HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction
Extracted source text · arXiv 2506.14229
\documentclass{article}
\usepackage[preprint]{neurips_2025} % 改为 [preprint] 或删去 final 即可匿名
\usepackage[table]{xcolor} % 必须启用 table 选项
\usepackage{colortbl} % 支持表格着色
% \input{macro}
\usepackage{graphicx} % 支持\scalebox
\usepackage{caption} % 支持\caption*
\usepackage{array} % 优化表格格式
\usepackage{booktabs} % 优化表格横线
\usepackage{xcolor}
\usepackage{placeins} % 修复：删除后面的zhong
\usepackage{amsmath} % 新增：数学公式
\usepackage{amsfonts} % 新增：数学字体
\usepackage{amssymb} % 新增：数学符号
\usepackage{multirow} % 新增：表格合并
\usepackage{enumitem}
\usepackage{wrapfig}
\newcommand{\vs}{vs.}
\newcommand{\chl}[1]{{\color{blue} {Hanlin: #1}}}
\newcommand{\zhd}[1]{{\color{orange} {Haodong: #1}}}
\title{
{HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction}
% AttriGLIP: A zero-shot medical detector via hierarchical knowledge bank
}
% \author{
% Changbai Li\thanks{Equal contribution.}\textsuperscript{1} \quad
% Haodong Zhu\footnotemark[1]\textsuperscript{1} \quad
% Tongfei Chen\textsuperscript{1} \quad
% Shuo Yang\textsuperscript{1} \quad
% Shuwei Shao\textsuperscript{1} \quad
% Wenhao Dong\textsuperscript{1} \quad
% Hanlin Chen\thanks{Corresponding author.}\textsuperscript{2} \quad
% Juan Zhang\footnotemark[2]\textsuperscript{1} \quad
% Baochang Zhang\textsuperscript{1} \\[1em]
% \textsuperscript{1}Beihang University \\
% \textsuperscript{2}National University of Singapore
% }
\author{
Changbai Li\thanks{Equal contribution.} \\
School of Artificial Intelligence \\
Beihang University \\
\texttt{changboli@buaa.edu.cn}
\And
Haodong Zhu\footnotemark[1] \\
School of Artificial Intelligence \\
Beihang University \\
\texttt{HaodongZhu@buaa.edu.cn}
\And
Hanlin Chen\thanks{Corresponding author.} \\
School of Computing \\
National University of Singapore \\
\texttt{hanlin.chen@u.nus.edu}
\And
Juan Zhang\footnotemark[2] \\
School of Artificial Intelligence \\
Beihang University \\
\texttt{zhang\_juan@buaa.edu.cn}
\And
Tongfei Chen \\
School of Artificial Intelligence \\
Beihang University \\
\texttt{tfchen@buaa.edu.cn}
\And
Shuo Yang \\
School of Automation Science and Electrical Engineering \\
Beihang University \\
\texttt{shuo1yang@buaa.edu.cn}
\And
Shuwei Shao \\
School of Automation Science and Electrical Engineering \\
Beihang University \\
\texttt{swshao@buaa.edu.cn}
\And
Wenhao Dong \\
School of Astronautics \\
Beihang University \\
\texttt{ZB2315207@buaa.edu.cn}
\And
Baochang Zhang \\
School of Artificial Intelligence \\
Beihang University \\
\texttt{bczhang@buaa.edu.cn}
}
\begin{document}
\maketitle
\begin{figure*}[h]
\centering
% \hspace*{-0.2cm} % 核心负向位移指令
% \vspace{1.0cm}
\includegraphics[width=1.0\linewidth]{fig1.pdf}
\caption{(a) High-resolution ($\sim$ 5K) renderings of the “bicycle” scene from the Mip-NeRF 360 dataset, with results from 3DGS, Mip-Splatting, and our method. Red dashed boxes highlight key details. (b) Performance on this scene: our method achieves the highest PSNR (25.6 dB) with significantly lower GPU memory (23 GB) and model size (567 MB) than 3DGS and Mip-Splatting. (c) Average resource usage across the full dataset shows our method maintains the smallest memory and model footprint.}
%
% \caption{{\color{red} (a) High-resolution (5K) training renderings of the “bicycle” scene from the Mip-NeRF 360 dataset alongside the reconstructions produced by 3DGS, Mip-Splatting and our method, with red dashed boxes highlighting key detail regions. (b) Integrated performance comparison on this single scene, showing average PSNR, VRAM consumption and model size. Our approach achieves the highest PSNR of 25.6 dB while requiring only 28 GB of GPU memory and a 567 MB model, substantially lower than 3DGS and Mip-Splatting. (c) Average resource consumption over full-resolution reconstructions across the entire dataset, comparing model size and GPU memory usage, where our method yields the smallest footprint relative to 3DGS and Mip-Splatting.}}
%
\label{fig:top}
\end{figure*}
\input{secs/abstract}
\input{secs/introduction}
\input{secs/related}
\input{secs/methodology}
\input{secs/experiment}
\input{secs/conclusion}
\newpage
{\small
\bibliographystyle{ieee_fullname}
\bibliography{reference}
}
\newpage
\input{secs/appendix}
% \newpage
% \input{Supplementary_Material}
