to improve the accuracy of segmentation across all images of the video sequence. Experiments demonstrate the efficacy of the proposed tracking method.

Cell nuclei detection in fluorescent microscopic images is an important and time consuming task for a wide range of biological applications. Blur, clutter, bleed through and partial occlusion of nuclei make this a challenging task for automated detection of individual nuclei using image analysis. This paper proposes a novel and robust detection method based on the active contour framework. The method exploits prior knowledge of the nucleus shape in order to better detect individual nuclei. The method is formulated as the optimization of a convex energy function. The proposed method shows accurate detection results even for clusters of nuclei where state of the art methods fail.

The advances in computing power, availability of large- capacity storage devices and research in computer vision have contributed to recent developments in gait recognition. The ease of acquiring human videos by low cost equipments makes gait recognition much easier and less intrusive than other biometric systems. In this paper, a gait recognition system using a space-model-based approach is proposed. The proposed mechanism is able to detect moving object of interest, while track them and analyzing their gait for recognition. The system captures videos of subjects in front a stationary camera. The identification module makes use of the shape and dynamics of the system using HMM. Then it models the gait properties by accepting the feature vectors as input and model the dynamics through state transitions and observation probabilities. The experimental results show that the proposed gait recognition system successfully recognizes humans using their gait.

Human action recognition is an important area of research in computer vision. Its applications include surveillance systems, patient monitoring, human-computer interaction, just to name a few. Numerous techniques have been developed to solve this problem in 2D and 3D spaces. However most of the existing techniques are view-dependent. In this paper we propose a novel view-independent action recognition algorithm based on the motion history of skeletons in 3D. First, we compute a skeleton for each volume and a motion history for each action. Then, alignment is performed using cylindrical coordinates- based Fourier transform to form a feature vector. A dimension reduction step is subsequently applied using Principle Component Analysis and action classification is carried out by using Euclidian distance, Mahalonobis distance, and Linear Discernment analysis. The proposed algorithm is evaluated on the benchmark IXMAS and i3DPost datasets where the proposed motion history of skeletons is compared against the traditional motion history of volumes. Obtained results demonstrate that skeleton representations improve the recognition accuracy and can be used to recognize human actions independent of view point and scale.

Image matting may be defined as the extraction and composition of foreground pixels from a given image using color or opacity estimation. The alpha matting still have their own problems due to statistical optimization from small number of features. In this paper, a novel alpha matting methodology using compressive sensing, which is less dependent on the given trimap and scribbles, is proposed. The experimental results obtained for many complex natural images show that our proposed matting method can provide good mattes from either scribbles or large parts of unknown regions in a trimap.

This paper presents a video completion framework with a novel foreground extraction method based on templates. Video completion is a process of filling in missing regions of a video with appropriate fragments from the input video. In existing video completion techniques, appropriate fragments are effectively detected by separating the video into foreground and background in advance. Any existing foreground/background separation used in the completion methods assumes that the moving objects are foreground. Accordingly, undesirable moving objects are detected as foreground. These moving objects are called “dynamic background”. This paper proposes a novel foreground extraction method exploiting templates which are originally used to accelerate video completion itself. Both the efficiency and accuracy of the process are demonstrated over existing methods.

Automatic action classification is a challenging task for a wide variety of reasons including unconstrained human motion, background clutter, and view dependencies. The introduction of affordable depth sensors allows opportunities to investigate new approaches for action classification that take advantage of depth information. In this paper, we perform action classification using sparse representations on 3D video sequences of spatio-temporal kinematic joint descriptors and compare the classification accuracy against spatio-temporal raw depth data descriptors. These descriptors are used to create over-complete dictionaries which are used to classify test actions using least squares loss L

1

-norm minimization with a regularization parameter. We find that the representations of raw depth features are naturally more sparse than kinematic joint features and that our approach is highly effective and efficient at classifying a wide variety of actions from the Microsoft Research 3D Dataset (MSR3D).

In many forensic and data analytics applications there is a need to detect whether and for how long a specific person is present in a video. Frames in which the person cannot be recognized by state of the art engines are of particular importance. We describe a new framework for detection and persistence analysis in noisy and cluttered videos. It combines a new approach to tagging individuals with dynamic person-specific tags, occlusion resolution, and contact re-acquisition. To assure that the tagging is robust to occlusions and partial visibility the tags are built from small pieces of the face surface. To account for the wide and unpredictable ranges of pose and appearance variations and environmental and illumination clutter the tags are continuously and automatically updated by local incremental learning of the object’s background and foreground.

