As a solution for the lost-in-space star identification problem we present Star Identification using Graph Neural Network (SIGNN), a novel approach using Graph Attention Networks. By representing the celestial sphere as a graph data structure, created from the ESA's Hipparcos catalogue, we are able to accurately capture the rich information and relationships within local star fields. Graph learning techniques allow our model to aggregate information and learn the relative importance of the nodes and structure within each stars local neighbourhood to it's identification. This approach, combined with our parametric data-generation and noise simulation, allows us to train a highly robust model capable of accurate star identification even under intensive noise, outperforming existing methods. Code and generation techniques will be available on https://github.com/FloydHepburn/SIGNN.

Point cloud completion is crucial for 3D computer vision tasks in autonomous driving, augmented reality, and robotics. However, obtaining clean and complete point clouds from real-world environments is challenging due to noise and occlusions. Consequently, most existing completion networks – trained on synthetic data – struggle with real-world degradations. In this work, we tackle the problem of completing and denoising highly corrupted partial point clouds affected by multiple simultaneous degradations. To benchmark robustness, we introduce the Corrupted Point Cloud Completion Dataset (CPCCD), which highlights the limitations of current methods under diverse corruptions. Building on these insights, we propose DWCNet (Denoising-While-Completing Network), a completion framework enhanced with a Noise Management Module (NMM) that leverages contrastive learning and self-attention to suppress noise and model structural relationships. DWCNet achieves state-of-the-art performance on both clean and corrupted, synthetic and real-world datasets. The dataset and code will be publicly available at https://github.com/keneniwt/DWCNET-Robust-Point-Cloud-Completion-against-Corruptions .

Point cloud completion is the task of producing a complete 3D shape given an input of a partial point cloud. It has become a vital process in 3D computer graphics, vision and applications such as autonomous driving, robotics, and augmented reality. These applications often rely on the presence of a complete 3D representation of the environment. Over the past few years, many completion algorithms have been proposed and a substantial amount of research has been carried out. However, there are not many in-depth surveys that summarise the research progress in such a way that allows users to make an informed choice of what algorithms to employ given the type of data they have, the end result they want, the challenges they may face and the possible strategies they could use. In this study, we present a comprehensive survey and classification of articles on point cloud completion untill August 2023 based on the strategies, techniques, inputs, outputs, and network architectures. We will also cover datasets, evaluation methods, and application areas in point cloud completion. Finally, we discuss challenges faced by the research community and future research directions.

The objective of this work was to investigate from a user perspective linkage between a 1D time-series view of data and a 2D representation provided by dimension reduction techniques. Our hypothesis is that when such interaction happens seamlessly, the use of these linked views, compared to only interacting with the 1D time-series view, for the ubiquitous task of selection and labelling, is more efficient and effective both in terms of performance and user experience. To this end we examine different dimension reduction techniques (UMAP, t-SNE, PCA and Autoencoder) and evaluate each technique within our experimental setting. Results demonstrate that there is a positive impact on speed and accuracy through augmenting 1D views with a dimension reduction 2D view when these views are linked and linkage is supported through coordinated interaction.

Data compression forms a central role in handling the bottleneck of data storage, transmission and processing. Lossless compression requires reducing the file size whilst maintaining bit-perfect decompression, which is the main target in medical applications. This paper presents a novel lossless compression method for 16-bit medical imaging volumes. The aim is to train a neural network (NN) as a 3D data predictor, which minimizes the differences with the original data values and to compress those residuals using arithmetic coding. We evaluate the compression performance of our proposed models to state-of-the-art lossless compression methods, which shows that our approach accomplishes a higher compression ratio in comparison to JPEG-LS, JPEG2000, JP3D, and HEVC and generalizes well.

