Conference Program

Location: in front of Cauerstrasse 11
Discover the fascinating intersection of art and artificial intelligence at our unique pre-VMV art tour near the labs of Visual Computing Erlangen at FAU! Basel artist Hans Furer and Junior Professor Bernhard Egger have launched an exciting AI experiment which is exhibited in our building at Cauerstrasse 11. An AI, trained on 50 years of Hans Furer's artistic creation, now presents its own artworks. Bernhard Egger will explain the inception of the project, how the artwork was generated and which role generative models played. After a short presentation we invite you to join us in exploring those physical artworks. This event is perfect for anyone interested in art, technology, and their convergence. Come by, be inspired, and engage in discussions with like-minded individuals about the future of art in the digital society. We can't wait to see you there!
If you are late, join us in Lecture Hall H13 on the first floor.

We present a method for computing alias-free, smooth shadows for alpha-tested geometry using a single ray per pixel. Without mipmap filtering, hard shadows from alpha-tested geometry appear very noisy under camera motion. Typically, many ray samples are required to soften the shadows and reduce noise. We propose using mipmaps instead, to achieve a fast and temporally stable solution. To determine the appropriate mipmap level, we introduce novel ray cone operations that account for directional and point light sources.

In this paper, we develop two methods for fast visualization of fully dynamic large-scale multi-layered height maps (MLHMs). MLHMs are a fairly uncommon data structure in computer graphics, which has been used as an efficient representation of 3D terrain, among other applications. Recently, a 3D hydraulic erosion simulation that utilizes this data structure effectively, allowing for real-time simulation of large scale terrain, was developed, but the fast simulation was paired with slow visualization. We extend this previous work with two efficient visualization methods. Rendering the MLHM as boxes is done using ray tracing, while a smooth surface is rendered by ray marching an implicit surface generated by smoothing the MLHM, following previous work. Both techniques are accelerated using a hierarchical data structure built directly from the MLHM, which enables quadtree-like traversal using 2D DDA, where each 2D cell contains 3D axis-aligned bounding boxes (AABBs). This data structure is adapted to accelerate ray marching by appropriately padding AABBs with the smoothing radius. We further propose a soft shadow method and geometric ambient occlusion that work in tandem with this data structure. Our visualization is fast enough to support fully dynamic terrain in real-time, where simulation, creation of our data structure and visualization are done every single frame in real-time for terrain resolutions up to 4096^2. For lower resolutions, it is possible to run expensive ray tracing and geometric ambient occlusion effects at full window resolution every frame with real-time or interactive frame rates.

We present a method to acquire and estimate the sub-surface scattering properties of light transport at a highly detailed level by learning the pixel footprint response at each point on the object surface. The reconstruction leverages 3D scanning techniques as input to a U-Net CNN. A stereo projector-camera setup using phase-shifted profilometry (PSP) patterns efficiently captures the data for a variety of scattering objects. Reconstructing dense pixel footprints allows for relighting with arbitrary high-resolution projector patterns. The final output is a relit image. Qualitative and quantitative comparison against illuminated real-world captured images demonstrate that the predicted footprints are almost identical to the actual responses. The same model is trained in a multi-view setting across multiple objects such that the learned representations can be used to generalize to unseen sub-surface scattering materials as well.

Implicit Neural Point Cloud (INPC) is a recent hybrid representation that combines the expressiveness of neural fields with the efficiency of point-based rendering, achieving state-of-the-art image quality in novel view synthesis. However, as with other high-quality approaches that query neural networks during rendering, the practical usability of INPC is limited by comparatively slow rendering. In this work, we present a collection of optimizations that significantly improve both the training and inference performance of INPC without sacrificing visual fidelity. The most significant modifications are an improved rasterizer implementation, more effective sampling techniques, and the incorporation of pre-training for the convolutional neural network used for hole-filling. Furthermore, we demonstrate that points can be modeled as small Gaussians during inference to further improve quality in extrapolated, e.g., close-up views of the scene. We design our implementations to be broadly applicable beyond INPC and systematically evaluate each modification in a series of experiments. Our optimized INPC pipeline achieves up to 25% faster training, 4x faster rendering, and 20% reduced VRAM usage paired with slight image quality improvements.