\newpage
% \input{secs/checklist}
\end{document}
% \begin{figure*}[t]
% \centering
% \hspace*{-0.2cm} % 核心负向位移指令
% \includegraphics[width=1.0\linewidth]{top.png}
% \caption{{\color{red}This figure illustrates the full-resolution training renderings of the ``bicycle'' scene in the mip-nerf 360 dataset, as well as the average GPU memory consumption and model sizes of 3DGS, Mip-Splatting, and the proposed method. The lower part includes two charts: the left-hand chart compares the model sizes, and the right-hand chart illustrates the GPU memory consumption, showing that Mip-Splatting has the highest usage and Ours has the lowest.}}
% \label{fig:top}
% \end{figure*}
\begin{abstract}
% 3D Gaussian Splatting (3DGS) has significantly driven progress in real-time 3D scene reconstruction. However, its application in high-resolution reconstruction scenarios encounters critical memory scalability limitations. To address this, we propose Hierarchically Refined Gaussian Splatting (HRGS), a memory-efficient framework with hierarchical block optimization from coarse to fine. Specifically, we first derive a coarse global Gaussian representation from low-resolution data, then partition the scene into blocks and refine each block using high-resolution data. The scene partitioning includes two procedures: Gaussian partitioning and training data partitioning. For Gaussian partitioning, we contract irregular scenes into normalized bounded cubic shapes and employ uniform grid division to distribute computational tasks evenly among various blocks. For training data partitioning, only the observations that reside within their corresponding blocks or significantly contribute to the rendering results are retained. Alignment of trained Gaussians from adjacent blocks is achieved by guiding each block's refinement with the coarse global Gaussian prior, ensuring seamless fusion. To further enhance surface reconstruction quality, we incorporate normal priors from a pretrained model. Finally, our method enables high-quality and high-resolution 3D scene reconstruction even under constrained memory capacities. Extensive experimental results on three public benchmarks show that our approach achieves state-of-the-art performance in 3D reconstruction, covering both high-resolution novel view systhesis (NVS) and surface reconstruction.
3D Gaussian Splatting (3DGS) has achieved significant progress in real-time 3D scene reconstruction. However, its application in high-resolution reconstruction scenarios faces severe memory scalability bottlenecks. To address this issue, we propose Hierarchical Gaussian Splatting (HRGS), a memory-efficient framework with hierarchical block-level optimization from coarse to fine. Specifically, we first derive a global, coarse Gaussian representation from low-resolution data; we then partition the scene into multiple blocks and refine each block using high-resolution data. Scene partitioning comprises two steps: Gaussian partitioning and training data partitioning. In Gaussian partitioning, we contract irregular scenes into a normalized, bounded cubic space and employ a uniform grid to evenly distribute computational tasks among blocks; in training data partitioning, we retain only those observations that lie within their corresponding blocks or make significant contributions to the rendering results. By guiding each block’s refinement with the global coarse Gaussian prior, we ensure alignment and seamless fusion of Gaussians across adjacent blocks. To reduce computational resource demands, we introduce an Importance-Driven Gaussian Pruning (IDGP) strategy: during each block’s refinement, we compute an importance score for every Gaussian primitive and remove those with minimal rendering contribution, thereby accelerating convergence and reducing redundant computation and memory overhead. To further enhance surface reconstruction quality, we also incorporate normal priors from a pretrained model. Finally, even under memory-constrained conditions, our method enables high-quality, high-resolution 3D scene reconstruction. Extensive experiments on three public benchmarks demonstrate that our approach achieves state-of-the-art performance in high-resolution novel view synthesis (NVS) and surface reconstruction tasks.