The achievement of neural networks in a variety of applications is accompanied by a dramatic increase in computational costs and memory requirements. In this paper, we propose an efficient method to simultaneously identify the critical neurons and prune the model during training without involving any pre-training or fine-tuning procedures. Unlike existing methods, which accomplish this task in a greedy fashion, we propose a majority voting technique to compare the activation values among neurons and assign a voting score to quantitatively evaluate their importance. This mechanism helps to effectively reduce model complexity by eliminating the less influential neurons and aims to determine a subset of the whole model that can represent the reference model with much fewer parameters within the training process. Experimental results show that majority voting efficiently compresses the network with no drop in model accuracy, pruning more than 79% of the original model parameters on CIFAR10 and more than 91% of the original parameters on MNIST. Moreover, we show that with our proposed method, sparse models can be further pruned into even smaller models by removing more than 60% of the parameters, whilst preserving the reference model accuracy.

As scanners produce higher-resolution and more densely sampled images, this raises the challenge of data storage, transmission and communication within healthcare systems. Since the quality of medical images plays a crucial role in diagnosis accuracy, medical imaging compression techniques are desired to reduce scan bitrate while guaranteeing lossless reconstruction. This paper presents a lossless compression method that integrates a Recurrent Neural Network (RNN) as a 3D sequence prediction model. The aim is to learn the long dependencies of the voxel's neighbourhood in 3D using Long Short-Term Memory (LSTM) network then compress the residual error using arithmetic coding. Experiential results reveal that our method obtains a higher compression ratio achieving 15% saving compared to the state-of-the-art lossless compression standards, including JPEG-LS, JPEG2000, JP3D, HEVC, and PPMd. Our evaluation demonstrates that the proposed method generalizes well to unseen modalities CT and MRI for the lossless compression scheme. To the best of our knowledge, this is the first lossless compression method that uses LSTM neural network for 16-bit volumetric medical image compression.

We seek to answer the question on whether different geometrical attributes within a glyph can bias interpretation of data. We focus on a specific visual encoding, the Emoji, and evaluate its effectiveness at encoding multidimensional features. Given the anthropomorphic nature of the encoding we seek to quantify the amount of bias the encoding itself introduces, and use this to balance the Emoji glyph to remove that bias. We perform our analysis by comparing Emoji with Chernoff faces, of which they can be seen as direct descendant. Results shed light on how this new approach of feature-tuning in glyph design can influence overall effectiveness of novel multidimensional encodings.

A fully-parallelized work-time optimal algorithm is presented for computing the exact Euclidean Distance Transform (EDT) of a 2D binary image with the size of n × n. Unlike existing PRAM (Parallel Random Access Machine) and other algorithms, this algorithm is suitable for implementation on modern SIMD (Single Instruction Multiple Data) architectures such as GPUs. As a fundamental operation of 2D EDT, 1D EDT is efficiently parallelized first. Specifically, the GPU algorithm for the 1D EDT, which uses CUDA (Compute Unified Device Architecture) binary functions, such as ballotO, ffs(), clzO, and shflO, runs in O(log32n) time and performs O(n) work. Using the 1D EDT as a fundamental operation, the fully-parallelized work-time optimal 2D EDT algorithm is designed. This algorithm consists of three steps. Step 1 of the algorithm runs in O(log32n) time and performs O(N) (N=n2) of total work on GPU. Step 2 performs O(N) of total work and has an expected time complexity of O(logn) on GPU. Step 3 runs in O(log32n) time and performs O(N) of total work on GPU. As far as we know, this algorithm is the first fully-parallelized and realized work-time optimal algorithm for GPUs. The experimental results show that this algorithm outperforms the prior state-of-the-art GPU algorithms.

There is a need for solutions which assist users to understand long time-series data by observing its changes over time, finding repeated patterns, detecting outliers, and effectively labeling data instances. Although these tasks are quite distinct and are usually tackled separately, we present an interactive visual analytics system and approach that can address these issues in a single system. It enables users to visualize, understand and explore univariate or multivariate long time-series data in one image using a connected scatter plot. It supports interactive analysis and exploration for pattern discovery and outlier detection. Different dimensionality reduction techniques are used and compared in our system. Because of its power of extracting features, deep learning is used for multivariate time-series along with 2D reduction techniques for rapid and easy interpretation and interaction with large amount of time-series data. We deploy our system with different time-series datasets and report two real-world case studies that are used to evaluate our system.

