Presents corrections to the paper, Move or Push? Studying Pseudo-Haptic Perceptions Obtained With Motion or Force Input.
Figurative glyph-based visualizations (FGVs) convey data through recognizable visual objects, but their creation is labor-intensive and conceptually demanding. A central challenge lies in enabling figurative mapping of data variables onto distinct components of a generated visual object while ensuring that automatically produced designs remain aligned with the user intent. We present Gen-lyph, a human-AI collaboration system that combines generative models with interactive refinement to support semantically rich FGV design. Informed by formative interviews and an analysis of existing FGVs, we developed a fourphase generative pipeline comprising ideation, decomposition, encoding, and placement, integrating automated generation with user control. Gen-lyph allows users to guide figurative glyph generation iteratively, segment and assign data attributes, and progressively refine encodings and arrangements. Gen-lyph was evaluated through an expert review, a free exploration study, and a usage scenario.
Current music-to-dance generation methods mainly rely on musical features, limiting precise control over generated movements. In particular, most existing methods with control mechanisms do not support example-based control, in which a user provides a reference motion sequence and the generated dances follow its fine-grained motion patterns while adapting to different musical pieces. While several motion-guided editing methods incorporate reference motion, they impose it as sparse positional constraints that capture local poses without modeling the reference's overall motion characteristics or adapting them to the target music. To tackle this limitation, we propose a novel framework that integrates reference motions as additional guidance for controllable dance generation while maintaining music synchronization. Our approach employs a hierarchical motion representation learning framework to capture both global characteristics and temporally coherent local details from reference sequences. A motion-music integration approach is then applied to combine the extracted motion features with a pretrained music to-dance diffusion model through efficient fine-tuning. Addition ally, cycle consistency regularization is incorporated to ensure robust generalization across diverse motion-music pairs. Extensive experiments demonstrate that our framework achieves effective controllability while maintaining high-quality motion generation comparable to state-of-the-art music-to-dance methods. The code is available at https://anonymous.4open.science/r/Controllable Music-to-Dance-FF76/
Transferring animations between characters with diverse skeletal structures is challenging. Traditional retargeting pipelines rely on fixed correspondences, canonical skeletons, or human-centric datasets, which can lead to artifacts when applied across heterogeneous morphologies. We introduce a framework for cross-morphology motion transfer with semantic style alignment that uses morphology-agnostic control signals (e.g., velocity, angular velocity, relative height) to align behaviors across species. Our method supports all-to-all retargeting: motions from any source can be mapped to any trained target while preserving target-specific style. For each target morphology, we train a Vector Quantized VAE and an autoregressive sequence model to construct a compact, morphology-specific codebook that captures stylistic priors. This modular design scales to new morphologies without retraining existing models and allows optional user control (e.g., phase, velocity scaling) for fine-grained alignment. Experiments across bipeds and quadrupeds demonstrate accurate, plausible, and style-faithful motion transfer, establishing a scalable approach to retargeting across arbitrary skeletal topologies.
Individuals with Color Vision Deficiency (CVD) face difficulties in accurately distinguishing between colors due to reduced perceptual contrast. To address this issue, various image recoloring methods have been proposed. They perform recoloring either by using a CVD simulation model or by training deep learning models on images generated with such simulation models. These simulation models are based on assumed severity levels of color vision deficiency. However, in reality, there is currently no reliable way to measure the actual severity, and human color perception is highly complex and subjective. As a result, existing methods may not fully capture individual perceptual characteristics or user-specific preferences.To address this, we adopt an end-to-end learning approach based on each user's subjective evaluation. A major challenge with such approaches is the large amount of training data typically required. To overcome this limitation, we propose a new model and training strategy, along with a dataset specifically designed to capture key features of individual color perception. This enables us to train a personalized model for each user using only a very small amount of teacher data. In addition, we will release the source code and an anonymized dataset comprising data from 20 protan and deutan CVD users to support reproducibility and future research on personalized color compensation.
