edge-regularized splatting

To address the imprecise surface reconstruction and compromised rendering realism of 3D Gaussian Splatting (3DGS) on complex geometries, this work introduces normal supervision into the differentiable 3DGS framework for the first time. We propose a signed distance function (SDF)-guided normal regularization method that jointly optimizes photometric fidelity and geometric accuracy. By unifying photometric consistency and explicit surface geometric constraints, our approach enables high-fidelity mesh extraction and real-time novel-view synthesis. Evaluated on the Mip-NeRF360 and Tanks and Temples benchmarks, our method surpasses existing mesh-based rendering approaches in photometric metrics (e.g., PSNR, SSIM), while preserving high-geometric-fidelity reconstructed meshes. This establishes a new paradigm that simultaneously delivers high-quality rendering and geometric usability—benefiting downstream applications such as video generation, AR, and VR.

本文针对3D高斯点云渲染中的梯度消失问题，提出了一种分段截断梯度更新方法，增强了优化稳定性和鲁棒性，并适用于静态和动态场景。

This work addresses depth bias and geometric holes in 2D Gaussian Splatting (2DGS) reconstructions of glossy surfaces, caused by reflectance discontinuities. To this end, we propose an unbiased depth modeling framework. Our method introduces three key innovations: (1) a depth convergence loss that replaces conventional depth distortion loss, effectively mitigating depth estimation bias at specular–diffuse transition regions; (2) a redesigned ray-Gaussian intersection criterion, incorporating ray-level multi-Gaussian depth weighting to enable full-intersection deep fusion; and (3) a reflection-aware optimization strategy to enhance geometric completeness in glossy regions. Evaluated on multiple standard benchmarks, our approach eliminates geometric holes significantly, reduces Chamfer distance by 32% on average, and improves both PSNR and SSIM. It achieves superior geometric fidelity and rendering quality compared to the original 2DGS.

To address inaccurate geometry reconstruction and surface artifacts in 3D Gaussian Splatting (3DGS) under multi-view scenes with significant color discrepancies, this paper proposes a multi-view stereo (MVS)-guided joint geometry-appearance optimization framework. Methodologically, it introduces the first MVS-driven Gaussian initialization strategy; incorporates a median depth relative loss with uncertainty-aware weighting; and jointly enforces MVS depth, normal, and RGB consistency constraints to model geometric complementarity between MVS priors and Gaussian optimization. Experiments across diverse indoor and outdoor scenes demonstrate substantial improvements in surface completeness and boundary sharpness—reducing Chamfer distance by 21.3%—while preserving high-fidelity rendering quality (surpassing state-of-the-art methods in PSNR and SSIM). This work establishes the first unified end-to-end co-optimization framework integrating MVS priors with 3DGS, enabling synergistic geometric refinement and appearance learning.

In sparse-view settings, 3D Gaussian Splatting (3DGS) suffers from low geometric reconstruction accuracy: existing methods rely on non-local depth regularization, limiting fine-grained structural modeling, while conventional smoothing strategies ignore semantic boundaries, degrading critical edges and textures. To address this, we propose a depth- and edge-aware multi-level regularization framework. Our approach innovatively integrates hierarchical depth supervision, Canny-edge-guided semantic masking, and RGB-guided total variation loss—jointly suppressing depth noise while explicitly preserving structural boundaries and high-frequency details. Evaluated on multiple sparse-view novel view synthesis benchmarks, our method significantly improves geometric consistency and visual fidelity, particularly around complex structures and object boundaries, outperforming current state-of-the-art approaches.

This study addresses the geometric distortions and visual artifacts in 3D Gaussian Splatting (3DGS) caused by sparse point cloud initialization in complex scenes. To overcome the limitations of conventional Structure-from-Motion (SfM) sparse initialization, this work proposes a regularized optimization framework that integrates surface normals with dense depth priors. By aligning Gaussian covariances with local geometric structures, the proposed method effectively guides the 3DGS optimization process. Experimental results demonstrate that our approach significantly enhances both geometric accuracy and rendering quality in highly reflective, complex environments such as street scenes. Ultimately, this framework achieves reliable novel view synthesis that simultaneously preserves high fidelity and real-time performance.

This study addresses the issue of floating artifacts in 3D Gaussian Splatting, which arise when occluded or weakly contributing Gaussians drift due to rasterization losses. To mitigate this, we propose a geometry-aware regularization framework based on differentiable local Principal Component Analysis (PCA). Moving beyond conventional view-dependent approaches, our method directly optimizes the neighborhood structure of Gaussian centers by constraining PCA eigenvalues and enforcing normal alignment, thereby achieving geometric correction even for Gaussians lacking gradient contributions. Experimental results demonstrate that the proposed approach significantly suppresses floating noise and enhances surface reconstruction quality. Furthermore, it effectively facilitates downstream tasks such as scene segmentation and Poisson surface reconstruction.

This work addresses the limitations of feedforward models in 3D Gaussian splatting under sparse-view settings—namely, insufficient reconstruction accuracy and the high computational cost of per-scene optimization—by proposing an end-to-end differentiable framework. The approach integrates a 3D Gaussian splatting optimization layer into the training pipeline, enabling the network to learn high-quality initializations that facilitate efficient test-time refinement. An uncertainty-aware mechanism is introduced to adaptively control parameter updates, while efficient backpropagation is achieved through an innovative combination of the implicit function theorem and a matrix-free preconditioned conjugate gradient (PCG) solver. A data-driven uncertainty model further enhances robustness. The method significantly improves reconstruction quality without sacrificing feedforward inference speed and effectively mitigates overfitting in under-constrained regions, performing well in both known and unknown camera pose scenarios.

This work addresses the challenge of effectively leveraging monocular depth priors to enhance geometric accuracy and rendering quality in Gaussian splatting when precise depth data are unavailable. The authors propose a weakly supervised training framework that incorporates scale-ambiguous and noisy monocular depth maps as priors. By analyzing geometric consistency, the method identifies ill-posed regions and applies selective depth regularization only within these areas, thereby preventing erroneous depth from corrupting well-constrained structures. Integrated with a scale-alignment strategy and off-the-shelf depth estimators, this approach seamlessly fits into existing Gaussian splatting pipelines. Experiments across multiple datasets demonstrate significant improvements in both geometry and rendering fidelity, while maintaining compatibility with various Gaussian splatting variants and depth backbone networks.

本文针对3D高斯点云方法在反射区域的表面塌陷问题，提出了一种基于物理的延迟渲染框架RGS，通过几何连续性学习和反射感知密集化策略来改善新视角合成的质量。