In this paper, we propose a clustering approach embedded in a deep convolutional auto-encoder (DCAE). In contrast to conventional clustering approaches, our method simultaneously learns feature representations and cluster assignments through DCAEs. DCAEs have been effective in image processing as it fully utilizes the properties of convolutional neural networks. Our method consists of clustering and reconstruction objective functions. All data points are assigned to their new corresponding cluster centers during the optimization, after that, clustering centers are iteratively updated to obtain a stable performance of clustering. The experimental results on the MNIST dataset show that the proposed method substantially outperforms deep clustering models in term of clustering quality.

Evaluating image quality in Monte Carlo rendered images is an important aspect of the rendering process as we often need to determine the relative quality between images computed using different algorithms and with varying amounts of computation. The use of a gold-standard, reference image, or ground truth is a common method to provide a baseline with which to compare experimental results. We show that if not chosen carefully, the quality of reference images used for image quality assessment can skew results leading to significant misreporting of error. We present an analysis of error in Monte Carlo rendered images and discuss practices to avoid or be aware of when designing an experiment.

Collecting sensor data results in large temporal data sets which need to be visualized, analyzed, and presented. Onedimensional time-series charts are used, but these present problems when screen resolution is small in comparison to the data. This can result in severe over-plotting, giving rise for the requirement to provide effective rendering and methods to allow interaction with the detailed data. Common solutions can be categorized as multi-scale representations, frequency based, and lens based interaction techniques. In this paper, we comparatively evaluate existing methods, such as Stack Zoom [15] and ChronoLenses [38], giving a graphical overview of each and classifying their ability to explore and interact with data. We propose new visualizations and other extensions to the existing approaches. We undertake and report an empirical study and a field study using these techniques.

We present a visualization tool aimed specifically at the development and optimization of photon map denoising methods. Our tool allows the rapid testing of hypotheses and algorithms through the use of parallel coordinates, domain‐specific scripting, colour mapping and point plots. Interaction is carried out by brushing, adjusting parameters and focus‐plus‐context and yields interactive visual feedback and debugging information. We demonstrate the use of the tool to explore high‐dimensional photon map data, facilitating the discovery of novel parameter spaces which can be used to dissociate complex caustic illumination. We then show how these new parametrizations may be used to improve upon pre‐existing noise removal methods in the context of the photon relaxation framework.

Hyperspectral image visualization reduces color bands to three, but prevailing linear methods fail to address data characteristics, and nonlinear embeddings are computationally demanding. Qualitative evaluation of embedding is also lacking. We propose faithful stochastic proximity embedding (FSPE), which is a scalable and nonlinear dimensionality reduction method. FSPE considers the nonlinear characteristics of spectral signatures, yet it avoids the costly computation of geodesic distances that are often required by other nonlinear methods. Furthermore, we employ a pixelwise metric that measures the quality of hyperspectral image visualization at each pixel. FSPE outperforms the state-of-art methods by at least 12% on average and up to 25% in the qualitative measure. An implementation on graphics processing units is two orders of magnitude faster than the baseline. Our method opens the path to high-fidelity and real-time analysis of hyperspectral images.

In this paper we introduce Order of Magnitude Markers (OOMMs) as a new technique for number representation. The motivation for this work is that many data sets require the depiction and comparison of numbers that have varying orders of magnitude. Existing techniques for representation use bar charts, plots and colour on linear or logarithmic scales. These all suffer from related problems. There is a limit to the dynamic range available for plotting numbers, and so the required dynamic range of the plot can exceed that of the depiction method. When that occurs, resolving, comparing and relating values across the display becomes problematical or even impossible for the user. With this in mind, we present an empirical study in which we compare logarithmic, linear, scale-stack bars and our new markers for 11 different stimuli grouped into 4 different tasks across all 8 marker types.