Neural fields offer continuous, learnable representations that extend beyond traditional discrete formats in visual computing. We study a fundamental inverse problem in this setting: learning repeated antiderivatives directly from sampled data, as a continuous analogue of summed-area tables. Although widely used in discrete domains, such cumulative schemes rely on grids, which prevents their applicability in continuous neural contexts. We introduce and systematically analyze a range of neural methods for repeated integration, including both adaptations of prior work and novel designs. Our evaluation spans multiple input dimensionalities and integration orders, assessing both reconstruction quality and performance in downstream tasks such as filtering and rendering. These results lay the groundwork for integrating classical cumulative operators into modern neural systems and offer broader insights into learning tasks involving differential and integral operators.

Quantised autoencoders usually split images into local patches, each encoded by one token. This representation is potentially inefficient, as the same number of tokens are spent per region, regardless of the visual information content in that region. To mitigate uneven distribution of information content, modern architectures provide an adaptive discretisation or add an attention mechanism to the autoencoder to infuse global information into the local tokens. Despite these improvements, tokens are still associated with a local image region. In contrast, our method is inspired by spectral decompositions which transform an input signal into a superposition of global frequencies. Taking the data-driven perspective, we train an encoder that produces a combination of tokens that are then decoded jointly, going beyond the simple linear superposition of spectral decompositions. We achieve this global description with an efficient transpose operation between features and channels and demonstrate how our global and holistic representation improves compression and can boost downstream tasks like generation.

Accurate depth estimation is at the core of many applications in computer graphics, vision, and robotics. Current state-of-the-art monocular depth estimators, trained on extensive datasets, generalize well but lack metric accuracy needed for many applications. In this paper, we combine the strength of those general monocular depth estimation techniques with multi-view data by framing this as an analysis-by-synthesis optimization problem to lift and refine such relative depth maps to accurate error-free depth maps. After an initial global scale estimation through structure-from-motion point clouds, we further refine the depth map through optimization enforcing multi-view consistency via photometric and geometric losses with differentiable rendering of the meshed depth map. In a two-stage optimization, first scaling is further refined, and afterwards artifacts and errors in the depth map are corrected via nearby-view photometric supervision. Our evaluation shows that our method is able to generate detailed, high-quality, metrically accurate depth maps, also in challenging indoor scenarios, and outperforms state-of-the-art multi-view depth reconstruction approaches on such datasets.

Computer Graphics Forum (Proc. Eurographics 2025)

ACM Transactions on Graphics (Proc. SIGGRAPH 2025)

IEEE Journal of Microwaves (2025)

Computer Vision and Pattern Recognition Conference (CVPR 2025)

International Conference on Learning Representations (ICLR 2025)

IEEE Open Journal of Vehicular Technology (OJVT 2025)

Computer Graphics Forum (CGF 2025)

International Conference on 3D Vision (3DV 2025)

Using multiple hand sensors, hand kinematics can be measured in Human-Computer Interaction (HCI) with the intention to predict the user's intention in a reach-to-grasp action, leading to multivariate time series data. Then, the goal is to classify the multivariate time series data, where the class shall be predicted as early as possible. To investigate the prediction evolution, detect and analyze challenging conditions, and identify the best trade-off between early prediction and prediction quality, we present XMTC. XMTC incorporates visualizations on accuracy over time, multi-variate time series classification probabilities, confusion matrices, and partial dependence plots for a trustworthy classification production. We employ XMTC to real-world HCI data in multiple scenarios to achieve good early classifications, as well as insights into which conditions are easy to distinguish, which multivariate time series measurements impose challenges, and which features have the most impact.

