http://localhost:1313/publication/PublicationsHugo Blox Builder (https://hugoblox.com)en-usMon, 01 Jun 2026 00:00:00 +0000http://localhost:1313/media/icon_hu10226378635250344412.pnghttp://localhost:1313/publication/http://localhost:1313/arcol-aspect-ratio-constrained-orthogonal-layout/Mon, 01 Jun 2026 00:00:00 +0000http://localhost:1313/arcol-aspect-ratio-constrained-orthogonal-layout/<h2 id="motivation">Motivation</h2> <p>Orthogonal layouts make graphs easier to read in domains such as software, process, and biological-network visualization. Existing algorithms optimize local layout quality, but offer little control over the overall shape—a problem when the same graph must fit a wide display, a tall screen, or a fixed interface region.</p> <h2 id="method">Method</h2> <p>ARCOL extends the Human-like Orthogonal Layout Algorithm (HOLA) with a soft aspect-ratio objective. It introduces aspect-ratio awareness during core stress minimization and the subsequent placement of tree components, encouraging a requested shape while retaining the geometric constraints and readability conventions of orthogonal drawing.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper compares layouts across target aspect ratios and studies the resulting trade-offs in area, bends, crossings, and edge length. The evaluation shows that ARCOL can steer the global layout shape without treating aspect ratio as a purely post-processing operation.</p> <h2 id="limitations">Limitations</h2> <p>Aspect ratio is an additional layout objective, not a guarantee that every graph can meet an arbitrary target without compromises. Highly constrained graphs may still require trade-offs against compactness or other conventional quality measures.</p>http://localhost:1313/efficientmonohair-fast-strand-level-reconstruction-from-monocular-video-via-multi-view-direction-fusion/Mon, 01 Jun 2026 00:00:00 +0000http://localhost:1313/efficientmonohair-fast-strand-level-reconstruction-from-monocular-video-via-multi-view-direction-fusion/<h2 id="motivation">Motivation</h2> <p>Strand-level hair reconstruction must balance global shape, fine directional detail, and runtime. Implicit methods can capture a coherent hair mass but lose individual strands, while explicit optimization can be accurate but expensive and difficult to scale.</p> <h2 id="method">Method</h2> <p>EfficientMonoHair combines an implicit neural representation with multi-view geometric fusion. It samples hair directions across views, aggregates them with a fusion-patch optimization, and enforces patch-level consistency to create a direction-aware outer-layer point cloud from which strands can be reconstructed.</p> <h2 id="evaluation">Evaluation</h2> <p>The work compares reconstruction quality and efficiency with existing monocular-video approaches across varied hairstyles. Its evaluation emphasizes the benefit of using multi-view direction evidence without the sequential optimization burden of earlier strand-growing pipelines.</p> <h2 id="limitations">Limitations</h2> <p>Very complex, heavily occluded, or highly curled hairstyles remain difficult because visible image evidence can be sparse or ambiguous. As with other monocular methods, reconstruction quality is coupled to capture coverage and input-image quality.</p>http://localhost:1313/clickaixr-on-device-multimodal-vision-language-interaction-with-real-world-objects-in-extended-reality/Wed, 01 Apr 2026 00:00:00 +0000http://localhost:1313/clickaixr-on-device-multimodal-vision-language-interaction-with-real-world-objects-in-extended-reality/<h2 id="motivation">Motivation</h2> <p>Vision-language assistants can answer questions about the physical world, but cloud inference introduces privacy, connectivity, and latency concerns. In XR, they must also know precisely which nearby object the user intends to ask about.</p> <h2 id="method">Method</h2> <p>ClickAIXR combines local vision-language inference with a controller-driven object-selection interface on an XR headset. Users place and adjust a 3D crop around a physical object using depth, width, and height controls; the selected visual context is then passed to an on-device VLM with the spoken or typed question.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper evaluates the interaction on real-world objects, considering the accuracy of the selected context as well as the practical responsiveness of the on-device workflow. The system is positioned against cloud-based and gaze-based alternatives that leave privacy or referential ambiguity unresolved.</p> <h2 id="limitations">Limitations</h2> <p>The approach depends on the capabilities and compute budget of the headset-resident model, and careful selection remains necessary when objects are occluded, small, or spatially close together.