Abstract Stream compaction is an important parallel computing primitive that produces a reduced (compacted) output stream consisting of only valid elements from an input stream containing both invalid and valid elements. Computing on this compacted stream rather than the mixed input stream leads to improvements in performance, load balancing and memory footprint. Stream compaction has numerous applications in a wide range of domains: e.g. deferred shading, isosurface extraction and surface voxelization in computer graphics and visualization. We present a novel In‐Kernel stream compaction method, where compaction is completed before leaving an operating kernel. This contrasts with conventional parallel compaction methods that require leaving the kernel and running a prefix sum kernel followed by a scatter kernel. We apply our compaction methods to ray‐tracing‐based visualization of volumetric data. We demonstrate that the proposed In‐Kernel compaction outperforms the standard out‐of‐kernel Thrust parallel‐scan method for performing stream compaction in this real‐world application. For the data visualization, we also propose a novel multi‐kernel ray‐tracing pipeline for increased thread coherency and show that it outperforms a conventional single‐kernel approach.

This paper presents a novel approach to detecting and preserving fine illumination structure within photon maps. Data derived from each photon's primal trajectory is encoded and used to build a high-dimensional kd-tree. Incorporation of these new parameters allows for precise differentiation between intersecting ray envelopes, thus minimizing detail degradation when combined with photon relaxation. We demonstrate how parameter-aware querying is beneficial in both detecting and removing noise. We also propose a more robust structure descriptor based on principal components analysis that better identifies anisotropic detail at the sub-kernel level. We illustrate the effectiveness of our approach in several example scenes and show significant improvements when rendering complex caustics compared to previous methods.

We introduce a novel algorithm for progressively removing noise from view-independent photon maps while simultaneously minimizing residual bias. Our method refines a primal set of photons using data from multiple successive passes to estimate the incident flux local to each photon. We show how this information can be used to guide a relaxation step with the goal of enforcing a constant, per-photon flux. Using a reformulation of the radiance estimate, we demonstrate how the resulting blue noise photon distribution yields a radiance reconstruction in which error is significantly reduced. Our approach has an open-ended runtime of the same order as unbiased and asymptotically consistent rendering methods, converging over time to a stable result. We demonstrate its effectiveness at storing caustic illumination within a view-independent framework and at a fidelity visually comparable to reference images rendered using progressive photon mapping.

Traditional sketch-based image or video search systems rely on machine learning concepts as their core technology. However, in many applications, machine learning alone is impractical since videos may not be semantically annotated sufficiently, there may be a lack of suitable training data, and the search requirements of the user may frequently change for different tasks. In this work, we develop a visual analytics systems that overcomes the shortcomings of the traditional approach. We make use of a sketch-based interface to enable users to specify search requirement in a flexible manner without depending on semantic annotation. We employ active machine learning to train different analytical models for different types of search requirements. We use visualization to facilitate knowledge discovery at the different stages of visual analytics. This includes visualizing the parameter space of the trained model, visualizing the search space to support interactive browsing, visualizing candidature search results to support rapid interaction for active learning while minimizing watching videos, and visualizing aggregated information of the search results. We demonstrate the system for searching spatiotemporal attributes from sports video to identify key instances of the team and player performance.

Streamline seeding rakes are widely used in vector field visualization. We present new approaches for calculating similarity between integral curves (streamlines and pathlines). While others have used similarity distance measures, the computational expense involved with existing techniques is relatively high due to the vast number of euclidean distance tests, restricting interactivity and their use for streamline seeding rakes. We introduce the novel idea of computing streamline signatures based on a set of curve-based attributes. A signature produces a compact representation for describing a streamline. Similarity comparisons are performed by using a popular statistical measure on the derived signatures. We demonstrate that this novel scheme, including a hierarchical variant, produces good clustering results and is computed over two orders of magnitude faster than previous methods. Similarity-based clustering enables filtering of the streamlines to provide a nonuniform seeding distribution along the seeding object. We show that this method preserves the overall flow behavior while using only a small subset of the original streamline set. We apply focus + context rendering using the clusters which allows for faster and easier analysis in cases of high visual complexity and occlusion. The method provides a high level of interactivity and allows the user to easily fine tune the clustering results at runtime while avoiding any time-consuming recomputation. Our method maintains interactive rates even when hundreds of streamlines are used.

