WSCG '2017: Short Papers Proceedings (2024)

Lens flare rendering in computer games has always been lacking. While physically motivated solutions based onreal camera systems exist, their inherent complexity makes them inappropriate for widespread use. Developers arenot trained to work with these complicated imaging systems, and even if they were, artists and designers preferintuitive parameters over these complicated optical system descriptions. More support is needed to start adoptingto these advanced models, and to show that with an appropriate tool the situation can be vastly improved. Thispaper describes OpenLensFlare, an open-source C++ framework for rendering convincing physically-based lensflare effects. Additionally, a supporting optical system editor is also provided, which is capable of producing andvisualizing optical systems and showing an example usage of the run-time library. Together, these two systems canbe used to replace the existing naive algorithms, with small effort and low integration complexity.

Persons with disabilities can face considerable challenges accessing many computing systems, such as cloud computing.We created six low-cost user interfaces using: keyboard-based, touchable, speech-based, touch-less gesture,tactile, and then combined them all in one user interface termed Crossover User Interface Model (CoUIM).We measured inclusiveness, error occurrence, user performance, and user satisfaction though an IRB approvedstudy of twenty-nine participants. We chose Xen cloud platform to evaluate our research. We focused on threegroups of users: persons with no disability, persons with blind and visually impairment (B/VI), and persons withmotor-impairment. When we combined several interactions in one user interface, results improved for persons withdisability. Using CoUIM improved inclusiveness, error rate, user performance and even user satisfaction. Personswith motor disability needed a little more time to complete the same tasks in our study. In particular, we show thatpersons with blind and visually impairment (B/VI) can compete on equal footing with their sighted peers based onerror rate and time to complete the tasks using CoUIM.

The United Nations Economic Commission for Europe defines a safety regulation based on the possible collisionbetween the driver’s knee and an automobile’s interior parts. The “knee-form” apparatus is used to evaluatecompliance with this regulation. Current software for analyzing possible collisions of the knee-form is notapplicable to the part whose surface is vertical and near parallel to the knee-form approaching direction. In thispaper, we propose a novel algorithm named “push-and-slide” for extracting the knee-form colliding area on theconsole part and door panel. In the first step of the algorithm, the target surface of the part is transformed to gridlikepoints in a high resolution. The knee-form models in various positions and orientations are prepared in thesecond step. Each knee-form model is pushed to the grid-like points. The model is then moved along the partsurface to detect possible collisions between the knee-form and the grid-like points. An experimental system isdeveloped and some computational experiments are performed.

This study improves the ergonomics of using the Leap Motion hand tracking device with an Oculus Rift. The improvementswere realised through the use of a 3D printed mount that angled the Leap Motion down by 30 degrees.This allowed for users to interact with a virtual environment in which their arms may be held in a biomechanicallyless stressful location, rather than up and in front of their face. To validate the configuration, 15 participants completeda specially designed task which involved pressing virtual buttons in a given location. The button pressingtask was performed in three configurations that compared the angled mount against the standard forward facingmount. Results indicate that the angled mount eliminates tracking loses, whilst producing comparable accuracyagainst the control condition and allowing the participant to interact in a more natural arm posture.

Digital images and digital image processing facilitated significant progress in numerous areas where medicineis an important one of them. Computer-aided detection and diagnostics systems are used to assist specialists ininterpretation of medical digital images. One of the important research issues is detection and classification of thechronic obstructive pulmonary disease in lung CT images. In this paper we proposed a method for emphysemaclassification based on texture and intensity features. Only six different characteristics of the uniform local binarypattern and intensity histogram were used as input vector for support vector machine that was used as classifier.Feature vector was significantly reduced compared to the other state-of-the-art methods while the classificationaccuracy was increased. On images from standard dataset global accuracy of our proposed algorithm was 98.18%compared to 95.24% and 93.9% of two other compared algorithms.