\end{abstract}
\section{Introduction}
3D scence reconstruction remains a longstanding challenge in computer vision and graphics. A significant advancement in this domain is the Neural Radiance Field (NeRF) \cite{nerf}, which effectively represents geometry and view-dependent appearance using multi-layer perceptrons (MLPs), demonstrating significant advancements in 3D reconstruction quality.
Recently, 3D Gaussian Splatting (3DGS)~\cite{3dgs} has gained considerable attention as a compelling alternative to MLP-based~\cite{nerf} and feature grid-based representations~\cite{tensorf,plenoxels,Neural,Instant_neural}. 3DGS stands out for its impressive results in 3D scene reconstruction and novel view synthesis while achieving real-time rendering at 1K resolutions.
This efficiency and effectiveness, combined with the potential integration into the standard GPU rasterization pipeline, marks a significant step toward the practical adoption of 3D reconstruction methods.
% \begin{figure*}[t]
% \centering
% \hspace*{-0.2cm} % 核心负向位移指令
% \includegraphics[width=1.0\linewidth]{top.png}
% \caption{{\color{red}This figure illustrates the full-resolution training renderings of the ``bicycle'' scene in the mip-nerf 360 dataset, as well as the average GPU memory consumption and model sizes of 3DGS, Mip-Splatting, and the proposed method. The lower part includes two charts: the left-hand chart compares the model sizes, and the right-hand chart illustrates the GPU memory consumption, showing that Mip-Splatting has the highest usage and Ours has the lowest.}}
% \label{fig:top}
% \end{figure*}
In particular, 3DGS models complex scenes as a collection of 3D Gaussian distributions, which are projected onto screen space using splatting-based rasterization. The characteristics of each 3D Gaussian, including position, size, orientation, opacity, and color, are optimized using multi-view photometric loss.
%Following this, a 2D dilation process is applied in screen space for low-pass filtering.
Although 3DGS has demonstrated impressive 3D reconstruction results, its application in high-resolution scenarios encounters critical memory scalability limitations.
Specifically,when reconstructing outdoor scenes at ultra-high resolutions approaching 5K (e.g.4978\,$\times$\,3300 pixels) in standardized benchmark datasets like Mip-NeRF 360~\cite{Mip-NeRF360}, conventional 3DGS implementations demand excessive VRAM, exceeding the capacity of mainstream GPUs with limited memory, such as the NVIDIA A5000 (24GB VRAM).
%This computational bottleneck originates from the inherent parametric scaling characteristics of 3DGS methodology: The technique encodes scene information through a dense population of anisotropic Gaussian primitives, whose cardinality exhibits superlinear growth relative to input resolution.
%This computational bottleneck originates from the inherent parametric scaling characteristics of 3DGS, which encodes scene information through a dense set of Gaussian primitives whose number grows superlinearly with the input resolution.
%High-fidelity scene reconstruction necessitates exponential expansion of Gaussian distributions to capture fine-grained geometric details and complex view-dependent radiance variations, thereby inducing quadratic memory complexity in parameter storage.
% \textcolor{red}{
This computational bottleneck arises from the increasing resolution: higher resolutions demand more GPU memory, as illustrated in Fig.~\ref{fig:top}.
% }
Such algorithmic behavior fundamentally conflicts with finite GPU memory resources, resulting in catastrophic memory overflow during optimization phases.
% To overcome these critical memory constraints while preserving reconstruction fidelity for high-resolution rendering, we present XXGaussian (HRGS), a memory-efficient framework with hierarchical block optimization from coarse to fine.