Abstract Noise removal for Monte Carlo global illumination rendering is a well known problem, and has seen significant attention from image-based filtering methods. However, many state of the art methods breakdown in the presence of high frequency features, complex lighting and materials. In this work we present a probabilistic image based noise removal and irradiance filtering framework that preserves this high frequency detail such as hard shadows and glossy reflections, and imposes no restrictions on the characteristics of the light transport or materials. We maintain per-pixel clusters of the path traced samples and, using statistics from these clusters, derive an illumination aware filtering scheme based on the discrete Poisson probability distribution. Furthermore, we filter the incident radiance of the samples, allowing us to preserve and filter across high frequency and complex textures without limiting the effectiveness of the filter.

Abstract Today real‐time sports performance analysis is a crucial aspect of matches in many major sports. For example, in soccer and rugby, team analysts may annotate videos during the matches by tagging specific actions and events, which typically result in some summary statistics and a large spreadsheet of recorded actions and events. To a coach, the summary statistics (e.g., the percentage of ball possession) lacks sufficient details, while reading the spreadsheet is time‐consuming and making decisions based on the spreadsheet in real‐time is thereby impossible. In this paper, we present a visualization solution to the current problem in real‐time sports performance analysis. We adopt a glyph‐based visual design to enable coaching staff and analysts to visualize actions and events “at a glance”. We discuss the relative merits of metaphoric glyphs in comparison with other types of glyph designs in this particular application. We describe an algorithm for managing the glyph layout at different spatial scales in interactive visualization. We demonstrate the use of this technical approach through its application in rugby, for which we delivered the visualization software,MatchPad, on a tablet computer. The MatchPad was used by the Welsh Rugby Union during the Rugby World Cup 2011. It successfully helped coaching staff and team analysts to examine actions and events in detail whilst maintaining a clear overview of the match, and assisted in their decision making during the matches. It also allows coaches to convey crucial information back to the players in a visually‐engaging manner to help improve their performance.

Abstract A new area of biological research is identifying and grouping patterns of behaviour in wild animals by analysing data obtained through the attachment of tri‐axial accelerometers. As these recording devices become smaller and less expensive their use has increased. Currently acceleration data are visualised as 2D time series plots, and analyses are based on summary statistics and the application of Fourier transforms. We develop alternate visualisations of this data so as to analyse, explore and present new patterns of animal behaviour. Our visualisations include interactive spherical scatterplots, spherical histograms, clustering methods, and feature‐based state diagrams of the data. We study the application of these visualisation methods to accelerometry data from animal movement. The reaction of biologists to these visualisations is also reported.

Abstract The photon mapping method is one of the most popular algorithms employed in computer graphics today. However, obtaining good results is dependent on several variables including kernel shape and bandwidth, as well as the properties of the initial photon distribution. While the photon density estimation problem has been the target of extensive research, most algorithms focus on new methods of optimising the kernel to minimise noise and bias. In this paper we break from convention and propose a new approach that directly redistributes the underlying photons. We show that by relaxing the initial distribution into one with a blue noise spectral signature we can dramatically reduce background noise, particularly in areas of uniform illumination. In addition, we propose an efficient heuristic to detect and preserve features and discontinuities. We then go on to demonstrate how reconfiguration also permits the use of very low bandwidth kernels, greatly improving render times whilst reducing bias.

In this paper, we present a new visual way of exploring state sequences in large observational time-series. A key advantage of our method is that it can directly visualize higher-order state transitions. A standard first order state transition is a sequence of two states that are linked by a transition. A higher-order state transition is a sequence of three or more states where the sequence of participating states are linked together by consecutive first order state transitions. Our method extends the current state-graph exploration methods by employing a two dimensional graph, in which higher-order state transitions are visualized as curved lines. All transitions are bundled into thick splines, so that the thickness of an edge represents the frequency of instances. The bundling between two states takes into account the state transitions before and after the transition. This is done in such a way that it forms a continuous representation in which any subsequence of the time series is represented by a continuous smooth line. The edge bundles in these graphs can be explored interactively through our incremental selection algorithm.We demonstrate our method with an application in exploring labeled time-series data from a biological survey, where a clustering has assigned a single label to the data at each time-point. In these sequences, a large number of cyclic patterns occur, which in turn are linked to specific activities. We demonstrate how our method helps to find these cycles, and how the interactive selection process helps to find and investigate activities.