In biology, the morphometric analysis is widely used to analyze the inter-organisms variations. It allows to classifyand to determine the evolution of an organism’s family. The morphometric methods consider features such asshape, structure, color, or size of the studied objects. In previous works [8], we have analyzed beetle mandiblesby using the centroid as feature, in order to classify the beetles. We have shown that the Probabilistic HoughTransform (PHT) is an efficient unsupervised method to compute the centroid. This paper proposes a new approachto precisely estimate the landmark geometry, points of interest defined by biologists on the mandible contours. Inorder to automatically register the landmarks on different mandibles, we defined patches around manual landmarksof the reference image. Each patch is described by computing its SIFT descriptor. Considering a query image, weapply a registration step performed by an Iterative Principal Component Analysis which identify the rotation andtranslation parameters. Then, the patches in the query image are identified and the SIFT descriptors computed.The biologists have collected 293 beetles to provide two sets of mandible images separated into left and right side.The experiments show that, depending on the position of the landmarks on the mandible contour, the performancecan go up to 98% of good detection. The complete workflow is implemented in the MAELab framework, freelyavailable as library on GitHub.

Comparing two random vectors by calculating a distance measure between the underlying probability densityfunctions is a key ingredient in many applications, especially in the domain of image processing. For this purpose,the recently introduced generalized Cramér-von Mises distance is an interesting choice, since it is well definedeven for the multivariate and discrete case. Unfortunately, the naive way of computing this distance, e.g., for twodiscrete two-dimensional random vectors ˜x; ˜y 2 [0; : : : ;n􀀀1]2;n 2 N has a computational complexity of O(n5) thatis impractical for most applications. This paper introduces fastGCVM, an algorithm that makes use of the wellknown concept of summed area tables and that allows to compute the generalized Cramér-von Mises distance witha computational complexity of O(n3) for the mentioned case. Two experiments demonstrate the achievable speedup and give an example for a practical application employing fastGCVM.

In this paper, we discuss the challenge of generating the screen-space ambient occlusion (SSAO) visual effect onthe state-of-the-art HoloVizio light field display. We detail the main features of modern SSAO techniques that arecurrently being applied during visualization on conventional 2D displays, and describe difficulties that potentiallycan appear when implementing this visual effect on 3D light field or similar systems. The main contribution of thispaper is our own modification of the SSAO algorithm for light field displays. The paper also includes suggestionson the possible ways to improve its visual quality and performance.

Vehicle localization is a fundamental issue in autonomous navigation that has been extensively studied by theRobotics community. An important paradigm for vehicle localization is based on visual place recognition whichrelies on learning a database, then consecutively trying to find matchings between this database and the actualvisual input. An increasing interest has been directed to visual place recognition in varying conditions like dayand night cycles and seasonal changes. A major approach dealing with such challenges is Sequence SLAM (SeqSLAM)based on matching a sequence of images to the database instead of a single image. This algorithm allowsglobal pose recovery at the expense of a higher computational time. To solve this problem with a certain amountof speedup, we propose in this work, a CUDA-based solution for real-time place recognition with SeqSLAM.We design a mapping of SeqSLAM to CUDA architecture and we describe, in detail, our hardware-specific implementationconsiderations as well as the parallelization methods. Performance analysis against existing CPUimplementation is also given, showing a speedup to six times faster than the CPU for common sized databases.More speedup could be obtained when dealing with bigger databases.

Estimating the volume of a 3D model of an object is an actual task in many scientific and engineering fields (forexample, CAD systems, biomedical engineering tasks etc.). Spline surfaces is one of the most powerful andflexible methods used to describe a 3D model. At the same time, it is rather difficult to estimate the volume of anobject described by spline surfaces. A model of a Bézier triangle is a simple type of a spline surface, but it ispractically advantageous.This paper describes a method of estimating the volume for 3D objects that are described by a set of Béziertriangles. The proposed method was tested on 3D models of objects of biomedical origin. A theorem is presentedin this paper for volume estimation, based on different properties of researched models, acquired by a projectionof a set vertices of a Bézier triangle onto a coordinate system axis. The proposed approach is based on usingmethods of differential geometry: surface integrals of scalar fields, Euler’s integral of the first kind and Betafunctions. Experimental results prove the accuracy of presented theorems. The proposed method can besuccessfully used to calculate the volume of different 3D models, including objects of biomedical origin.

In this study, a method for parameterizing unorganized cylindrical point clouds is suggested. The proposed methodcreates an initial base surface onto which points are projected to estimate parameter values for each point. To producean initial base surface, we suggest the concept of a virtual turntable and apply multilevel B-splines. Gridpoints are projected to the point cloud to increase the accuracy of the initial base surface. During the projectionprocess, a modified weight factor is introduced. Lastly, global B-spline surface interpolation is executed to generatean initial base surface, which undergoes a process of refinement and relaxation. Parameter values are thenobtained by projecting the point cloud onto the surface orthogonally. Experiments show that the proposed methodsuccessfully estimates parameters and solves the problems of self-loop and crossover. Furthermore, the results ofexperiment show that the proposed weight factor is more effective than the existing weight factor for point directedprojection.

