Computer Graphics Forum最新文献

3D shape matching plays a fundamental role in applications such as texture transfer and 3D animation. A key requirement for many scenarios is that matchings exhibit geometric consistency, which ensures that matchings preserve neighbourhood relations across shapes. Despite the importance of geometric consistency, few existing methods explicitly address it, and those that do are either local optimisation methods requiring accurate initialisation, or are severely limited in terms of shape resolution, handling shapes with only up to 3,000 triangles. In this work, we present a scalable approach for geometrically consistent 3D shape matching that, for the first time, scales to high-resolution meshes with up to 10,000 triangles. Our method follows a two-stage procedure: (i) we compute a globally optimal and geometrically consistent mapping of surface patches on the source shape to the target shape via a novel integer linear programming formulation. (ii) we find geometrically consistent matchings of corresponding surface patches which respect correspondences of boundaries of patches obtained from stage (i). With this, we obtain dense, smooth, and guaranteed geometrically consistent correspondences between high-resolution shapes. Empirical evaluations demonstrate that our method is scalable and produces high-quality, geometrically consistent correspondences across a wide range of challenging shapes. Our code is publicly available: https://github.com/NafieAmrani/SuPa-Match.

Reconstructing three-dimensional shapes from point clouds remains a central challenge in geometry processing, particularly due to the inherent uncertainties in real-world data acquisition. In this work, we introduce a novel Bayesian framework that explicitly models and propagates uncertainty from both input points and their estimated normals. Our method incorporates the uncertainty of normals derived via Principal Component Analysis (PCA) from noisy input points. Building upon the Smooth Signed Distance (SSD) reconstruction algorithm, we integrate a smoothness prior based on the curvatures of the resulting implicit function following Gaussian behavior. Our method reconstructs a shape represented as a distribution, from which sampling and statistical queries regarding the shape's properties are possible. Additionally, because of the high cost of computing the variance of the resulting distribution, we develop efficient techniques for variance computation. Our approach thus combines two common steps of the geometry processing pipeline, normal estimation and surface reconstruction, while computing the uncertainty of the output of each of these steps.

We present a straightforward and robust method for resolving the mesh intersection problem. We focus specifically on the challenge caused by the intersections resulting from the conversion of the vertices coordinates from their exact mathematical values to a fixed-precision floating-point format. Our method takes as input a soup of triangles and outputs intersection-free models whose vertices coordinates are all represented with double-precision floating-point format. We evaluated our approach thoroughly, considering a large collection of meshes. In particular, we can process all the 4524 models in Thingi10K [ZJ16] that contain self-intersections. This outperforms previous state-of-the-art approaches: On the 527 models of Thingi10K for which naive rounding fails, Zhou et al.'s approach [ZGZJ16] is capable of handling 91% of them, and Valque's 94% [Val24]. In terms of time efficiency, our approach handles about 50k vertices per second on average, which is faster to that of Zhou et al. by a factor 1.4 on these non-trivial models and is faster than that of Valque by several order of magnitude.

The generalized winding number is an essential part of the geometry processing toolkit, allowing to quantify how much a given point is inside a surface, even when the surface has boundaries and noise. We propose a new universal method to compute a generalized winding number, based only on the surface boundary and the intersections of a single ray with the surface, supporting any oriented surface representations that support a ray intersection query. Due to the focus on the boundary, our algorithm has a unique set of properties. For 2D parametric curves, on a regular grid of query points, our method is up to 4× faster than the current state of the art, maintaining the same precision. In 3D, our method can compute a winding number of a surface without discretizing it, including parametric surfaces. For some meshes with many triangles and a simple boundary, our method is faster than the hierarchical evaluation of the generalized winding number while still being precise. Similarly, on some parametric surfaces with a simple boundary, our method can be faster than adaptive quadrature. We validate our algorithms theoretically, numerically, and by demonstrating a gallery of results on a variety of parametric surfaces and meshes, as well uses in a variety of applications, including voxelizations and boolean operations.

For reconstructing high-fidelity human 3D models from monocular videos, it is crucial to maintain consistent large-scale body shapes along with finely matched subtle wrinkles. This paper explores how per-frame rendering results can be factorized into a pose-independent component and a corresponding pose-dependent counterpart to facilitate frame consistency at multiple scales. Pose adaptive texture features are further improved by restricting the frequency bands of these two components. Pose-independent outputs are expected to be low-frequency, while high-frequency information is linked to pose-dependent factors. We implement this with a dual-branch network. The first branch takes coordinates in the canonical space as input, while the second one additionally considers features outputted by the first branch and pose information of each frame. A final network integrates the information predicted by both branches and utilizes volume rendering to generate photo-realistic 3D human images. Through experiments, we demonstrate that our method consistently surpasses all state-of-the-art methods in preserving high-frequency details and ensuring consistent body contours. Our code is accessible at https://github.com/ChunjinSong/facavatar.