</p>http://localhost:1313/rara-clipper-a-clipper-for-gaussian-splatting-based-on-ray-tracer-and-rasterizer/Mon, 01 Dec 2025 00:00:00 +0000http://localhost:1313/rara-clipper-a-clipper-for-gaussian-splatting-based-on-ray-tracer-and-rasterizer/<h2 id="motivation">Motivation</h2> <p>Clipping is essential for inspecting interior structure in 3D scenes. For Gaussian Splatting, however, a primitive can contribute to many pixels across a volume, so simply discarding primitives at a clipping plane can produce imprecise boundaries and distracting artifacts.</p> <h2 id="method">Method</h2> <p>RaRa Clipper uses a hybrid renderer: rasterization provides efficient visibility handling, while ray tracing resolves the volumetric contributions needed for accurate clipping. The resulting pipeline is designed to preserve both the speed expected from splatting and the spatial fidelity needed at a cut boundary.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper compares RaRa Clipper with hard clipping across a set of clipping scenarios, combining image-based measures with participant feedback. The evaluation focuses on the visual clarity of the cut, preservation of surrounding context, and dataset-dependent preferences.</p> <h2 id="limitations">Limitations</h2> <p>Clipping quality remains dependent on the underlying Gaussian representation and viewing configuration. The work also identifies cases where retaining contextual background or choosing the right clipping configuration requires further user control.</p>http://localhost:1313/s2d-sparse-to-dense-keymask-distillation-for-unsupervised-video-instance-segmentation/Mon, 01 Dec 2025 00:00:00 +0000http://localhost:1313/s2d-sparse-to-dense-keymask-distillation-for-unsupervised-video-instance-segmentation/http://localhost:1313/a-survey-on-quality-metrics-for-text-to-image-generation/Wed, 01 Oct 2025 00:00:00 +0000http://localhost:1313/a-survey-on-quality-metrics-for-text-to-image-generation/<h2 id="motivation">Motivation</h2> <p>Text-to-image systems must be judged on more than photorealism: an image can look plausible while failing to depict the requested objects, relationships, or style. Classical image-quality measures alone do not capture this joint visual-and-linguistic requirement.</p> <h2 id="method">Method</h2> <p>This survey organizes text-to-image evaluation metrics around two main criteria: compositional image quality and semantic consistency with the prompt. It reviews metric families, the datasets used to validate them, and how evaluation choices interact with the intended generation task.</p> <h2 id="evaluation">Evaluation</h2> <p>Beyond cataloguing metrics, the paper discusses the evidence behind them, surveys available evaluation datasets, and examines a selection of human-preference metrics in supplementary experiments. The result is a practical map of what a reported score can—and cannot—support.</p> <h2 id="limitations">Limitations</h2> <p>No single automatic metric captures all of human judgment. Metric selection remains task-dependent, and the survey emphasizes the need to combine appropriate automatic measures with carefully designed human evaluation when stakes are high.</p>http://localhost:1313/cuts3d-cutting-semantics-in-3d-for-2d-unsupervised-instance-segmentation/Wed, 01 Oct 2025 00:00:00 +0000http://localhost:1313/cuts3d-cutting-semantics-in-3d-for-2d-unsupervised-instance-segmentation/http://localhost:1313/attention-guided-masked-autoencoders-for-learning-image-representations/Sat, 01 Feb 2025 00:00:00 +0000http://localhost:1313/attention-guided-masked-autoencoders-for-learning-image-representations/http://localhost:1313/hpscan-human-perception-based-scattered-data-clustering/Sat, 01 Feb 2025 00:00:00 +0000http://localhost:1313/hpscan-human-perception-based-scattered-data-clustering/<h2 id="motivation">Motivation</h2> <p>Conventional clustering algorithms optimize mathematical criteria that do not necessarily match how people perceive groups in a scatterplot. HPSCAN asks a different question: which separations would human viewers recognize as clusters?</p> <h2 id="method">Method</h2> <p>The authors crowdsourced cluster annotations for 7,320 bivariate datasets from 384 participants. HPSCAN learns from these labels using a modified PointNet++ architecture that receives the scattered point coordinates directly, avoiding the resolution and rendering choices of image-based scatterplot analysis.</p> <h2 id="evaluation">Evaluation</h2> <p>The work reports annotation statistics and compares HPSCAN’s separations with human judgments on held-out and real-world data. This makes the target of evaluation perceptual agreement, rather than only agreement with a geometric clustering heuristic.</p> <h2 id="limitations">Limitations</h2> <p>The model is trained on bivariate scatterplots and reflects the judgments represented in the collected dataset. Extending this notion of perceptual clustering to other visual encodings or higher-dimensional interaction contexts remains an open direction.