Real-time deformable simulation relies on bounding volume hierarchies (BVHs) that must be updated continuously under mesh deformation. Existing BVH maintenance strategies are executed as separate post-processing stages, requiring repeated global scans of mesh geometry regardless of deformation locality. This design introduces unnecessary computational overhead and limits scalability in interactive applications. We revisit this design assumption and propose a solver-integrated BVH framework that embeds hierarchical maintenance directly into the physics solver's per-vertex update loop. By performing lightweight boundary checks during vertex updates and propagating changes only when bounds are violated, the method eliminates redundant full-mesh traversals and restricts recomputation to deformation-affected regions. We establish the correctness of the update for triangle meshes with shared vertices through a one-to-many vertex-to-leaf mapping, and we evaluate the approach across meshes ranging from a few thousand to 7.2 million triangles, multiple deformation regimes, a real XPBD cloth solver, baked skeletal character animation, parallel multi-core CPU execution, and a GPU compute implementation. Under static and localized deformation the method reduces BVH update cost by up to roughly 20 times relative to the refitting, kinetic, and dynamic fat-bound baselines, for example from 19 to 62 ms down to 2.9 ms on an 871 thousand triangle mesh at rest, while preserving collision-detection correctness on every frame. We further report end-to-end collision cost and identify the globally contact-rich regime in which a dynamic fat-bound hierarchy remains competitive, giving a clear account of where solver-integrated maintenance helps and where it does not. These results show that integrating acceleration-structure maintenance into the simulation loop is a more efficient and scalable alternative to traditional post-processing pipelines for real-time deformable systems.
Gaussian Splatting has emerged as a prominent technique in computer vision and graphics, enabling high-fidelity scene reconstruction from a vast collection of 3D Gaussian primitives. Each primitive's appearance is governed by a viewindependent opacity parameter and a view-dependent color derived from spherical harmonics. A fundamental challenge, however, lies in the non-unique and heterogeneous nature of this parameterization, which limits physical plausibility and degrades expressiveness. To address this, we introduce the Extinction Coordinator for Gaussian Splattings (ECGS), which enforces a crucial consistency: a Gaussian's intrinsic opacity must align with its maximum alpha blending weight observed across all views. This constraint effectively encourages primitives to distribute along thin, surface-like shells, thereby enhancing the representation's expressiveness. Furthermore, we propose an anisotropic morphology regularization to promote planar Gaussian shapes while suppressing elongated, needle-like artifacts, leading to a more compact model. Extensive experiments on multiple benchmarks demonstrate that our approach reduces the Gaussian count by up to 75% while simultaneously boosting geometric accuracy and rendering speed. As a lightweight and modular component, ECGS can be seamlessly integrated into existing Gaussian Splatting pipelines. Code is available at: https://anonymous.4open.science/r/ECGS-F774.
Line art colorization is a critical yet labor-intensive bottleneck in professional animation production, requiring artists to maintain rigorous color consistency across frames. While contemporary methods leveraging feature alignment or generative diffusion have improved efficiency, they often fail when confronted with occlusions and large inter-frame deformations. The fundamental difficulty arises from the inherent sparsity of binary line-art images: the absence of internal texture limits the discriminative power of local feature descriptors, resulting in ambiguous region correspondences during propagation. To address these limitations, we propose a region-based feature enhancement framework built upon a Topology-aware Segment Graph (TSG). By explicitly modeling topological adjacency—a robust structural cue that remains invariant across frames despite significant geometric changes—our method leverages spatial context to increase feature separability across semantically distinct regions, enabling more robust alignment. Furthermore, to mitigate temporal drift and the diminishing relevance of static reference frames in long sequences, we introduce a Dual-source Temporal Selection and Refinement strategy. This mechanism adaptively integrates information from both the global reference keyframe and the immediate previous frame, balancing color fidelity with spatial continuity. Extensive experiments on industrial benchmarks demonstrate that our approach achieves superior color accuracy and temporal stability compared to state-of-the-art frameworks, particularly in scenarios involving complex character motion and topological variation.