Understanding how behavior changes under genetic or experimental conditions is a key challenge in behavioral neuroscience. High-throughput tracking enables the collection of high-dimensional datasets describing locomotion, posture, and stimulus orientation in Drosophila melanogaster larvae (fruit fly). However, exploring relations across numerous dimensions remains challenging. We present a Visual Analytics system that integrates coordinated views, type-aware relation metrics, and hierarchical clustering to support relation discovery and validation in behavioral data. The system was initially developed based on prior experience and refined through evaluation with domain experts to address key analysis tasks, including grouping dimensions, exploring behavioral patterns, and validating hypotheses. We demonstrate how it supports both confirmatory and exploratory workflows, enabling users to confirm known effects and uncover novel patterns—such as an unexpected correlation between head-casting behavior and locomotion speed. This work highlights how tailored visual analysis can advance behavioral research.

Large Language Models (LLMs) are emerging as promising approaches for tabular data generation and enrichment, helping ease constraints related to data availability. However, the reliable use of LLM-generated data remains challenging due to hallucinations, and inconsistencies. While some validation approaches exist, key challenges remain: lack of explanations and transparency in how values are generated, balancing fine-grained accurate with coarse-grained scalable validation, validating without ground truth, and evaluating plausibility, semantic relevance, and downstream utility. To address these challenges, we present Val-LLM, a novel Visual Analytics approach for validating LLM-generated tabular data. Val-LLM enables users to contextualize generated values with explanations, externalize knowledge, and relate LLM outputs with existing data. We conducted a user study to evaluate the approach. Results highlight the usefulness of supporting multiple levels of granularity and enabling human knowledge externalization for validation. The study also indicates the need to study validation workflows and workflow flexibility, based on user domain experience and user preferences. Our work supports the trustworthy and effective use of LLM-generated tabular data by integrating visual analytics for systematic validation.

We present an in-depth investigation of the Apple Vision Pro as a platform for large-scale volume visualization, focusing on both its technical capabilities and practical limitations in real-world immersive rendering scenarios. Our study centers on BorgVR, a custom-built volume rendering system that implements a modern bricked, ray-guided, and out-of-core rendering pipeline tailored to the unique characteristics of the Vision Pro and the visionOS graphics stack. BorgVR is designed to overcome memory and performance bottlenecks associated with rendering structured grids that exceed device-local memory. Through dynamic data streaming, hierarchical bricking, and GPU-accelerated early ray termination, the system achieves interactive frame rates for gigabyte-scale datasets, even under the constraints of mobile spatial computing. We analyze how well the Apple Vision Pro supports such workloads across its distinct rendering modes, ranging from video see-through AR to fully immersive VR. Beyond demonstrating system performance, we critically evaluate the Vision Pro’s suitability for scientific visualization—highlighting its strengths in display fidelity and sensor integration, while also documenting friction points such as GPU architecture constraints, memory management, and platform-specific development hurdles. The open-source release of BorgVR provides a reusable foundation for the community, facilitating future research and application development in immersive volume visualization.

Location: Room 2.018, Kollegienhaus, Universitätsstraße 15, 91054 Erlangen

Location: Room 0.015, Kollegienhaus, Universitätsstraße 15, 91054 Erlangen

Location: Room 00.027, Orangerie, Schlossgarten 1, 91054 Erlangen

We present a study of how to integrate color (RGB) and multi-spectral imagery (red, green, red-edge and near-infrared) into the 3D Gaussian Splatting (3DGS) framework -- a leading explicit radiance-field method for fast and high-fidelity 3D reconstruction from multi-view images. While 3DGS excels on RGB data, naive per-band optimization of additional spectra yields poor reconstructions due to inconsistently appearing geometry in the spectral domain. This problem is prominent, even though the actual geometry is the same, regardless of spectral modality. To investigate this, we evaluate three strategies: 1) Separate per-band reconstruction with no shared structure; 2) Splitting optimization, in which we first optimize RGB geometry, copy it, and then fit each new band to the model; and 3) Joint, in which the modalities are jointly optimized, optionally with an initial RGB-only phase. We showcase through quantitative metrics and qualitative novel-view renderings on multi-spectral datasets the effectiveness of our dedicated optimized Joint strategy, increasing overall spectral reconstruction as well as enhancing RGB results through spectral cross-talk. We therefore propose the integration of multi-spectral data into the spherical harmonics color components to model each Gaussian’s multi-spectral reflectance compactly. Our analysis reveals key trade-offs in when and how to introduce spectral bands during optimization, offering practical hints for robust multi-modal 3DGS reconstruction.