To overcome these critical memory constraints while preserving reconstruction fidelity for high-resolution scene reconstruction, we present Hierarchically Gaussian Splatting (HRGS), a memory-efficient framework with hierarchical block optimization from coarse to fine.
Specifically, we first obtain a coarse global Gaussian representation using low-resolution images.
Subsequently, to minimize memory usage on a single GPU, we partition the scene into spatially adjacent blocks and parallelly refined.
Each block is represented with fewer Gaussians and trained on reduced data, allowing further optimization with high-resolution images. The partitioning strategy operates at two levels: Gaussian primitives and training data.
To achieve a more balanced partition of Gaussians and avoid blocks with sparse Gaussians, we begin by contracting unbounded Gaussians. In detail, we define a bounded cubic region and use its boundary to normalize the Gaussian positions. Within this region, Gaussians are contracted via a linear mapping, while those outside undergo nonlinear contraction, yielding a more compact Gaussian representation. We then apply a uniform grid subdivision strategy to this contracted space, ensuring an even distribution of computational tasks.
%For Gaussian partitioning, we contract irregular scenes into normalized bounded cubic shapes and employ uniform grid division to distribute computational tasks evenly among various blocks.
%During training data partitioning，为了挑选出与要refine块强相关的数据，我们利用该pose分别利用全局Gaussian和全局gaussian去掉这一block部分的Gaussian渲染出的图片计算SSIM loss，只保留对这块贡献较大的pose
%During training data partitioning, to select data highly relevant to the block under refinement, we render images for each pose using the full Gaussian representation and again with the block’s Gaussians removed, then compute the SSIM loss~\cite{mss} between the two renders.
% During data partitioning for training, we identify data most relevant to the block under refinement by rendering images for each pose both with the complete global Gaussians and with the block’s Gaussians removed, then computing the SSIM loss~\cite{mss} between the two outputs.
% A higher SSIM loss indicates that the pose contributes more significantly to the block, so we set a threshold on SSIM loss and retain only those poses whose values exceed it.
During data partitioning for training, we compute the SSIM loss~\cite{mss} for each observation by comparing two renderings:One rendering is implemented with the complete global Gaussian representation, while the other is executed subsequent to the elimination of Gaussians within the target block. A more pronounced SSIM loss denotes that the observation exerts a more substantial contribution to the target block, so we set a threshold on SSIM loss and retain only observations whose values exceed it.
%Only those poses that contribute significantly to the block are kept.
To mitigate artifacts at the block boundaries, we further include observations that fall within the region of the considered block.
Finally, to prevent overfitting, we employ a binary search algorithm during data partitioning to expand each block until the number of Gaussians it contains exceeds a specified threshold.
%During training data partitioning, only the poses that reside within their corresponding blocks or significantly contribute to the rendering results are retained.
This innovative strategy effectively reduces interference from irrelevant data while improving fidelity with decreased memory usage, as demonstrated in Tab.~\ref{tab4}.
After partitioning the Gaussian primitives and data, we initialize each block in the original, uncontracted space using the coarse global Gaussian representation. To accelerate convergence and reduce computational overhead during block-level refinement with high-resolution data, we introduce an Importance-Driven Gaussian Pruning (IDGP) strategy. {Specifically, we evaluate the interaction between each Gaussian and the multi-view training rays within the corresponding block, and discard those with negligible rendering contributions. All blocks are then refined in parallel, and subsequently integrated into a unified, high-resolution global Gaussian representation.}
%
% we compute the interaction between each Gaussian and the block’s multi-view training rays, then remove Gaussians with low rendering contributions.
% We refine all blocks in parallel and finally assemble them into a global high-resolution Gaussian representation.
% %
In addition to novel view synthesis, we evaluate our method on another key subtask of 3D reconstruction: 3D surface reconstruction. To further enhance the quality of the reconstructed surfaces, we incorporate the View-Consistent Depth-Normal Regularizer~\cite{vcR}, which is applied both during the initialization of the coarse global Gaussian representation and throughout the subsequent block-level refinement.
%
% To further enhance surface reconstruction quality, we employ the View Consistent Depth-Normal Regularizer~\cite{vcR} to obtain the initial coarse global Gaussian representation and during subsequent block refinements.