Real-time visualization of massive 3D virtual environments, such as city-scale digital twins and immersive simulations, remains a core challenge in computer graphics. While 3D Gaussian Splatting (3DGS) achieves photorealistic rendering at interactive rates, its explicit Gaussian representation incurs substantial GPU memory overhead, limiting deployment on resource-constrained devices. Existing 3DGS Level-of-Detail (LoD) methods are often data-dependent, requiring access to large collections of original training images, and they rarely adapt to dynamic hardware budgets, leading to unstable performance on low-VRAM GPUs. We present InfoLoD, a data-free and hardware-aware framework for hierarchical LoD construction and large-scale 3DGS rendering. InfoLoD introduces a Fisher-guided self-distillation scheme that uses the Fisher Information Matrix (FIM) to select geometrically valid, information-rich pseudo viewpoints, enabling LoD training directly from a pre-trained 3DGS model without any original images. To operate under strict VRAM constraints, we further propose a budget-aware out-of-core streaming system with spatio-temporal working-set management and resource-adaptive LoD scheduling to balance fidelity and throughput. Experiments on challenging outdoor scenes show that InfoLoD enables stable real-time rendering on resource-constrained GPUs, delivering higher efficiency than prior baselines while maintaining competitive visual quality.
In complex and hazardous environments, robots often require human's fine teleoperation to ensure accurate decision-making and optimal manipulation. To improve teleoperation efficiency, we design and develop 3D predictive Virtual Reality (VR) interfaces with four distinct automation levels. These 3D interfaces, coupled with physics-based simulation and task-level automatic planning, assist users in controlling robots for fine manipulation. Our comparative user study evaluates the efficiency of these interaction strategies, investigating the feasibility of solely relying on simulation results for robot control. We further compare VR controller-based teleoperation with direct hand gesture manipulation, assessing efficiency and user interaction experience. Our results demonstrate that predictive interfaces and intelligent assistance significantly improve operational efficiency. VR controller-based teleoperation outperforms hand gesture interaction in terms of efficiency and user experience for fine manipulation. In particular, even novice users can effectively utilize the predictive VR interface for precise remote robot control. This work presents a novel approach of integrating predictive interaction, physics-based simulation, and task-level automatic planning for fine teleoperation.
VR research predominantly focuses on upright postures, such as sitting or standing. Recent research has begun exploring VR use while lying down, revealing numerous benefits. However, existing interaction techniques are tailored for upright postures, creating challenges when applied to a lying-down position. We conducted an experiment to understand user performance across different spatial dimensions (azimuth, inclination, and distance) during a target selection task in a lying-down posture. Our findings revealed that hand movement increases nearly linearly with target offset and that users experience noticeable fatigue and a desire to minimize physical effort. Motivated by these insights, we identified two prevalent hand redirection methods, Go-Go and RBNL, and, based on empirical performance data, developed an empirical data-driven compensation scaling (DDCS) technique that dynamically scales virtual hand movements. Evaluations indicate that both DDCS and RBNL can significantly reduce hand movement and accelerate interaction in supine VR, though at the cost of increased spatial offsets that may compromise precision; however, users generally prioritize comfort over high accuracy. We further discussed the potential reasons and future implications of adopting VR in a supine posture.
Novel view synthesis methods, such as neural radiance fields and 3D Gaussian splatting, offer a promising solution for photorealistic rendering. However, they remain challenged in few-shot settings, where models tend to overfit the limited supervised views, leading to artifacts such as quality fluctuations, degradation in distant views, and geometric inconsistencies. To address these issues, we introduce Human Perceptual Preference Optimization (HuPPO), a framework that incorporates human perceptual guidance into model training. HuPPO mitigates distortions by regularizing training dynamics with perceptual preference cues, thereby reducing the reliance on extensive supervised views. Specifically, HuPPO leverages human perception to identify and select candidate novel views, and introduces a corresponding objective function that steers optimization toward perceptually preferred outcomes. In addition, a meta-learning pipeline is integrated to promote the learning of generalizable representations. The framework is flexible and can be seamlessly applied to a wide range of neural rendering models without incurring additional inference overhead. Extensive experiments and analyses demonstrate that HuPPO achieves consistent improvements over state-of-the-art baselines.