Accurate camera calibration is crucial for high-quality 3D reconstruction in computer vision applications. In industrial measuring scenarios, turntable sequences are often captured using telecentric lenses to overcome the foreshortening effect. While specialized Structure-from-Motion (SfM) solutions exist for orthographic projection, these methods are limited to textured objects. Approaches that leverage the scanned object's silhouette for camera calibration are independent of texture but are often restricted to smooth objects or require non-trivial optimization initializations to converge. In this work, we present a novel silhouette-based approach to estimate the rotation axis of a turntable under orthographic projection, extending the applicability to complex geometries, while requiring little to none parameter adjustments. By identifying the symmetry axis of the object's contour envelope and establishing frontier point correspondences on circular trajectories, we robustly estimate the azimuth and inclination angles of the rotation axis, enabling accurate camera pose computation. We evaluate our approach on synthetic datasets comprising four models with varying characteristics and compare it to a state-of-the-art orthographic SfM method, achieving comparable accuracy, while reducing computational cost 37-fold and eliminating reliance on object texture.

Image segmentation has gained a lot of attention in the past. When working with photometric stereo data, we discovered that shadow cues provide valuable spatial information, especially when combining multiple images of the same scene under different lighting conditions. In the following, we present a robust method to pre-segment images, relying heavily on shadow masks as the main input. We first detect object contours from light to shadow transitions. In the second step, we run an image segmentation algorithm based on Delaunay triangulation that is capable of closing the gaps between contours. Our method requires spatial input data but is free from training data. Initial results look promising, generating pre-segmentations close to recent data-driven image segmentation algorithms.

Zero-shot anomaly localization is a rising field in computer vision research, with important progress in recent years. This work focuses on the problem of detecting and localizing anomalies in textures, where anomalies can be defined as the regions that deviate from the overall statistics, violating the stationarity assumption. The main limitation of existing methods is their high running time, making them impractical for deployment in real-world scenarios, such as assembly line monitoring. We propose a real-time method, named QFCA, which implements a quantized version of the feature correspondence analysis (FCA) algorithm. By carefully adapting the patch statistics comparison to work on histograms of quantized values, we obtain a 10x speedup with little to no loss in accuracy. Moreover, we introduce a feature preprocessing step based on principal component analysis, which enhances the contrast between normal and anomalous features, improving the detection precision on complex textures. Our method is thoroughly evaluated against prior art, comparing favorably with existing methods.

Interactive editing of character images with diffusion models remains challenging due to the inherent trade-off between fine-grained control, generation speed, and visual fidelity. We introduce CharGen, a character-focused editor that combines attribute-specific Concept Sliders, trained to isolate and manipulate attributes such as facial feature size, expression, and dec- oration with the StreamDiffusion sampling pipeline for more interactive performance. To counteract the loss of detail that often accompanies accelerated sampling, we propose a lightweight Repair Step that reinstates fine textures without compromising structural consistency. Through extensive ablation studies, comparisons against open-source InstructPix2Pix and closed-source Google Gemini, and a comprehensive user study, CharGen achieves two-to-four-fold faster edit turnaround with precise editing control and identity-consistent results.