Photon mapping is an efficient method for producing high-quality, photorealistic images with full global illumination. In this paper we present a more accurate and efficient approach to final gathering using the photon map based upon hierarchical evaluation of the photons over each surface. We use the footprint of each gather ray to calculate the irradiance estimate area rather than deriving it from the local photon density. We then describe an efficient method for computing the irradiance from the photon map given an arbitrary estimate area. Finally, we demonstrate how the technique may be used to reduce variance and increase efficiency when sampling diffuse and glossy-specular BRDFs.

Abstract A sampled object representation (SOR) defines a graphical model using data obtained from a sampling process, which takes a collection of samples at discrete positions in space in order to capture certain geometrical and physical properties of one or more objects of interest. Examples of SORs include images, videos, volume datasets and point datasets. Unlike many commonly used data representations in computer graphics, SORs lack in geometrical, topological and semantic information, which is much needed for controlling deformation and animation. Hence it poses a significant scientific and technical challenge to develop deformation and animation methods that operate upon SORs. Such methods can enable computer graphics and computer animation to benefit enormously from the advances of digital imaging technology. In this state of the art report, we survey a wide range of techniques that have been developed for manipulating, deforming and animating SORs. We consider a collection of elementary operations for manipulating SORs, which can serve as building blocks of deformation and animation techniques. We examine a collection of techniques that are designed to transform the geometry shape of deformable objects in sampled representations and pay particular attention to their deployment in surgical simulation. We review a collection of techniques for animating digital characters in SORs, focusing on recent developments in volume animation.

A distance field is a representation where, at each point within the field, we know the distance from that point to the closest point on any object within the domain. In addition to distance, other properties may be derived from the distance field, such as the direction to the surface, and when the distance field is signed, we may also determine if the point is internal or external to objects within the domain. The distance field has been found to be a useful construction within the areas of computer vision, physics, and computer graphics. This paper serves as an exposition of methods for the production of distance fields, and a review of alternative representations and applications of distance fields. In the course of this paper, we present various methods from all three of the above areas, and we answer pertinent questions such as How accurate are these methods compared to each other? How simple are they to implement?, and What is the complexity and runtime of such methods?

Abstract If we were to have a Grid infrastructure for visualization, what technologies would be needed to build such an infrastructure, what kind of applications would benefit from it, and what challenges are we facing in order to accomplish this goal? In this survey paper, we make use of the term ‘visual supercomputing’ to encapsulate a subject domain concerning the infrastructural technology for visualization. We consider a broad range of scientific and technological advances in computer graphics and visualization, which are relevant to visual supercomputing. We identify the state‐of‐the‐art technologies that have prepared us for building such an infrastructure. We examine a collection of applications that would benefit enormously from such an infrastructure, and discuss their technical requirements. We propose a set of challenges that may guide our strategic efforts in the coming years.

This paper presents the design and implementation of artistic effects in modelling and rendering of volume datasets. Following different stages of a volume-based graphics pipeline, we examine various properties of volume data, and illustrate how expressive and non-photorealistic effects can be implemented. We demonstrate that the true 3D nature of volume data makes it particularly applicable for this use, allowing the addition of complex effects at the modelling stage as well as during rendering.

Voxelisation is the process of converting a source object of any data type into a three-dimensional grid of voxel values. This voxel grid should represent the original object as closely as possible, although some inaccuracies will occur due to the discrete nature of the voxel grid representation. In this paper we report our ongoing research into methods for representing objects as voxelised distance fields, in particular we report fast methods for accurate distance field production. A review of current alternative voxelisation methods is also given.

Abstract This paper presents a new approach to the construction of a surface from a stack of contour slices. Unlike most existing methods, this new approach handles ambiguous conditions consistently without employing an algorithm to establish a correspondence between vertices on one contour and those on the next. It is easy to implement and fast to compute, requiring only basic geometric properties, namely closedness and simplicity, to be available with contour data. The advantages of this new approach have also been demonstrated with solutions to a few classical problems from the literature and some practical problems in medical imaging. It can also be applied to geographical surveying and keyframe animations.