arXiv:2610.35347v1 Announce Type: new Abstract: Voronoi tessellations have recently emerged as an adaptive, space-partitioning discretization for shape and appearance reconstruction from multi-view i...
LoBE-GS is a Load‑Balanced and Efficient 3D Gaussian Splatting framework designed for large‑scale scene reconstruction. It introduces a load‑balanced KD‑tree partitioning scheme that balances camera counts per block, uses depth‑based back‑projection to speed up camera assignment, and applies visibility cropping and selective densification to cut training costs. Experiments on urban and outdoor datasets show up to an order‑of‑magnitude faster scene‑partitioning and a 3.2× reduction in worst‑block fine‑training time compared to CityGaussian, while keeping reconstruction quality competitive.
CoMVS‑GS is a surface‑reconstruction framework that fuses Multi‑View Stereo (MVS) with 3D Gaussian splatting. It initializes Gaussian primitives from dense MVS points, uses PatchMatch‑3DGS mutual supervision to refine depths and normals, and replaces voxel‑based meshing with a Delaunay graph‑cut pipeline. Experiments on DTU, GauU‑Scene V2, and MatrixCity demonstrate competitive object‑level results and improved geometric accuracy and mesh compactness in outdoor scenes while preserving high rendering quality.
VolFill is a generative framework that reconstructs the complete 3D geometry of a scene from a single RGB image, addressing the challenge of inferring hidden structures. It uses a hybrid 3D VAE to encode sparse truncated unsigned distance function grids into a latent space, and a latent Diffusion Transformer to denoise and recover the full scene. By conditioning on a geometry foundation model, VolFill leverages spatial priors and provides a structured representation that enables direct surface extraction and scalable occupancy queries, outperforming existing amodal baselines on SCRREAM and NRGB-D datasets.
Scaling 3D Gaussian Splatting (3DGS) to large outdoor scenes is costly in both data acquisition and computation. Adopting panoramic images with equirectangular projection (ERP) can reduce capture effort via their full $360^{\circ}$ field of view, yet the resulting omnipresent visibility invalidates existing partitioning strategies that rely on local camera frustums, causing block-wise optimization to degenerate into global training.
arXiv:2608.20687v1 Announce Type: new Abstract: 3D Gaussian Splatting has achieved remarkable success in novel view synthesis. However, extracting high-fidelity surfaces directly from 3DGS remains ch...
The paper introduces SVRecon, a generalizable neural surface reconstruction framework that uses sparse volumetric representations to achieve high-resolution 3D reconstruction. It employs a two-stage architecture: first predicting occupied voxels with an occupancy network, then rendering only within those regions using specialized sparse algorithms. This approach allows reconstruction at resolutions up to 512³ on 32 GB hardware, producing smoother and more precise surfaces, especially in sparse-view scenarios.
PCAsplat introduces a geometry‑aware regularization framework for Gaussian splatting that uses differentiable local PCA to update Gaussians even when they are occluded or contribute little to sampled camera rays. By regularizing PCA eigenvalues and aligning Gaussian normals with neighborhood normals, the method encourages Gaussians to move onto the underlying surface with isotropic tangent‑plane coverage, thereby reducing floaters. Experiments on DTU, Tanks and Temples, and NeRF Synthetic demonstrate that PCAsplat produces surface‑aligned splats that improve geometry‑processing tasks such as point cloud segmentation and Poisson reconstruction while remaining competitive in novel view synthesis and mesh extraction.
STARS-GS is a new structure‑aware 3D Gaussian Splatting framework designed for large‑scale aerial surface reconstruction. It introduces a scene partitioning strategy that preserves continuous scene elements, a neighborhood‑aware Gaussian organization that extends geometric constraints to local neighborhoods, and an adaptive surface regularization that tailors regularization strength to local geometry. Experiments on aerial photogrammetry benchmarks show that STARS‑GS improves the average F1‑score from 0.640 to 0.698, a relative gain of about 9.1%.
ZipMVS is a multi‑view stereo (MVS) approach that focuses on reducing memory usage while maintaining high‑quality 3D reconstructions. It introduces a new depth‑hypothesis strategy that compresses the cost volume, significantly lowering GPU memory consumption. Experiments on the DTU and Tanks and Temples datasets demonstrate that ZipMVS delivers competitive reconstruction quality compared to other efficiency‑oriented MVS methods.
The paper introduces VAD, a lightweight, network‑free method for computing Unsigned Distance Fields (UDFs) from unoriented point clouds. It assigns bi‑directional normals using two Voronoi‑based criteria, diffuses these normals to approximate a UDF gradient field, and then integrates to recover the final UDF. Experiments show VAD handles watertight, open, non‑manifold, and non‑orientable geometries efficiently and stably.
PointGauss is a 3D-native framework that performs semantic parsing and instance segmentation on 3D Gaussian splatting representations by treating Gaussian primitives as unstructured point sets and extracting scale‑invariant geometric features with Point Transformer V3. It introduces an adaptive region‑of‑interest cropping strategy and an instance‑aware distance‑constrained rasterization pipeline to enable scalable, view‑consistent pixel‑level projections. The authors also release SplatSeg‑360, a cross‑scale benchmark with 32 complex scenes and over 6,300 aligned 2D‑3D masks, and show that PointGauss achieves real‑time performance with state‑of‑the‑art 3D‑mIoU (~90%) and 2D‑mIoU (~80%) scores.