</p>http://localhost:1313/depthg/Tue, 27 Feb 2024 00:00:00 +0000http://localhost:1313/depthg/http://localhost:1313/deep-learning-in-volume-rendering/Mon, 01 Jan 2024 00:00:00 +0000http://localhost:1313/deep-learning-in-volume-rendering/http://localhost:1313/vit-tf/Sun, 03 Sep 2023 00:00:00 +0000http://localhost:1313/vit-tf/<h2 id="motivation">Motivation</h2> <p>Transfer functions determine which structures are visible in a volume rendering, yet designing them from intensity and gradient values is often tedious and unintuitive. The goal is to let users specify structures directly rather than manually engineer a mapping in feature space.</p> <h2 id="method">Method</h2> <p>The workflow exposes a slice viewer in which a user marks a structure of interest. A self-supervised, pre-trained vision transformer supplies high-level features; similar features are selected across the volume to form a segmentation that drives the transfer function. Because the method relies on pre-trained features rather than per-volume training, edits can be reflected immediately.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper demonstrates annotation-guided transfer-function design on volume data and evaluates whether feature-based selections isolate target structures from only a small amount of user input. It also contrasts the interactive workflow with previous learning-based approaches that require costly optimization for each volume.</p> <h2 id="limitations">Limitations</h2> <p>The result depends on the discriminative power of the pre-trained features and on the ambiguity of the selected structure. Sparse annotations are powerful, but may need refinement when visually similar structures should receive different optical properties.</p>http://localhost:1313/stdepth/Mon, 27 Jun 2022 00:00:00 +0000http://localhost:1313/stdepth/<h2 id="motivation">Motivation</h2> <p>Depth is ambiguous in semi-transparent volume rendering: several structures can contribute to one pixel, and there may be no single visible surface. A useful depth estimate nevertheless enables labels, geometry, and additional shading to be composed with a rendering.</p> <h2 id="method">Method</h2> <p>The work studies depth definitions for volume-rendered images and adapts monocular depth networks to predict a layered decomposition. From one RGB rendering, the model estimates front semi-transparent content, the depth of a surface of interest, and structures behind it; the layers can be recomposited with new scene elements.</p> <h2 id="evaluation">Evaluation</h2> <p>The evaluation compares depth-estimation approaches across different opacity and illumination settings, and assesses the additional color and opacity predictions needed for layered reconstruction. The resulting layers support downstream re-composition tasks rather than serving only as a scalar depth map.</p> <h2 id="limitations">Limitations</h2> <p>The decomposition is learned from rendered imagery and inherits ambiguities from the rendering setup. It is therefore most useful as an image-space representation for interactive composition, not as a replacement for the original volumetric scene model.</p>http://localhost:1313/difftem/Sun, 26 Jun 2022 00:00:00 +0000http://localhost:1313/difftem/<h2 id="motivation">Motivation</h2> <p>Electron-microscopy data are expensive to acquire, noisy, and difficult to annotate. Faithful simulation can supply controlled ground truth, but conventional simulators are often too slow or too rigid to support optimization and learning.</p> <h2 id="method">Method</h2> <p>DiffTEM simulates micrographs from atomistic models at a scale suitable for many viral particles. Both the deterministic image-formation stages and the stochastic signal-and-noise stages are differentiable, allowing gradients to flow from an image-space objective back to simulation and detector parameters.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper demonstrates two uses of differentiability: estimating detector parameters automatically and using simulated data to support denoising of real microscopy observations. It also evaluates the simulator’s visual plausibility and scalability relative to previous simulation approaches.</p> <h2 id="limitations">Limitations</h2> <p>A simulator remains an approximation of a complex physical acquisition process. Its usefulness depends on how well the modeled specimen, microscope, and noise assumptions match the data to which it is applied.</p>http://localhost:1313/nano-oetzi/Tue, 06 Apr 2021 00:00:00 +0000http://localhost:1313/nano-oetzi/<h2 id="motivation">Motivation</h2> <p>Cryo-electron tomography contains valuable nanoscale structure but suffers from severe noise, making direct volume rendering and transfer-function design difficult. Experts need a way to expose meaningful structure without manually tuning a complex visual mapping for every dataset.