With the recent advances in remote sensing of objects and environments, point cloud processing has become amajor field of study. Three-dimensional point cloud collected with remote sensing instruments may be very large,containing up to several tens of billions of points. This imposes the use for efficient and automatic algorithms toextract geometric or structural elements of the scanned surfaces. In this paper, we focus on the estimation of normaldirections in an unorganized point cloud and provide a curvature indicator. We avoid point-wise operations to acceleratethe running time for normals estimation. Instead, our method rely on an innovative anisotropic partitioningof the point cloud using an octree structure guided by the geometric complexity of the data and generates patchesof points. These patches are then approximated by a quadratic surface in order to estimate the normal directionsand curvatures. Our method has been applied to six models of various types presenting different characteristics andperforms, in average, 2.65 times faster than multi-threads implementations available in current pieces of software.The results obtained are a compromise between running time efficiency and normals accuracy. Moreover, thiswork opens up promising perspectives and can be easily inserted in wide range of workflows.

In this paper, we propose a novel approach to efficient rendering of an unlimited number of 3D objects in real-time.We present a rendering pipeline that is based on a new computer graphics programming paradigm implementinga holistic approach to the virtual scene definition. Using Signed Distance Functions (SDF) for a virtual scenerepresentation, we managed to control the content and complexity of the virtual scene with the use of mathematicalequations. In order to solve the limited hardware problem, especially the limited capacity of the GPU memory,we propose a scene element repository which extends the idea of the data based amplification. The content ofthe repository strongly depends on a 3D object visualization method. One of the most important requirementsof the developed pipeline is the possibility to render 3D objects created by artists. In order to achieve that, theobject visualization method uses Sparse Voxel Octree (SVO) ray casting. The developed rendering pipeline is fullycompatible with the available SVO algorithms. We show how to avoid occlusion errors which can occur in theSDF and SVO integration single-pass rendering pipeline. Finally, in order to control the content and complexityof the virtual scenes in an unlimited way, we propose a collection of global operators applicable to the virtualscene distance function. Developed Unlimited Object Instancing rendering pipeline can be easily integrated withtraditional visualization methods, e.g. the triangle rasterization. The only hardware requirement for our approachis the support for compute shaders or any GPGPU API.

Photosynthesis is one of turning points to shape the world. Plants use this process to convert light energy intochemical energy. Some of the early microorganisms evolved a way to use the energy from sunlight to makesugar out of simpler molecules, but unlike green plants today, the first photosynthesizing organisms did notrelease oxygen as waste product, so there was no oxygen in the air. Plants are very busy factories and leaves arethe main place for production. A useful plant recognition system is capable of identification of different speciesin natural environment. In natural environment, plants and leaves grow in different regions and climates. Duringday, variation of light intensity can be considered as an important factor. Thus, recognition of species in differentconditions is a real need as plants are ubiquitous in human life. A dataset of natural images has been utilized.The dataset contains four different plant species of Siegerland, Germany. Modern combined descriptionalgorithms, SURF, FAST-SURF, and HARRIS-SURF, have been carried out to implement a reliable system forplants species recognition and classification in natural environment. One of well known methods in machinelearning community, Support Vector Machine, has been applied in the implemented systems. All steps ofsystem’s implementation are described in related sections. The highest obtained accuracy belongs to theimplemented system by means of SURF algorithm and equals to 93.9575.

Vector representation of the images, maps, schematics and other information is widely used, and in computer processingof these data, comparison and similarity evaluation of two sets of line segments is often necessary. Varioustechniques are already in use, but these mostly rely on the algorithmic functions such as minimum/maximum oftwo or more variables, which limits their applicability for many optimization algorithms. In this paper we proposea novel area based criterion function for line segment similarity evaluation, which is easily differentiable and thederivatives are continuous in the whole domain of definition. The second important feature is the possibility ofpreprocessing of the input data. Once finished, it takes constant time to evaluate the criterion for different transformationsof one of the input sets of line segments. This has potential to greatly speed up iterative matchingalgorithms. In such case, the computational complexity is reduced from O(pt) to O(p+t), where p is the numberof line segment pairs being examined and t is the number of transformations performed.