Audio-driven 3D facial animation is essential for advancing immersion and interactivity in virtual experiences. Although recent advances have shown promising capabilities, the training and evaluation of existing methods typically rely on ground-truth-based errors, which fall short of aligning with human preferences. To address this, we present a comprehensive framework that learns an automatic perceptual model from human preference data and leverages it to improve and evaluate the perceptual quality of audio-driven 3D facial animation. To begin with, we construct FMPair (Facial Motion Pairwise preference), the first human preference dataset for audio-driven 3D facial animation, which is built through a systematic annotation pipeline and comprises 65,574 annotated 3D facial motion pairs from 8,834 distinct in-the-wild audio clips. Based on the pairwise comparison dataset, we propose a Facial Motion Reward model, termed FMReward, which takes audio and 3D facial motion as inputs and predicts a perceptual quality score aligned with human preferences. Building upon FMReward, we further introduce Facial Motion reward Feedback Learning (FMFL), a direct fine-tuning algorithm that leverages a pretrained reward model to optimize diffusion-based audio-driven 3D facial animation models for better alignment with human preferences. Extensive experiments demonstrate the superiority of FMReward over other metrics in aligning with human preferences and the effectiveness of FMFL in improving the perceptual quality of audio-driven 3D facial animation. The dataset and codes will be released at: https://github.com/wsj-sjtu/FMReward.
We explore how running speed affects target selection in immersive environments. Through a Fitts' tapping task, we investigate direct touch selections while participants are standing or running (at speeds up to 9 km/h). We conduct a multifaceted controlled study (n=22) to understand the effect of speed, pointing phase, and foot-strike on selection performance. We find that speed dramatically increases movement time, error rate, perceived exertion, and task load. The increased movement time is driven by the correction phase, which doubles in duration at 9 km/h, while the ballistic phase remains stable. Our exploratory analysis further reveals that a higher cadence reduces the error rate at lower speeds (4.5 km/h), while this effect disappears at higher speeds (9 km/h), as participants' motor control in target selection is significantly disrupted by the movement oscillations. As the selection error rate quickly exceeds 30% when speed increases, even with accurate, fast external tracking, we argue that direct touch interaction is likely inappropriate for running. Further efforts are needed to facilitate selection interaction at higher running speeds.
Counterfactual explanations offer an intuitive way to understand the prediction behavior of machine learning models. However, they still face major challenges in practice. Explanation quality is often unclear due to feature redundancy and the lack of systematic evaluation. The resulting information is fragmented and hard to integrate into a coherent decision logic. To address these issues, this paper proposes a new analysis paradigm. It reframes counterfactual explanation as an auditable and iterative investigative reasoning workflow. The paradigm consists of three tightly coupled steps. First, it focuses on causal features to reduce the initial complexity of the explanation space. This is achieved through the Minimal Feature Boundary (MFB) and interactive constraint modeling. Second, it introduces multi-dimensional quality audit views. These views incorporate metrics such as the Counterfactual Equilibrated Quality Score (CEQS). They support evidence-based filtering and comparison of counterfactual candidates. Third, it establishes a visualization-driven hypothesis testing loop. This loop links coordinated views to distribution comparison and causal exploration. It allows users to dynamically validate and iteratively refine their understanding of model's behavior. The paradigm is instantiated in the visual analytics system PeCE. Its effectiveness is demonstrated through case studies and expert interviews. The results show that the proposed paradigm improves transparency and traceability in the counterfactual explanation process. As a result, counterfactual explanations become more credible and practical as diagnostic tools.
Dynamic four-dimensional (4D) Gaussian Splatting has emerged as a powerful explicit representation for highquality view synthesis, yet existing methods still require tens to hundreds of megabytes per scene due to their heavy reliance on large multi-resolution hash tables and high-dimensional Gaussian attributes. This paper presents CC-4DGS, a storage-efficient and scalable framework that rethinks both deformation modeling and canonical attribute storage. First, we introduce a computational deformation field (CDF) that replaces large multi-resolution learnable hash tables with deterministic dense hash encoding and compact neural decoders, enabling on-the-fly synthesis of deformation features while reducing deformation storage to only 1-3 MB per scene. Second, we propose a compression of canonical point-cloud attributes (CCA) pipeline that compresses highdimensional spherical harmonic appearance terms and auxiliary Gaussian attributes via conditional autoencoding, selective quantization, and residual codebooks, achieving 3-5× point-cloud reduction with negligible quality loss. Together, these components yield a unified representation that preserves real-time rendering performance while reducing total storage to 20-30 MB. Extensive experiments across the N3DV and Technicolor Light Field datasets demonstrate that CC-4DGS achieves reconstruction accuracy comparable to state-of-the-art methods such as Swift4D, while offering significantly improved storage efficiency and favorable runtime-memory trade-offs.