%
% After partitioning the Gaussian primitives and data, we initialize each block in the original, uncontracted space using a coarse global Gaussian representation. We then refine each block in parallel with high-resolution data and finally assemble these refined blocks to obtain the global high-resolution Gaussian representation.
%Alignment of trained Gaussians from adjacent blocks is achieved by guiding each block's training with the coarse global Gaussian prior, ensuring seamless fusion.
% To further enhance the quality of surface reconstruction, we adopt the View Consistent Depth-Normal Regularizer from VCR-GauS~\cite{vcR} to obtain the initial coarse global Gaussian representation and subsequent refinement.
Finally, our method enables high-quality and high-resolution scene reconstruction even under constrained memory capacities (e.g., NVIDIA A5000 with 24GB VRAM).
We validate our method on two sub-tasks of 3D reconstruction: high-resolution NVS and surface reconstruction, and demonstrate that it delivers superior high-resolution reconstruction performance. In summary, the main contributions of this paper are:
\begin{itemize}[leftmargin=*]
\item We propose HRGS, a memory-efficient coarse to fine framework that leverages low-resolution global Gaussians to guide high-resolution local Gaussians refinement, enabling high-resolution scene reconstruction with limited GPU memory.
\item A novel partitioning strategy for Gaussian primitives and data is introduced, optimizing memory usage, reducing irrelevant data interference, and enhancing reconstruction fidelity.
\item We propose a novel dynamic pruning strategy, Importance-Driven Gaussian Pruning (IDGP), which evaluates the contribution of each Gaussian primitive during training and selectively removes those with low impact. This approach significantly improves training efficiency and optimizes memory utilization.
%
% \item {\color{red}A novel data importance assessment-based dynamic pruning strategy, termed Importance-Driven Gaussian Pruning (IDGP), is introduced to remove low-contribution Gaussian primitives during the training process, significantly enhancing training efficiency and memory utilization.}
%
\item Extensive experiments on three public datasets demonstrate that our approach achieves state-of-the-art performance in high-resolution rendering and surface reconstruction.
\end{itemize}
\input{images/blank}
% \begin{figure*}[t]
% \centering
% \includegraphics[width=0.5\linewidth]{Fig0.png}
% \caption{Memory usage on single GPU and rendering results of our method and 3DGS at different resolution levels: 1/8, 1/4, 1/2 and full (original) resolution. The figure compares the memory consumption and visual quality of the rendered images at each resolution for both methods on the MipNeRF 360 dataset.
% }
% \end{figure*}
\section{Related Work}
\noindent\textbf{3D Reconstruction.}
Recent 3D reconstruction research can be broadly categorized into traditional geometry-based and deep learning methods. The former relies on multi-view stereo (MVS)~\cite{YAN} and structure from motion (SfM)~\cite{sfm} to estimate scene depth and camera poses, producing point clouds and subsequent surface meshes. The latter integrates implicit functions (e.g., SDF, Occupancy)~\cite{Tri} with volumetric rendering for high-fidelity reconstruction, as exemplified by Neural Radiance Fields (NeRF)~\cite{nerf}. However, NeRF-based approaches often struggle with real-time performance in large-scale or dynamic scenarios. In contrast, 3D Gaussian Splatting~\cite{3dgs} encodes scenes as 3D Gaussians (with position, scale, and color), using differentiable point-based rendering to achieve fast training and inference while balancing accuracy and quality. Balancing high fidelity, scalability, and real-time capability remains a key challenge in 3D reconstruction.
Within the field of 3D reconstruction, there are primarily two main sub-tasks: novel view synthesis (NVS) and surface reconstruction.