The paper introduces a methodology for the automatic classification of ancient Greek inscriptions to cutters by exploiting the three dimensional digital representation of each inscription. In particular, the authors employed surface information features extracted from 3D datasets of the letters depicted on each inscription. Therefore, implementations of various alphabet symbols are used to extract a three dimensional “ideal” prototype of the symbol for each inscription separately. Next, statistical criteria are introduced so as to reject the hypothesis that two inscriptions have been carved by the same writer, determining thus the distinct number of cutters who carved a given set of inscriptions. The remaining inscriptions are then classified to the (be) determined by the previous step cutters by maximizing resemblance likelihood of their underlined alphabet symbols “ideal” prototypes. The methodology has been applied to twenty eight Ancient Athenian inscriptions and classified them to eight different cutters. The classification results have been fully confirmed by expert epigraphists.

The concentration of the cones and ganglion cells is much higher in the fovea than the rest of the retina. This non-uniform sampling results in a retinal image that is sharp at the fixation point, where a person is looking, and blurred away from it. This difference between the sampling rates at the different spatial locations presents us with the question of whether we can employ this biological characteristic to achieve better image compression. This can be achieved by compressing an image less at the fixation point and more away from it. It is, however, known that the vision system employs more that one fixation to look at a single scene which presents us with the problem of combining images pertaining to the same scene but exhibiting different spatial contrasts. This article presents an algorithm to combine such a series of images by using image fusion in the gradient domain. The advantage of the algorithm is that unlike other algorithms that compress the image in the spatial domain our algorithm results in no artifacts. The algorithm is based on two steps, in the first we modify the gradients of an image based on a limited number of fixations and in the second we integrate the modified gradient. Results based on measured and predicted fixations verify our approach.

This paper presents an original problem of knot detection in 3D X-ray Computer Tomography images of wood stems. This image type is very different from classical medical images and presents specific geometric structures. These ones are characteristic of wood stems nature. The contribution of this work is to exploit the original geometric structures in a simple and fast algorithm to automatically detect and analyze the wood knots. The proposed approach is robust to different wood qualities, like moisture or noise, and more simple to implement than classical deformable models approaches.

We demonstrate that one can adapt recent diffusion-based image compression techniques such that they become ideally suited for steganographic applications. Thus, the goal is to embed secret images within arbitrary cover images. We hide only a small number of characteristic points of the secret in the cover image, while the remainder is reconstructed with edge-enhancing anisotropic diffusion inpainting. Even when using significantly less than 1% of all pixels as characteristic points, sophisticated shapes of the secret can be clearly identified. Selecting more characteristic points results in improved image quality. In contrast to most existing approaches, this even allows to embed large colour images into small grayscale images. Moreover, our approach is well-suited for uncensoring applications. Our evaluation and a web demonstrator confirm these claims and show advantages over JPEG and JPEG 2000.

The fruit fly,

Drosophila melanogaster

, is a well established model organism used to study the mechanisms of both learning and memory

in vivo

. This paper presents video analysis algorithms that generate data that may be used to categorize fly behaviors. The algorithms aim to replace and improve a labor-intensive, subjective evaluation process with one that is automated, consistent and reproducible; thus allowing for robust, high-throughput analysis of large quantities of video data. The method includes tracking the flies, computing geometric measures, constructing feature vectors, and grouping the specimens using clustering techniques. We also generated a Computed Courtship Index (CCI), a computational equivalent of the existing Courtship Index (CI). The results demonstrate that our automated analysis provides a numerical scoring of fly behavior that is similar to the scoring produced by human observers. They also show that we are able to automatically differentiate between normal and defective flies via analysis of their videotaped movements.

This paper presents an image construction tool for biological image visualization and education using image matching and stitching approaches. The image matching technique is based on the algorithm SURF (Speeded-up Robust Feature) [3, 4], a successor to the popular feature detection algorithm SIFT (Scale Invariant Feature Transform) [1, 2]. Unlike a traditional image stitching approach, our tool assumes that biological images are taken on a linear model with similar degrees of overlap and orientation angle towards ground from air. With these aspects in mind, generated panoramas will display less distortion and more raw valuable details. Such a tool will facilitate the scientific research and education through applications of visual information processing in fields of Biology, Astronomy, Geology, etc.

Modern time-varying flow visualization techniques that rely on advection are able to convey fluid transport, but cannot provide an accurate insight into local flow behavior over time or locally corresponding patterns in unsteady vector fields. We overcome these limitations of purely Lagrangian approaches by generalizing the concept of function fields to time-varying flows. This representation of unsteady vector-fields as stationary function fields, where every position in space is a vector-valued function supports the application of novel analysis techniques based on function correlation, and allows to answer data analysis questions that remain unanswered with classic time-varying vector field analysis techniques. Our results demonstrate how analysis of time-varying flow fields can benefit from a conversion into function field representations and show the robustness of our presented clustering techniques.

We present a novel way to efficiently compute Riemannian geodesic distance over a two-dimensional domain. It is based on a previously presented method for computation of geodesic distances on surface meshes. Our method is adapted for rectangular grids, equipped with a variable anisotropic metric tensor. Processing and visualization of such tensor fields is common in certain applications, for instance structure tensor fields in image analysis and diffusion tensor fields in medical imaging.