Kernel density estimation (KDE) is a fundamental tool across statistics, data science, visualization and machine learning. In particular, it is widely used in analyzing and visualizing multi-dimensional data sets, including, e.g., geo-spatial data. Spatio-temporal KDE is an important extension to the basic KDE problem that also takes into account the temporal relation of the data points, and provides a method to visualize evolving data patterns over time. With the explosive growth of Big Data in both size and complexity, it is typically not feasible to run naive exact algorithms due to the extremely high run time and memory-space requirements. Especially challenging is the streaming setting where the data points arrive in an online fashion, one time step at a time, and we can only access the data in one pass with a fixed, sub-linear memory. Such a challenging setting renders most of the existing methods to be inadmissible. In this paper, we take on this challenge by providing scalable approximate frameworks with low memory footprints under the streaming setting. We present randomized algorithms for efficient computation of spatio-temporal KDE with strong theoretical guarantees and empirical support. Specifically, we introduce a weighted sampling scheme and discuss a random-feature-based sketching method, and analyze their complexity and use cases. The proposed methods are sub-linear in memory and do not require the knowledge of the length of the data stream in advance. Also, the methods provide nice trade-offs between speed and accuracy. Moreover, the sketching method can easily support deletions of data points from the stream, which is not trivial for most sampling methods to perform robustly. In addition, we show how the sketching method can be easily used to answer aggregate density queries, with a computational advantage over any sampling approaches. Experiments on synthetic and real world datasets demonstrate the quality and efficiency of these proposed techniques.
Neural implicit methods reconstruct surfaces from point clouds by learning a continuous function, usually signed or unsigned distances. For unsigned distance field (UDF) reconstruction, many self-supervised methods typically estimate distance through local, gradient-guided displacement regression from query points to the surface. Because these corrections are local, they can become unstable near the zero level set and accumulate errors under sparse or non-uniform sampling. We address this limitation by reformulating UDF learning as a geometrically constrained trajectory-evolution problem, shifting from static local regression to global path-consistent modeling. Specifically, we construct a bidirectional linear flow that promotes shortest-path trajectories between surface samples and query points. Under this formulation, distance prediction is interpreted as continuous state evolution along deterministic trajectories, inherently preserving geometric consistency and mitigating local error accumulation. This perspective connects UDF estimation to neural ODEs, where geometry is recovered through continuous dynamics, with the learned transport following structured trajectories induced by an implicit geometric field. To address sparse and non-uniform sampling, we further introduce a flow-guided iterative densification strategy that progressively upgrades shortest-path modeling from discrete point-to-point approximations to accurate point-to-surface projections, enabling implicit geometric completion. Extensive experiments on synthetic benchmarks and real-world scans demonstrate that our method achieves state-of-the-art performance in both reconstruction accuracy and robustness.
We contribute PERCEIV (PERCeption of visual Encodings in InfoVis), a large-scale dataset to support studies of user responses to visual encoding variants as a foundation for adaptive information visualization interfaces. The dataset includes recordings from 120 participants who completed tasks of varying difficulty, while simultaneous multimodal data were collected from brain, eye, and electrodermal activity. All data are time-synchronized and accompanied by relevant annotations and participant metadata. We release raw sensor streams together with fully processed event-aligned derivatives and rich trial-level annotations. We also provide baseline analyses and scripts for ex tracting cognitive load features across modalities. In addition, we make the code for data pre-processing, feature extraction, qual ity control, and training/evaluating baseline Machine Learning classifiers available. Standardized data splits and comprehensive documentation further enable reproducible benchmarking and rapid reuse. We hope our dataset will facilitate novel studies of sensor modalities, development of computational models, and deeper understanding of cognitive state dynamics in controlled information-processing contexts.