\vspace{0.5em}
\noindent\textbf{Novel View Synthesis.}
Novel View Synthesis (NVS) aims to generate a target image from an arbitrary camera pose, given source images and their camera poses~\cite{light, lum}. NeRF~\cite{nerf} integrates implicit representations with volume rendering\cite{volume, eff}, demonstrating impressive results in view synthesis. However, dense point sampling remains a major bottleneck for rendering speed. To address this, various methods accelerate NeRF by replacing the original multi-layer perceptrons (MLPs)~\cite{learining, deep} with discretized representations, such as voxel grids~\cite{dir}, hash encodings~\cite{Instant}, or tensor radiation fields~\cite{tensorf}. Additionally, some approaches~\cite{bake, merf} distill pretrained NeRFs into sparse representations, enabling real-time rendering. Recent advancements in 3D Gaussian Splatting (3DGS) have significantly improved real-time rendering, demonstrating that continuous representations are not strictly necessary. However, directly optimizing and rendering at high resolutions drastically increase memory overhead, making it challenging to achieve real-time reconstruction of high-quality scenes on mainstream GPUs with limited memory (24GB). Our approach specifically addresses this challenge by reducing the computational cost of high-resolution processing while preserving reconstruction fidelity.
\vspace{0.5em}
\textbf{Multi-View Surface Reconstruction.}
In recent years, multi-view reconstruction methods have evolved from traditional geometric approaches to neural implicit representations. Traditional multi-view stereo methods~\cite{re_MVS1,re_MVS2,re_MVS3,re_MVS4,re_MVS5,re_MVS6} primarily rely on extracting dense depth maps~\cite{re_depth1,re_depth2} from multiple images, fusing them into point clouds~\cite{re_point,re_point2}, and generating scene models through triangulation or implicit surface fitting~\cite{re_fit}. Although these techniques have been widely adopted in both academia and industry, they are often susceptible to artifacts and loss of detail due to matching~\cite{re_matching1} errors, noise, and local optimum issues during reconstruction.
Recent advances in neural implicit representations, such as NeRF~\cite{re_nerf} and its SDF-based variants~\cite{NeuS,MonoSDF}, have shifted the reconstruction paradigm from explicit depth estimation toward learning continuous volumetric or surface fields directly from images. These approaches inherently model geometry and appearance jointly, offering better robustness to occlusions and textureless regions.
While neural implicit methods offer high-quality reconstructions, they remain computationally intensive and face scalability challenges. As an alternative, 3D Gaussian Splatting (3DGS)~\cite{3dgs} adopts an explicit representation by projecting anisotropic Gaussians onto image space, enabling efficient and differentiable rasterization~\cite{yifan2019differentiable}. Despite its ability to support real-time rendering with high visual fidelity, 3DGS often suffers from insufficient geometric supervision, particularly in sparse-view or large-scale scenarios~\cite{neusg}. To address this limitation, recent methods such as VCR-GauS~\cite{vcr-gaus}, Vastgaussian~\cite{vast}, and SuGaR~\cite{SuGaR} introduce view-consistent depth and normal constraints~\cite{turkulainen2024dnsplatter,bae2024dsine,yin2019enforcing}, significantly enhancing both reconstruction accuracy and convergence stability. These developments position 3DGS as a promising solution for high-resolution and scalable surface reconstruction.
%With the advancement of deep learning, NeRF~\cite{re_nerf}and its variants have emerged, modeling the color and density in continuous space using multi-layer perceptrons and generating novel views via volume rendering, which significantly enhances rendering quality.Subsequent methods aimed to optimize both quality and speed. Techniques like position encoding and band-limited coordinate networks are combined with neural radiance fields for pre-filtered scene representation~\cite{re_mip,re_bacon}. Innovations to speed up rendering included leveraging spatial data structures and adjusting MLP size~\cite{re_mlpsize1,re_mlpsize2,re_mlpsize3,re_mlpsize4,re_mlpsize5}. However, high-quality rendering with NeRF also comes at the cost of increased computational overhead and longer training times, posing challenges for real-time applications and large-scale scenarios~\cite{Tri-miprf,mega-nerf}. To mitigate these limitations, subsequent research has proposed various acceleration strategies, such as Zip-NeRF~\cite{zip-nrf}, InstantNGP~\cite{InstantNGP}, and Plenoxels~\cite{plenoxels}, which employ techniques like low-pass filtering, multi-resolution hash grids, and sparse voxel grids to substantially boost computational efficiency~\cite{grid-nerf} while preserving rendering quality.