The included benchmark study shows that our method provides significantly better results in anisotropic regions and is faster than current stat-of-the-art solvers. Additionally, our method is straightforward to code; the test implementation is less than 150 lines of C++ code.

Different taxi drivers may use different strategies to choose operating regions and find customers, which is called mobility intelligence. In this paper, we present a visualization system to analyze a large amount of spatial-temporal multi-dimensional trajectory data and identify some key factors that differentiate the top drivers and ordinary drivers according to their income. Two novel encoding schemes, Choice-of-Location graph and Move/Wait Strategy tree, have been proposed to analyze drivers’ behaviors when choosing operating locations and drivers’ move/wait strategies when their taxis are vacant.We have applied our system to the trajectories of thousands of taxis in a major city and have gained some interesting findings on taxi drivers’ mobility intelligence.

Multi-frame rate systems decouple viewing from rendering in an asynchronous pipeline. Multiple GPUs can be used as frame sources, while a primary GPU is responsible for viewing and display update. Conventionally, the last rendering result is used for display. However, modern GPUs are equipped with a fairly large amount of memory which allows frames to be cached in video memory. As long as the data is static, caching allows for a more sophisticated reference frame selection that increases the output quality. With a growing frame database, images for most viewpoints can be queried from the cache and the system converges into a conventional image-based rendering system. However, multi-frame rate systems use purely virtual image sources. As a consequence, the rendering process can be actively steered by the viewing process, which allows for advanced strategies. Moreover, by picking multiple reference frames from the cache, we can avoid display artifacts arising from occlusions.

Detecting suspicious or malicious user behavior in large networks is an essential task for administrators which requires significant effort due to the huge amount of log data to be processed. However, several of these activities can be rapidly identified since they usually demonstrate periodic behavior. For instance, periodic activities by specific users accessing the billing system of a financial institution may conceal fraud. Detecting periodicity in user behavior not only offers security to the network, but may prevent future malicious activities. In this paper, we present visualization techniques that aim to detect authorized (or unauthorized) user activities that seem to appear at regular time intervals.

We present a practical implementation of a variation of the Turing Test for realistic computer graphics. The test determines whether virtual representations of objects appear as real as genuine objects. Two experiments were conducted wherein a real object and a similar virtual object is presented to test subjects under specific restrictions. A criterion for passing the test is presented based on the probability for the subjects to be unable to recognise a computer generated object as virtual. The experiments show that the specific setup can be used to determine the quality of virtual reality graphics. Based on the results from these experiments, future versions of the Graphics Turing Test could ease the restrictions currently necessary in order to test object telepresence under more general conditions. Furthermore, the test could be used to determine the minimum requirements to achieve object telepresence.

The aim of this paper is to present a 3D hybrid shape construction that benefits from discrete and continuous modeling approaches. The proposed technique addresses the problem of automated modeling of virtual scene components such as caves, buildings and clouds. The approach combines two independent methods well known in three-dimensional computer graphics: shape grammar and shape morphing. The modeled structures are characterized by geometrical complexity with inner graph structure more optimized than in classical CSG approach. In this paper, we mainly focus on the description of the algorithm.

In applications of augmented reality it is an essential task to retrieve the camera pose for correct overlay with virtual content. This can be realized by using a model-based camera tracking approach that fits a given model of the scene to the images captured by the camera. These systems have to be initialized properly for the pose estimation process of continuous tracking. We present a two-step concept for the global initialization of such model-based tracking systems. With a model database and known GPS coordinates as well as compass orientation, it is possible to determine which part of the scene is visible and to obtain a first rough pose. We also introduce a method to refine the initialization pose to overcome GPS inaccuracies. It has been successfully tested in an urban context.

We present a technique for real-time rendering of teeth with no need for computational or artistic preprocessing. Teeth constitute a translucent material consisting of several layers; a highly scattering material (dentine) beneath a semitransparent layer (enamel) with a transparent coating (saliva). In this study we examine how light interacts with this multilayered structure. In the past, rendering of teeth has mostly been done using image-based texturing or volumetric scans. We work with surface scans and have therefore developed a simple way of estimating layer thicknesses. We use scattering properties based on measurements reported in the optics literature, and we compare rendered results qualitatively to images of ceramic teeth created by denturists.

We present a comparison of five freely available physics engines with specific focus on robotic assembly simulation in virtual reality (VR) environments. The aim was to evaluate the engines with generic settings and minimum parameter tweaking. Our benchmarks consider the minimum collision detection time for a large number of objects, restitution characteristics, as well as constraint reliability and body inter-penetration. A further benchmark tests the simulation of a screw and nut mechanism made of rigid-bodies only, without any analytic approximation. Our results show large deviations across the tested engines and reveal benefits and disadvantages that help in selecting the appropriate physics engine for assembly simulations in VR.