Conventional UAV (abbr. Unmanned Air Vehicle) auto-pilot systems uses GPS signal for navigation. While theGPS signal is lost, jammed or the UAV is navigating in GPS-denied environment conventional autopilot systemsfail to navigate safely. UAV should estimate it’s own position without the need of external signals. Localization,the process of pose estimation relatively to known environment, may solve the problem of navigation without GPSsignal. Downward looking camera on a UAV may be used to solve pose estimation problem in combination withvisual odometry and other sensor data. In this paper a vision-based particle filter application is proposed to solveGPS-denied UAV localization. The application uses visual odometry for motion estimation, correlation coefficientfor apriori known map image matching with aerial imagery, KLD (abbr. Kueller-Leiblach distance) sampling forparticle filtering. Research using data collected during real UAV flight is performed to investigate: UAV headinginfluence on correlation coefficient values when matching aerial imagery with the map and measure localizationaccuracy compared to conventional GPS system and state-of-the-art odometry.

Estimation of design rainfall in unobservable places is important in hydrological engineering. The aim of thispaper is to use genetic algorithms to find the optimal global and local shape parameters of radial basis functions(RBFs) to create an interpolation model to estimate scaling exponents of short term rainfalls across selected regionsof Slovakia. Scaling exponents can be used later to estimate rainfalls intensity in places without observations. Inthis paper, we have used interpolation methods based on RBFs to model interpolation surfaces. We investigate theproperties of shape parameters in RBFs, and we test some methods for finding an optimal shape parameter. Thechoice of the best basis function along with the optimal shape parameter has a significant impact on the accuracyof the interpolation models which best approximate the real model. We have found that Hardy’s multiquadricsinterpolant with the optimal local shape parameters can be used for estimation the rainfall intensities in areaswithout direct observation.

In recent years, reinforced concrete has been widely used as a building material having high strength. However,when the arrangement of the reinforcing bars embedded in the inside are not correct, there is a problem that thestrength drops greatly. Therefore, when use reinforced concrete, it is necessary to confirm whether the reinforcingbars are correctly arranged. In this paper, we propose a method of thinning of point clouds scanned by rangescanner from reinforcing bars, and extracting the core wires of point clouds by introducing a concentrationdistribution function. This process moves points according to a gradient field of a concentration field. In addition,we conducted an experiment to confirm the effectiveness of the proposed method. As a result of the experiment,it was confirmed that the core wires can be extracted from the point clouds of the reinforcing bars.

Human action recognition is a topic widely studied over time, using numerous techniques and methods to solvea fundamental problem in automatic video analysis. Basically, a traditional human action recognition systemcollects video frames of human activities, extracts the desired features of each human skeleton and classify themto distinguish human gesture. However, almost all of these approaches roll out the space-time information of therecognition process. In this paper we present a novel use of an existing state-of-the-art space-time technique, theSpace-Time Interest Point (STIP) detector and its velocity adaptation, to human action recognition process. UsingSTIPs as descriptors and a Support Vector Machine classifier, we evaluate four different public video datasets tovalidate our methodology and demonstrate its accuracy in real scenarios.

Modelling dynamic pose deformations of human subjects is an important topic in many research applications.Existing approaches of human pose deformations can be classified as volume-based, skeletal animation andexample-based methods. These approaches have both strengths and limitations. However, for models incustomized shapes, it is very challenging to deform these models into different poses rapidly and realistically. We10 propose a conceptual model to realize rapid and realistic pose deformation to customized human models by theintegration of skeletal-driven rigid deformation and example-learnt non-rigid surface deformation. Based on thisframework, a method for rapid automatic pose deformation is developed to deform human models of variousbody shapes into a series of dynamic poses. A series of algorithms are proposed to complete the pose deformationautomatically and efficiently, including automatic segmentation of body parts and skeleton embedding, skeletal15driven rigid deformation, training of non-rigid deformation from pose dataset; shape mapping of non-rigiddeformation, and integration of rigid and non-rigid deformations. Experiment has shown that the proposedmethod can customize accurate human models based on two orthogonal-view photos and also efficiently generaterealistic pose deformations for the customized models.