Monocular video provides an accessible source of human motion, but reconstructed body-centric motions do not preserve spatial relationships when transferred to furniture with different geometries. To address this problem for sitting interactions, we propose a furniture-adaptive motion retargeting framework. Our framework predicts interaction importance to represent the spatial relationships between the body-centric motion and the target furniture. We use the predicted pelvis importance to guide furniture adaptation and determine the interaction phase based on the detected sitting onset. To this end, we introduce interaction-guided geometry alignment and phase-aware motion composition, which adapt the motion to the target furniture and integrate interaction-aware key-joint motion according to the interaction phase. Quantitative and qualitative evaluations show that interaction-phase-aware adaptation reduces unnecessary motion changes outside the interaction phase, while phase-aware composition improves adaptation to the target furniture while preserving the characteristics of the source motion.
Synthetic data are most valuable where general-purpose datasets cannot provide the domain-specific priors a task requires, and where manual annotation is expensive, imprecise, or infeasible. In this article we argue that the central question for specialized vision systems is not how to generate more data, but which data to generate. We therefore discuss curated synthetic data, whose scene content, appearance variations, sensing characteristics, and annotations are deliberately designed around a given task. We examine three complementary curation paradigms. Procedural rendering offers explicit control over scene parameters and the annotations follow by construction. Physically-based simulation encodes the mechanism behind an observed effect and yields exactly aligned supervision pairs. Generative AI learns sensor-specific appearance from small real seed sets and attains plausible realism, though it remains prone to hallucination and to inaccurate annotation. These paradigms are illustrated with examples from industrial surface defect detection, restoration of degraded digitized autochrome plates, and 6DoF pose estimation from RGB and event data. Using these examples, we analyze different data regimes and training strategies that combine synthetic and real data across these paradigms. We conclude that curated synthetic data are best understood as a complement to real observations, and that hybrid pipelines combining controllable supervision with learned appearance are the most promising direction for reliable sim-to-real transfer.
In 3D graphics, texture mapping is an essential technique. Over the past three decades, various texture codecs have been developed to efficiently compress data, balancing compression quality, ratio, speed, and flexibility. Among them, ASTC (Adaptive Scalable Texture Compression) is widely adopted on mobile devices. Its high flexibility brings structural complexity, requiring careful consideration during encoding. This study analyzes the main stages, characteristics, and design methodology of ASTC encoding. We discuss key techniques, including partitioning algorithms, color encoding, and block size trade-offs, and how to balance encoding quality and speed. Furthermore, we experimentally evaluate quality and encoding speed by varying the encoder, block size, and modes. Finally, we provide a structured overview of representative studies on ASTC. This work aims to enhance understanding of ASTC among app developers and graphics researchers to facilitate its more effective use.
Open standards for representing and exchanging interactive 3D scenes and data is a critical capability needed to publish visualization and graphics to the web. X3D, the successor to VRML, is a leading provider of such capabilities, enabling integration with web services, supporting cross-platform execution, and facilitating inter-application data exchange. Building on the capabilities of X3D and other open data standards, the addition of cross-platform, shareable execution modules implemented via generalized programming languages provides additional powerful web deployment capabilities. To this end, we have developed execution models based on the emerging open standards WebAssembly (WASM) and WebGPU. WASM transforms complex code, in our case C++, into portable, secure, and performant modules supported by web browsers; while WebGPU enables cross-platform graphics and processing capabilities. In this paper we describe how we use WASM and WebGPU to build cross-platform web applications using the C++ VTK toolkit as the underlying graphics and visualization engine.
Recent advances in 3D Gaussian Splatting (3DGS) have enabled high-resolution real-time rendering; however, the unstructured nature of 3DGS complicates dynamic deformation. Without inherent topology, 3DGS often suffers from structural tearing and visual artifacts during manipulation. We present APG-GS, a framework enabling semantically selective deformation while preventing non-physical interference between adjacent objects. To establish reliable topology, we construct a Gaussian Similarity Graph using a multi-attribute metric that integrates spatial distance, orientation, scale, and Spherical Harmonics. A Point-consistent ChainMail algorithm then propagates user-driven displacements through the graph to ensure interactive responsiveness. Furthermore, we maintain geometric consistency by estimating local deformation gradients, refined via polar decomposition, to adapt Gaussian covariance matrices in real time. Experimental results demonstrate that APG-GS effectively reduces visual artifacts and preserves structural coherence without sacrificing rendering efficiency. This approach extends 3DGS from static reconstruction to interactive, user-driven dynamic applications.