This paper describes the setup and development of an innovative framework for conducting user tests in a virtual environment. As it is the main purpose to evaluate the user’s reactions to user interfaces on mobile devices, an Android-based smartphone and a tablet computer were linked to a Virtual Reality (VR) system. The framework allows user interaction to trigger certain events in the immersive environment and vice versa interaction with the virtual environment can trigger interface behaviour. For hand-free navigation through the virtual scene a Wii Balance Board is used. The framework is the basis for user tests to be conducted within the uTRUSTit (Usable Trust in the Internet of Things) project supported by the EU under the Seventh Framework Programme.

Automatic pain recognition is an evolving research area with promising applications in health care. In this paper, we propose the first fully automatic approach to continuous pain intensity estimation from facial images. We first learn a set of independent regression functions for continuous pain intensity estimation using different shape (facial landmarks) and appearance (DCT and LBP) features, and then perform their late fusion. We show on the recently published UNBC-MacMaster Shoulder Pain Expression Archive Database that late fusion of the afore-mentioned features leads to better pain intensity estimation compared to feature-specific pain intensity estimation.

Automatic recognition of face sketches is a challenging problem. It has application in forensics. An artist drawn sketch based on the descriptions from the witnesses can be used as the test image to recognize a person from the photo database of suspects. In this paper, we propose a novel method for face sketch recognition. We use the edge features of a face sketch and face photo image to create a feature string called ’edge-string’. The edge-strings of the face photo and face sketch are then compared using the Smith-Waterman algorithm for local alignments. The results on CUHK (Chinese University of Hong Kong) student dataset show the effectiveness of the proposed approach in face sketch recognition.

Psychological evidence suggests that the human ability to recognize facial expression improves with the addition of temporal stimuli. While the facial action coding community has largely migrated towards temporal information, the facial expression recognition community has been slow to utilize facial dynamics. This paper contrasts the contributions of static vs. temporal features, including both dense and sparse facial tracking methodologies in combination with sparse representation classification. The temporal methods of facial feature point tracking, motion history images, free form deformation, and SIFT flow are adapted for facial expression classification. Dense optical flow for facial expression recognition is successfully utilized. We show that when used in isolation, the best temporal methods are just as good as static methods. However, when fusing temporal dynamics with static imagery significant increases in facial expression classification are achieved.

We show that the results from searching 2D images or a video sequence with a 3D head model can be improved by using detailed sub-models. These parts are initialised with the full model result and are allowed to search independently of that model, and each other, using the same algorithm. The final results for the sub-models can be reported exactly, or optionally fed back into the full model to be constrained by its parameter space. In the case of a video sequence this can then be used in the initialisation of the next frame. We tested various data sets, constrained and unconstrained, including a variety of lighting conditions, poses, and expressions. Our investigation showed that using the sub-models improved on the original full model result on all but one of the data sets.

Gender recognition from facial images plays an important role in biometric applications. We investigated Dyadic wavelet Transform (DyWT) and Local Binary Pattern (LBP) for gender recognition in this paper. DyWT is a multi-scale image transformation technique that decomposes an image into a number of subbands which separate the features at different scales. On the other hand, LBP is a texture descriptor and represents the local information in a better way. Also, DyWT is a kind of translation invariant wavelet transform that has better potential for detection than DWT (Discrete Wavelet Transform). Employing both DyWT and LBP, we propose a new technique of face representation that performs better for gender recognition. DyWT is based on spline wavelets, we investigated a number of spline wavelets for finding the best spline wavelets for gender recognition. Through a large number of experiments performed on FERET database, we report the best combination of parameters for DyWT and LBP that results in maximum accuracy. The proposed system outperforms the stat-of-the-art gender recognition approaches; it achieves a recognition rate of 99.25% on FERET database.

Domes are architectural structural elements characteristic for ecclesiastical and secular monumental buildings, like churches, basilicas, mosques, capitols and city halls. In the scope of building facade architectural style classification the current paper addresses the problem of architectural style classification of facade domes. Building facade classification by architectural styles is achieved by classification and voting of separate architectural elements, like domes, windows, towers, etc. Typical forms of the structural elements bear the signature of each architectural style. Our approach classifies domes of three architectural styles - Renaissance, Russian and Islamic. We present a three-step approach, which in the first step analyzes the height and width of the dome for the identification of Islamic saucer domes, in the second step detects golden color in YCbCr color space to determine Russian golden onion domes and in the third step performs classification based on dome shapes, using clustering and learning of local features. Thus we combine three features - the relation of dome width and height, color and shape, in a single methodology to achieve high classification rate.

In this paper we deal only contour detection based on image analogy principle which has been used in super resolution images, texture, curves synthesis and interactive editing. Human is able to hand drawn best outlines that may considered as benchmarks for contour detection and image segmentation algorithms. Our goal is to model this expertise and to pass on it at the computer for contour detection. Giving a reference image where outlines are drawn by human, we propose a method based on the learning of this expertise to locate outlines of a query image in the same way that is done for the reference. Experiments are conducted on different data sets and the obtained results are presented and discussed.