Location: Redoutensaal, Theaterplatz 1, 91054 Erlangen

From: Busparkplatz Vestnertor, Vestnernorgraben 4, 90408 Nürnberg (Directions)
To: Bratwurst Röslein, Rathausplatz 6, 90403 Nürnberg (Directions)

Location: Bratwurst Röslein, Rathausplatz 6, 90403 Nürnberg (Directions)

Many geometry processing algorithms rely on solving PDEs on discrete surface meshes. Their accuracy and robustness crucially depend on the mesh quality, which oftentimes cannot be guaranteed -- in particular when automatically processing geometries extracted from implicit representations in machine learning applications. Through extensive numerical experiments, we evaluate the robustness of various Laplacian implementations across modern geometry processing libraries on synthetic and ``in-the-wild'' surface meshes with degenerate or near-degenerate elements, revealing their strengths, weaknesses, and failure cases. To improve numerical stability, we extend the recently proposed tempered finite elements method (TFEM) to meshes with strongly varying element sizes, to arbitrary polygonal elements, and to gradient and divergence operators. Our resulting differential operators are simple to implement, efficient to compute, and robust even in the presence of fully degenerate mesh elements.

Computing maps between data sequences is a fundamental problem with various applications in the fields of geometry and signal processing. As such, a multitude of approaches exist, that make trade-offs between flexibility, performance, and accuracy. Even recent approaches cannot be applied to periodic data, such as contours, without significant compromises due to their map representation or method of optimization. We propose a universal method to optimize maps between periodic and non periodic univariate sequences. By continuously optimizing a piecewise linear approximation of the smooth map on a common intermediate domain, we decouple the map and input resolution. Our optimization offers bijectivity guarantees and flexibility with regards to applications and data modality. To robustly converge towards a high quality solution we initially apply a low- pass filter to the input. This creates a scale space that suppresses local features in the early phase of the optimization (global phase) and gradually adds them back later (local phase). We demonstrate the versatility of our method on various scenarios with different types of sequences, including multi-contour morphing, signature prototypes, symmetry detection, and 3D motion- capture-data alignment.

Accurate cloth simulation is a vital component in computer graphics, virtual reality, and fashion design. Position-Based Dynamics (PBD) and its extension (XPBD) offer robust and efficient methods for simulating deformable objects like cloth. This paper details the evaluation and comparison of cloth simulations based on XPBD, including its "small steps" variant and an Energy-Aware (EA) modification. The XPBD variants are evaluated for their physical plausibility, energy conservation, and constraint satisfaction to analyze their suitability for inverse problems. Furthermore, we explore the implementation of a differentiable XPBD simulator, enabling automatic parameter estimation for material properties and external forces. The differentiable simulator is assessed for its capability to estimate parameters in scenarios of increasing complexity. Results indicate that small time steps with single iterations in XPBD offer good energy behavior, while the EA modification exhibits undesired characteristics. The differentiable simulator successfully estimates single parameters but identifies challenges with multi-parameter optimization due to compensatory effects.

We present binned variable block compressed sparse row (Bin-VBSR), a novel GPU-optimized sparse matrix data structure and associated sparse matrix-vector multiplication algorithm for matrices with variable-size dense blocks. This includes a novel approach to handling long rows in the binned compressed sparse row (Bin-CSR) family of GPU-optimized sparse matrix data structures. We demonstrate speedups of up to 9.9× over Bin-BCSR* and extend its data compression advantages over compressed sparse row (CSR) to variable block size, resulting in an improvement of up to 50%.

This paper introduces a geometry-aware method for analyzing swarm behavior in Particle Swarm Optimization (PSO) based on ellipsoidal modeling. Inspired by the n-ball hitting probability, we propose an abstraction of the search space covered by particles over time. Using principal component analysis (PCA), we approximate the particle distribution at each iteration with ellipsoids, enabling a visual and quantitative assessment of how well the swarm explores and concentrates its search effort. We apply this technique to investigate a PSO variant with negative inertia weights, which has shown promising performance in prior empirical analysis. While negative inertia may appear counterintuitive, our ellipsoidal analysis reveals that it introduces oscillatory search dynamics that balance exploration and exploitation more effectively than standard strategies such as constant inertia or linear decreasing inertia. Our experiments include a six-dimensional medical image registration task and an illustrative two-dimensional Rastrigin function, which serves to visually demonstrate how the swarm structure evolves. The proposed analysis framework provides new insight into swarm dynamics and offers a tool for understanding and comparing the behavior of PSO variants beyond conventional performance metrics.