</p> <h2 id="method">Method</h2> <p>Nano-Ötzi decomposes visual mapping into two tasks. A semi-supervised pipeline turns sparse user input into soft segmentation through pseudo labels and a stronger classifier; an iterative thresholding method then combines raw-data detail with soft labels to estimate opacity. The result is a volume-rendering workflow tailored to noisy cryo-ET data.</p> <h2 id="evaluation">Evaluation</h2> <p>The paper demonstrates the pipeline on challenging cryo-ET data and evaluates its individual components. Feedback from two domain experts complements the technical assessment by examining whether the resulting visualizations support real scientific interpretation.</p> <h2 id="limitations">Limitations</h2> <p>The learned segmentation stage benefits from substantial pre-training data, and sparse labels may not resolve every ambiguous structure. The system is designed to support expert exploration rather than remove the need for domain judgment.</p>http://localhost:1313/pbtvis/Mon, 15 Mar 2021 00:00:00 +0000http://localhost:1313/pbtvis/<h2 id="motivation">Motivation</h2> <p>Visualization software can expose many rendering parameters, while conventional regression tests usually cover only a small set of hand-picked configurations. Bugs may therefore remain hidden until a particular combination of values is encountered.</p> <h2 id="method">Method</h2> <p>The work applies property-based testing to visualization development. It generates many parameter settings automatically, renders paired configurations, and checks whether the expected visual effect of a controlled parameter change is present. These properties can be defined alongside an algorithm, making the approach compatible with test-driven development.</p> <h2 id="evaluation">Evaluation</h2> <p>The approach is integrated into Inviwo’s regression-testing infrastructure and exercised through visualization use cases. The examples show how automatically generated parameter combinations can expose errors that are unlikely to be covered by fixed image baselines alone.</p> <h2 id="limitations">Limitations</h2> <p>A property-based test is only as useful as the property it encodes. Designing robust visual assertions and diagnosing a failed generated case still require developer insight, especially when rendering differences are subtle or data-dependent.</p>http://localhost:1313/dvao/Fri, 01 Jan 2021 00:00:00 +0000http://localhost:1313/dvao/<h2 id="motivation">Motivation</h2> <p>Ambient occlusion improves depth perception in direct volume rendering, but high-quality volumetric illumination can be expensive to compute. DVAO investigates whether a learned model can precompute an ambient-occlusion volume for the current transfer function, leaving view changes interactive.</p> <h2 id="method">Method</h2> <p>DVAO is a 3D convolutional encoder–decoder that takes a scalar volume and transfer-function context and predicts one ambient-occlusion value per voxel. The strongest variant <strong>preclassifies</strong> the volume with transfer-function opacity, giving the network a spatially aligned, implicit representation of the global transfer function from its first layer.</p> <p> <figure > <div class="d-flex justify-content-center"> <div class="w-100" ><img alt="Diagram of the selected preclassified DVAO encoder–decoder architecture" srcset=" /dvao/method_hu16237650332852266575.webp 400w, /dvao/method_hu14055903595399580284.webp 760w, /dvao/method_hu3313981667586537401.webp 1200w" src="http://localhost:1313/dvao/method_hu16237650332852266575.webp" width="760" height="241" loading="lazy" data-zoomable /></div> </div></figure> </p> <p><em>The selected implicit/preclassified strategy combines opacity and intensity volumes before the encoder and predicts a 128³ ambient-occlusion volume. Adapted from Fig. 6 of the arXiv version.</em></p> <h2 id="evaluation">Evaluation</h2> <ul> <li>Across 30 training runs covering six transfer-function injection strategies and five loss functions, the implicit preclassified input performed best; the selected DSSIM-2D + MSE model reached <strong>SSIM 0.875</strong> and <strong>MSE 0.008</strong> on held-out data.</li> <li>The default 16-filter model required <strong>411 ms</strong> per transfer-function update on an RTX 2070. It is not evaluated every frame, so camera interaction remains responsive after the occlusion volume is produced.</li> <li>Although trained only on human head CT data, it obtained SSIM scores of <strong>0.786</strong> on mouse micro-CT, <strong>0.866</strong> on chameleon CT, <strong>0.751</strong> on MRI, and <strong>0.869</strong> on electron microscopy data.</li> </ul> <h2 id="limitations">Limitations</h2> <p>The prototype operates on 128³ volumes and needs substantial memory for 3D CNN inference. It was a research exploration of learned volumetric illumination rather than a replacement for optimized algorithmic ambient-occlusion methods.</p>