This paper presents a new texture representation, the volume Trace transform, based on several Trace transform. The volume Trace transform (VTT) is constructed using multi-trace functional to produce salient features. The VTT is transformed to a distinctive compact representation using our proposed method, the discriminant feature transform (DFT). DFT is a 2-D histogram. The histogram is evaluated by chi-square test statistics. The experimental result was conducted on Brodatz texture database.

We present a new unsupervised metric for segmentation result evaluation based on Bayes classification error and image global contrast. First, we presented a comparative study between several unsupervised metrics in order to prove their limits. The qualitative study was performed to make a preliminary selection and to discard some measures unsuitable for evaluation of foreground/background segmentation on flower images. For the quantitative study, we proposed a validation protocol based on the vote technique and involving a comparison to the ground truth. Experiments were performed on Oxford flower dataset in order to select the best result between different segmentation results. The obtained result showed that our proposed metric gives the best results.

Robust hand tracking is highly demanded for many real-world applications relevant to human machine interface. However, current methods achieve no satisfactory robustness in real environments. In this paper a novel hand tracking method was proposed integrating online Hough Forest and Flocks-of-Features tracking. Skin color was integrated in the Hough Forest framework to gain more robustness against drastic hand appearance and pose changes, especially against partial occlusions. Also a novel multi-cue Flocks-of-Features tracking algorithm based on computer graphics was integrated in to enhance the framework’s robustness against distractors and background clutter. Additionally, recovery from tracking failure was addressed. Lots of experiments were carried out to evaluate our method, also to compare it with CAMShift, Hough Forest tracker, and the original Flocks-of-Features Tracker, and showed the effectiveness of our method.

Whereas common Vickers indentation segmentation algorithms are precise with high quality images, low quality images often cannot be segmented appropriately. We investigate an approach, where unfocused images are segmented. On the one hand, the segmentation accuracy of low quality images can be improved. On the other hand we aim in reducing the overall runtime of the hardness testing method. We introduce one approach based on single unfocused images and one gradual enhancement approach based on image series.

We propose a GPU-based algorithm for texture analysis and synthesis of nearly-regular patterns, in our case scanned textiles or similar manufactured surfaces. The method takes advantage of the highly parallel execution on the GPU to generate correlation maps from captured template images. In an analysis step a lattice encoding the periodicity of the texture is computed. This lattice is used to synthesize the smallest texture tile describing the underlying pattern. Compared to other approaches, our method analyzes and synthesizes a valid lattice model without any user interaction. It is robust against small distortions and fast compared to other, more general approaches.

We develop a higher dimensional version of the Canny edge detection algorithm. The Canny operation detects the zero-crossing of the gradient of the Gaussian-convolved image. The segment edge curve detected by the Canny operation is an approximation of zero-crossing of bilinear form defined by second order derivative of an image. This definition of edge points of segment is dimension independent. This definition also allows us to extend the filtering operation from the Gaussian convolution to general linear and non-linear ones.

Estimation of human body orientation is an important cue to study and understand human behaviour, for different tasks such as video surveillance or human-robot interaction. In this paper, we propose an approach to simultaneously estimate the body orientation of multiple people in multi-view scenarios, which combines a 3D human body shape and appearance model with a 2D template matching approach. In particular, the 3D model is composed of a generic shape made up of elliptic cylinders, and a 3D colored point cloud (appearance model), obtained by back-projecting pixels from foreground images onto the geometric surfaces. In order to match the reconstructed appearance to target images in arbitrary poses, the appearance is re-projected onto each of the different views, by generating multiple templates that are pixel-wise, robustly matched to the respective foreground images. The effectiveness of the proposed approach is demonstrated through experiments in indoor sequences with manually-labeled ground truth, using a calibrated multi-camera setup.

We here present a new approach to characterize similar areas of two 2D point clouds, which is a major issue in Pattern Recognition and Image Analysis.

To do so, we define a similarity measure that takes into account several criteria such as invariance by rotation, outlier elimination, and one-dimensional structure enhancement. We use this similarity measure to associate locations from one cloud to the other, to use this result in the frame of a registration process between these two point clouds.

Our main contributions are the integration of various one-dimensional structure representations into a unified formalism, and the design of a robust estimator to evaluate the common information related to these structures.

Finally, we show how to use this approach to register images of different modalities.

In this paper we examine the effectiveness of using k-means clustering to build a codebook for image classification. Our findings suggest that k-means clustering does not result in a better codebook than a random sampling of training features. An alternative strategy that uses k-means to cluster features, then ultimately selects a feature from each cluster, also fails to significantly outperform random sampling. We present data in support of our claims, gathered using the training set for the PASCAL Visual Object Classes Challenge. Additionally, we provide a theoretical justification for our findings.