\begin{figure*}[t]
\centering
\centering \includegraphics[width=1.0\textwidth]{fig_main.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Illustrative diagram of the hierarchical block optimization framework.} We first derive a global coarse Gaussian representation using low-resolution data, which is then contracted into a bounded cubic region. Subsequently, the contracted Gaussian primitives are partitioned into blocks, each paired with corresponding data. Leveraging the global coarse Gaussian as initialization, we parallelly refine each block in the original uncontracted space using high-resolution data. During this refinement process, an Importance-Driven Gaussian Pruning strategy is employed to compute the interaction between each Gaussian primitive and training view rays, removing low-contribution primitives to accelerate convergence and reduce redundancy. The optimized blocks are then concatenated to form the final global Gaussian representation, which is validated through novel view synthesis (NVS) and surface reconstruction tasks. }
\label{fig:main}
\vspace{-0.2cm}
\end{figure*}
\section{Method}
% 写一下各个subsect分别干了啥
Our proposed HRGS efficiently reconstructs high-resolution scenes.
We first review 3DGS in Section~\ref{3.1}.
Next, in Section~\ref{3.2}, we present the memory-efficient coarse-to-fine framework, detailing the partitioning of Gaussian primitives and data, along with the proposed Importance-Driven Gaussian Pruning (IDGP) strategy.
%
% Next, in Section~\ref{3.2}, we introduce the memory-efficient coarse-to-fine framework and detail how Gaussian primitives and data are partitioned, {\color{red}as well as the Importance-Driven Gaussian Pruning strategy.}
%
Finally, Section~\ref{3.3} describes the loss function employed in our approach.
\subsection{Preliminary}
\label{3.1}
We begin with a brief overview of 3D Gaussian Splatting (3DGS)~\cite{3dgs}.
In the 3DGS framework, a scene is represented as a set of discrete 3D Gaussian primitives, denoted by \( G_K = \{ G_k \mid k = 1, \dots, K \} \), where \( K \) is the total number of Gaussians in the scene.
Each Gaussian \( G_k \) is defined by a set of learnable parameters, including its 3D position \( \mathbf{p}_k \in \mathbb{R}^{3 \times 1} \), opacity \( \sigma_k \in [0, 1] \), and geometric properties, which typically consist of scaling and rotation parameters that define the Gaussian covariance matrix \( \Sigma_k \in \mathbb{R}^{3 \times 3} \). %i.e. \( \Sigma_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top \mathbf{R}_k^\top \).
Furthermore, spherical harmonic (SH) features \( f_k \in \mathbb{R}^{3 \times 16} \) are used to encode view-dependent color information \( c_k \in \mathbb{R}^{3 \times 1} \), allowing for a realistic depiction of color variations as a function of the viewing angle.
For rendering purposes, the combined color and opacity contributions from multiple Gaussians at a given pixel are weighted according to their respective opacities. The color blending for overlapping Gaussians is computed as follows:
\begin{equation}
\hat{C} = \sum_{k \in M} c_k \alpha_k \prod_{j=1}^{k-1} \left( 1 - \alpha_j \right),
\label{eq1}
\end{equation}
where \( c_k \) and \( \alpha_k = \sigma_kG_k \) denote the color and density of the \( k \)-th Gaussian primitive, respectively.
% \begin{figure*}[htp]
% \centering
% \centering \includegraphics[width=1.0\textwidth]{fig_main.png}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{\textbf{Illustrative diagram of the hierarchical block optimization framework.} We first derive a global coarse Gaussian representation using low-resolution data, which is then contracted into a bounded cubic region. Subsequently, the contracted Gaussian primitives are partitioned into blocks, each paired with corresponding data. Leveraging the global coarse Gaussian as initialization, we parallelly refine each block in the original uncontracted space using high-resolution data. During this refinement process, an Importance-Driven Gaussian Pruning strategy is employed to compute the interaction between each Gaussian primitive and training view rays, removing low-contribution primitives to accelerate convergence and reduce redundancy. The optimized blocks are then concatenated to form the final global Gaussian representation, which is validated through novel view synthesis (NVS) and surface reconstruction tasks. }
% \label{fig:main}
% \vspace{-0.2cm}
% \end{figure*}
\subsection{Hierarchical Block Optimization Framework}
\label{3.2}
Traditional 3D Gaussian methods~\cite{3dgs,vcR} rely on global iterative optimization for scene reconstruction but struggle with memory inefficiency in high-resolution settings, such as the Mip-NeRF 360~\cite{mip} dataset. To address this, we propose a hierarchical optimization framework that balances coarse global representation and fine-grained local refinement, as shown in Fig~\ref{fig:main}. We first construct a low-resolution global Gaussian prior, guiding block-wise high-resolution optimization to enhance geometric detail while maintaining memory efficiency. This approach enables precise reconstruction under constrained memory conditions. The following subsections detail the coarse global Gaussian generation, Gaussian and data partitioning strategies, as well as refinement and post-processing procedures.