This paper studies the classical problem of 3D shape alignment, namely computing the relative rotation between two shapes (centered at the origin and normalized by scale) by aligning spherical harmonic coefficients of their spherical function representations. Unlike most prior work, which focuses on the regime in which the inputs have approximately the same shape, we focus on the more general and challenging setting in which the shapes may differ. Central to our approach is a stability analysis of spherical harmonic coefficients, which sheds light on how to align them for robust rotation estimation. We observe that due to symmetries, certain spherical harmonic coefficients may vanish. As a result, using a robust norm for alignment that automatically discards such coefficients offers more accurate rotation estimates than the widely used L2 norm. To enable efficient continuous optimization, we show how to analytically compute the Jacobian of spherical harmonic coefficients with respect to rotations. We also introduce an efficient approach for rotation initialization that requires only a sparse set of rotation samples. Experimental results show that our approach achieves better accuracy and efficiency compared to baseline approaches.

Diffusion models have emerged as a popular method for 3D generation. However, it is still challenging for diffusion models to efficiently generate diverse and high-quality 3D shapes. In this paper, we introduce OctFusion, which can generate 3D shapes with arbitrary resolutions in 2.5 seconds on a single Nvidia 4090 GPU, and the extracted meshes are guaranteed to be continuous and manifold. The key components of OctFusion are the octree-based latent representation and the accompanying diffusion models. The representation combines the benefits of both implicit neural representations and explicit spatial octrees and is learned with an octree-based variational autoencoder. The proposed diffusion model is a unified multi-scale U-Net that enables weights and computation sharing across different octree levels and avoids the complexity of widely used cascaded diffusion schemes. We verify the effectiveness of OctFusion on the ShapeNet and Objaverse datasets and achieve state-of-the-art performances on shape generation tasks. We demonstrate that OctFusion is extendable and flexible by generating high-quality color fields for textured mesh generation and high-quality 3D shapes conditioned on text prompts, sketches, or category labels. Our code and pre-trained models are available at https://github.com/octree-nn/octfusion.

We present the first combinatorial algorithm for efficiently computing the Reeb space in all dimensions. The Reeb space is a higher-dimensional generalization of the Reeb graph, which is standard practice in the analysis of scalar fields, along with other computational topology tools such as persistent homology and the Morse-Smale complex. One significant limitation of topological tools for scalar fields is that data often involves multiple variables, where joint analysis is more insightful. Generalizing topological data structures to multivariate data has proven challenging and the Reeb space is one of the few available options. However, none of the existing algorithms can efficiently compute the Reeb space in arbitrary dimensions and there are no available implementations which are robust with respect to numerical errors. We propose a new algorithm for computing the Reeb space of a generic piecewise linear map over a simplicial mesh of any dimension called arrange and traverse. We implement a robust specialization of our algorithm for tetrahedral meshes and evaluate it on real-life data.

3D Gaussian Splatting has markedly impacted neural rendering by achieving impressive fidelity and performance. Despite this achievement, it is not readily applicable to developing interactive applications. Real-time applications like XR apps and games require functions such as animation, UV mapping and level of detail (LOD) simultaneously manipulated through a 3D model. To address this need, we propose a modelling strategy analogous to typical 3D models, which we call 3D Gaussian Model (3DGM). 3DGM relies on attaching 3D Gaussians on the triangles of a mesh proxy, and the key idea is to bind sheared 3D Gaussians in texture space and re-projecting them back to world space through implicit shell mapping; this design naturally enables deformation and UV mapping via the proxy. Further, to optimize speed and fidelity based on different viewing distances, each triangle can be tessellated to change the number of involved 3D Gaussians adaptively. Application-wise, we will show that our proxy-based 3DGM is capable of enabling novel deformation without animated training data, texture transferring via UV mapping of the 3D Gaussians, and LOD rendering. The results indicate that our model achieves better fidelity for deformation and better optimization of fidelity and performance given different viewing distances. Further, we believe the results indicate the potential of our work for enabling interactive applications for 3D Gaussian Splatting.

As smartwatches become increasingly prevalent, their built-in sensors provide a rich source for gathering various personal data, including physical activity and health metrics. We found that different brands and models use various visualization techniques. However, the effectiveness of these visualizations within the limited display space of smartwatches remains unclear. Therefore, this paper compares four popular visualizations—bar charts, radial bar charts, donut charts and multi-donut charts—used for displaying activity data on smartwatches. The evaluation focuses on their performance in three common user tasks: counting completed goals, estimating completion percentage and estimating exercise duration. Additionally, the study investigates the impact of the positioning of the target data item, within these visualizations on user performance. Our results indicate that bar charts are superior in terms of task completion time across all tasks. Radial bar charts and multi-donut charts are most effective in helping users perceive the completion ratio (percentage) of each activity and understand the time taken for each activity metric (in minutes). Interestingly, we found that the positioning of data items within the visualizations significantly influences user performance in many cases. Furthermore, it was noted that the visualizations users favoured the most were generally those that enabled them to achieve the highest accuracy in task completion. These insights provide valuable guidelines for future designs in visualizing exercise data on smartwatches. Supplementary material is available at https://osf.io/5u2ph/.