In planning navigation schemes for social robots, distinguishing between humans and other obstacles is crucial for obtaining a safe and comfortable motion. A Kinect camera is capable of fulfilling such a task but unfortunately can only deliver a limited field of view (FOV). Recently a lens that is capable of improving the Kinect’s FOV has become commercially available from Nyko. However, this lens causes a distortion in the RGB-D data, including the depth values. To address this issue, we propose a two-staged undistortion strategy. Initially, pixel locations in both RGB and depth images are corrected using an inverse radial distortion model. Next, the depth data is post-filtered using 3D point cloud analysis to diminish the noise as a result of the undistorting process and remove the ground/ceiling information. Finally, the depth values are rectified using a neural network filter based on laser-assisted training. Experimental results demonstrate the feasibility of the proposed approach for fixing distorted RGB-D data.

The paper describes the methodology of computer-based measurement for taking anthropometric dimensions. We focus on a 2D anthropometry system which is being designed for the purposes of a clothing company. The system consists of computer software, a digital camera and a background board with calibration dots. Only a few steps are needed to calibrate such a system and prepare it for the measurements. The measured person is captured, his silhouette is extracted and body dimensions are computed. Our research is intended to enrich existing techniques for automatic detection of anatomical landmarks (and body parts such as waist, chest, etc.) on the extracted silhouette. This step is very important to create a complete system without any need of user interaction and it must correspond to the body parts definitions given by clothing standards and tailors. Several potential sources of problems are discussed and some possible solutions are proposed.

This paper presents a novel implementation of the JPEG2000 standard as a system on a chip (SoC). While most of the research in this field centers on acceleration of the EBCOT Tier I encoder, this work focuses on an embedded solution for EBCOT Tier II. Specifically, this paper proposes using an embedded softcore processor to perform Tier II processing as the back end of an encoding pipeline. The Altera NIOS II processor is chosen for the implementation and is coupled with existing embedded processing modules to realize a fully embedded JPEG2000 encoder. The design is synthesized on a Stratix IV FPGA and is shown to out perform other comparable SoC implementations by 39% in computation time.

As part of a strategy for creating 3D models of engineering objects from sketched input, we attempt to identify

design features

, geometrical structures within objects with a functional meaning. Our input is a 2D B-Rep derived from a single view sketch of a polyhedral shape. In this paper, we show how to use suitable cues to identify algorithmically two additive engineering design features, angular and linear ribs.

Integrated sensor orientation (ISO) is an alternative approach in determining the exterior orientation parameters of the camera at the time of exposure. The technique combines the advantages of the direct geocoding and the conventional aerial triangulation. It exploits directly measured orientation parameters as constraints for a bundle adjustment process. This technique is more timely efficient than the traditional adjustment in which ground control points can be eliminated. Also, its accuracy is superior to that of direct geocoding and shows the potential in large scale applications. The accuracy of ISO depends heavily on the accuracy of the directed measured EOs and the distribution and number of tie-points. Various combinations of input produce varying results of ISO. This paper thus presents a design framework of an ISO simulator that can assist operators to investigate real ISO systems with controllable parameters.

Automatic Design of Algorithms through Evolution (ADATE) is a system for fully automatic programming that has the ability to either generate algorithms from scratch or improve existing ones.

In this paper, we employ ADATE to improve a standard image processing algorithm, namely graph based segmentation (GBS), which has emerged as one of the very most popular methods for image segmentation, that is partitioning an image into regions.

The key contribution of the paper is to show that a proven and well-known computer vision code is easy to improve through automatic programming. This may presage a change to the entire field of computer vision where automatic programming becomes a routine way of improving standard as well as state-of-the art image processing and pattern analysis algorithms.

GBS was mostly chosen as case study to investigate how useful the ADATE automatic programming system may be in computer vision. Numerous other algorithms in the field could have been chosen instead.

This paper presents the possibilities of using a compact digital camera, which is amateur class of equipment, to a simple analysis of the deformation of mining chains during the tests performed at percussive, dynamic load. When the mining chain works during underground coal-bed exploitation, in the scraper conveyor, which is one of the elements of a longwall system, there could be observed frequent chains fractures due to its dynamic percussive load. It often occurs as a result of an emergency chain locking in the troughs of the chain conveyor. For the purpose of analysis special software was created. Thus it was possible to obtain at low cost, with the assumed accuracy of deformation measurement, deformations as a function of time, without installing on the monitored elements acceleration sensors (e.g. inductive, potentiometric), which were usually destroyed during the tests, mostly due to their low resistance to shock.

We introduce a spatial-based feature approach to locating and recognizing objects in cases where several identical or similar objects are accumulated together. With respect to cognition, humans specify such cases with a group-based reference system, which can be considered as an extension of conventional notions of reference systems. On the other hand, a spatial-based feature is straightforward and distinctive, making it more suitable for object recognition tasks. We evaluate this approach by testing it on eight diverse object categories, and thereby provided comprehensive results. The performance exceeds the state of art by high accuracy, less attempts and fast running time.

This paper proposes how to utilize pose exemplars in the prediction step of particle filter for efficient human pose estimation. The prediction of particle filter is only dependent on the previous posterior distribution. If observation data and reference dataset are used in prediction, the prediction range can be more compact and precise. We use adaptive exemplars for prediction. To do so the similarity between pose exemplar and the pose silhouette observation are measured. Based on the similarity of exemplars corresponding number of particles are predicted either from exemplars and previous posterior distribution. After pose estimation with the likelihoods of predicted particles, the finally estimated pose is used for updating adaptive exemplar dataset for improving performance of next prediction. Therefore, the proposed method efficiently estimates the pose of articulated full body as resultant images represent.

To use omni-directional images obtained by a rotating line camera for indoor services, we should know the position and attitude of the camera at the acquisition time to register the images with respect to an indoor coordinate system. In this study, we thus develop a method for the estimation of the extrinsic orientation parameters of an omni-directional image. First, we derive a collinearity equation for the omni-directional image by geometrically modeling the rotating line camera. We then estimate the extrinsic orientation parameters (EOP) through the collinearity equations with indoor control points which are stochastic constraints. The experimental results indicate that the extrinsic orientation parameters are estimated with the precision of ±1.4 mm and ±0.05° for the position and attitude, respectively. The residuals are within ±3.11 and ±9.20 pixels in horizontal and vertical directions. Using the proposed method for estimating EOP of indoor omni-directional images, we can generate sophisticated indoor 3D models and offer precise indoor services to users based on the models.

In laser drilling, an assist gas is often used to remove material from the drilling point. In order to design assist gas nozzles to minimize spatter formation, measurements of spatter trajectories are required.

We apply computer vision methods to measure the 3D trajectories of spatter particles in a laser cutting event using a stereo camera configuration. We also propose a novel method for calibration of a weak perspective camera that is effective in our application.

The proposed method is evaluated with both computer-generated video and video taken from actual laser drilling events. The method performs well on different workpiece materials.

This paper presents a novel method that can detect license plates which have large variations including perspective distortion, size variation, blurring. Spatial combinations of covariance descriptors in different positions are used with feed-forward network to extract plate-like region and HOG descriptor is used with LDA for validation. From this method, we could achieve high detection rate 94% while maintaining low FPPW(2.5

− 6

) in road view image.

In this paper, we proposed a fast intra prediction algorithm for H.264/AVC. Although the H.264/AVC achieved superior coding performance compared with previous video coding standards, the coding computational complexity is also considerable. To reduce the complexity, we select proper candidate intra prediction modes using the angle of the pixel differences along the horizontal and vertical direction to be involved in the RDO, and it results in around 77% of encoding time reduction with negligible coding performance loss. Thus, the proposed algorithm can be utilized usefully for real time high quality high quality video application.

In this paper, we introduce an efficient global image registration for medical images. We use the nearest neighbour method for searching appropriate reference image. The NNS based image registration uses pre-computed references in the image-dictionary as the targets of image registration and search the best matched reference using NNS. For speeding up, we apply the random projection to image for reduction of the image size. For the reduction of the number of data in the dictionary, we use interpolation of references images in the image dictionary.

Most commercial visualization tools support an easy and quick creation of conventional time-oriented visualizations such as line charts, but customization is limited. In contrast, some academic visualization tools and programming languages support the creation of some customizable time-oriented visualizations but it is time consuming and hard. To combine

efficiency

, the effort required to develop a visualization, and

customizability

, the ability to tailor a visualization, we developed time-oriented building blocks that address the specifics of time (e.g. linear vs. cyclic or point-based vs. interval-based) and consist of inner customizable parts (e.g. ticks). A combination of the time-oriented and other primitive graphical building blocks allowed the creation of several customizable advanced time-oriented visualizations. The appearance and behavior of the blocks are specified using spreadsheet-like formulas. We compared our approach with other popular visualization tools. Evaluation showed that our approach rates well in customizability.

Bioinformatics research in general and the exploration of metabolic networks in particular rely on processing data from different sources. Visualization in this context supports the exploration process and helps to evaluate the data quality of the used sources.

In this work, we extend our existing metabolic network visualization toolbox and hereby address the fundamental task of comparing metabolic networks from two major bioinformatics resources for the purpose of data validation and verification. This is done on different levels of granularity by providing an overview on retrieval rates of chemical compounds and reactions per pathway on the one hand, as well as giving a detailed insight into the differences in the biochemical reaction networks on the other.

Good information design is very important for human computer interfaces since it improves productivity and enhances human understanding. This paper proposes a novel algorithm for estimation of deployment of the visual elements of the GUI. The interface layout can be either created manually or be the output of computer simulation locating uniform rectangular blocks on the layout. Those ‘black-box’ blocks can be replaced by the system dependent objects representations, e.g. visual metaphors of the heating system. The proposed “shadows” algorithm can be used in visual rating of various GUI models. The paper discusses the theoretical background, the properties of the proposed algorithm together with the sample prototype application.

This paper presents a novel framework for optimizing the visual analysis of network related information, and in particular of Border Gateway Protocol (BGP) updates, using information theoretic measures of both the underlying data and the visual information. More precisely, a hierarchical visualization scheme is proposed using a graph metaphor that is optimized, with respect to information theoretic metrics of several visual mapping parameters. Experimental demonstration in state-of-the-art BGP events, illustrate the flexibility of the proposed framework and the significant analytics effect of the proposed optimization scheme.

We present InShape, an in-situ shape-based multiple-view selection interface for interactive exploration of dense tube-based diffusion magnetic resonance imaging (DMRI) visualizations. An optimal experience in such exploration demands concentration on the tract of interest (TOI). InShape facilitates such workflow by leveraging three design principles: (1) shape-enabled precise selection; (2) in-the-flow multi-views for comparison; (3) sculpture-based removal. Results of a pilot study suggested that users have the best interaction experience when the widget shapes match the targeted selection shape. We also found that widget design without losing the flow of operations facilitates focused control. Finally, quick sculpture can help reach the target selection fibers quickly. The contributions of this work are the design principles, together with discussions of usability considerations in interactive exploration in dense 3D DMRI environments.

We present an algorithm to generate an interpolation or approximation model consisting of many patches from a triangle mesh, and each patch is a weighted combination of the three surfaces associated with the vertices of a triangle. Moreover, to make the whole surface include fewer patches, mesh simplification is introduced into the process of surface construction. The algorithm takes a triangle mesh and a given error as input, and iteratively deletes vertex whose distance to the surface model constructed from the simplified mesh is less than or equal to the given error until convergence. Since the method is based on surface approximation and vertex deletion, it allows us to control the error between the generated model and the original mesh precisely. Furthermore, many experimental results show that the generated models approximate the original models well.

Generating quadrilateral meshes is very important in many industrial applications such as finite element analysis and B-spline surface fitting. However, it is still a challenging task to design appropriate vertex connectivity in the quadrilateral meshes by respecting the shapes of the target object and its boundary. This paper presents an approach for interactively editing such mesh topology in quadrilateral meshes by introducing a 2D diffusion tensor field to the interior of the target object. The primary idea is to track the two principal directions of the tensor field first and then construct the dual graph of the quadrilateral mesh, so that we can control the mesh topology through the design of the underlying 2D diffusion tensor field. Our method provides interactive control of such mesh topology through editing the orientations of the tensor samples on the boundary of the target object. Furthermore, it also allows us to intentionally embed degeneracy inside the object to introduce extraordinary (i.e., non-degree-four) vertices according to user requirements.

Finding which triangle in a planar triangle mesh contains a query point is one of the most frequent tasks in computational geometry. Usually, a large number of point locations has to be performed, and so there is a need for fast algorithms resistant to changes in triangulation and having minimal additional memory requirements. The so-called walking algorithms offer low complexity, easy implementation and negligible additional memory requirements, which makes them suitable for such applications. In this paper, we propose a walking algorithm which significantly improves the current barycentric approach and propose how to effectively combine this algorithm with a suitable hierarchical structure in order to improve its computational complexity. The hierarchical data structure used in our solution is easy to implement and requires low additional memory while providing a significant acceleration thanks to the logarithmic computational complexity of the search process.

In the paper, we present a real-time algorithm optimization based on a gaze point position. The data is provided by the eye tracker and allows the effective rendering of algorithms in real-time where their accuracy depends on the distance from the point-of-regard. We present the approach based on the algorithm simulating the subsurface scattering effect. The model complexity depends on the number of parameters that should be taken into account to simulate a real effect. In order to specify model parameters correctly dependent on the gaze point position, a series of perceptual experiments must be conducted. The quality of the results will be evaluated on the basis of the single stimulus perceptual metrics.

An approach for auto-calibration and validation of depth measurements gained from range cameras is introduced. Firstly, the geometry of the scene is reconstructed and its surface normals are computed. These normal vectors are segmented in 3D with the Mean-Shift algorithm and large planes like walls or the ground plane are recovered. The 3D reconstruction of the scene geometry is then utilized in a novel approach to derive principal camera parameters for range or depth cameras. It operates based on a single range image alone and does not require special equipment such as markers or a checkerboard and no specific measurement procedures as are necessary for previous methods. The fact that wrong camera parameters deform the geometry of the objects in the scene is utilized to infer the constant depth error (the phase offset for continuous wave ToF cameras) as well as the focal length. The proposed method is applied to ToF cameras which are based on the Photonic Mixer Device to measure the depth of objects in the scene. Its capabilities as well as its current and systematic limitations are addressed and demonstrated.