\noindent\textbf{Coarse Global Gaussian Representation.}
This stage establishes the foundation for subsequent Gaussian and data partitioning. Initially, we train the COLMAP~\cite{schoenberger2016mvs,schoenberger2016sfm} points using all observations at a low resolution for 30,000 iterations, generating a coarse representation of the global geometric structure. The resulting Gaussian primitives are represented as \( G_K = \{ G_k \mid k = 1, \ldots, K \} \), where \( K \) denotes the total number of Gaussians. In the following block-wise high-resolution refinement, this robust global geometric prior ensures that Gaussians are positioned accurately, thereby preventing drift and eliminating inter-block discontinuities, minimizing significant fusion artifacts.
\noindent\textbf{Primitives and Data Division.}
Directly applying uniform grid division in the original 3D space may lead to uneven Gaussian distribution in local regions (e.g. many nearly empty grid cells alongside overly dense ones).
% \textcolor{red}{
% To avoid this imbalance, the coarse global Gaussian prior is first contracted to a bounded cubic region.
% For contraction, we define a internal region in the central one-third of the space with the minimum and maximum positions of its corners $\mathbf{p}_{\text{min}}$ and $\mathbf{p}_{\text{max}}$.
% The remainder of the space is then designated as the external region.
% To mitigate this imbalance, we first define a bounded cubic region and contract all Gaussians toward it.
% Within this cubic region, central one-third of the space bounded by minimum and maximum positions of its corners $\mathbf{p}_{\text{min}}$ and $\mathbf{p}_{\text{max}}$, is designated as the internal region, while the remainder is classified as the external region.
%The remainder of the space is then designated as the external region.
To address this imbalance, we define a bounded cubic region and contract all Gaussians within it. Within this region, the central one-third of the space is designated as the internal region, while the surrounding area is classified as the external region. The internal region is bounded by the minimum and maximum corner positions, $\mathbf{p}_{\text{min}}$ and $\mathbf{p}_{\text{max}}$, which define the limits of the central one-third of the entire region.
% }
To standardize the representation of global Gaussians, we introduce a normalization step:
\(
\hat{\mathbf{p}}_k = 2\left( {\mathbf{p}_k - \mathbf{p}_{\text{min}}}\right)/\left({\mathbf{p}_{\text{max}} - \mathbf{p}_{\text{min}}} \right)-1.
\)
As a result, the coordinates of Gaussians located in the internal region are constrained within the range $[-1, 1]$.
To achieve more effective contraction of the global Gaussians, we apply a linear mapping for the Gaussians in the internal region, while a nonlinear mapping is employed for the external region (as shown in Fig.~\ref{fig:main}). The final contraction step is performed using the function described in \cite{scanerf}:
\small{
\begin{equation}
% \small
\text{contract} (\hat{\mathbf{p}}_k) =
\begin{cases}
\hat{\mathbf{p}}_k, & \text{if } \|\hat{\mathbf{p}}_k\|_\infty \leq 1, \\
\left( 2 - \frac{1}{\|\hat{\mathbf{p}}_k\|_\infty} \right) \frac{\hat{\mathbf{p}}_k}{\|\hat{\mathbf{p}}_k\|_\infty}, & \text{if } \|\hat{\mathbf{p}}_k\|_\infty > 1.
\end{cases}
\label{12}
\end{equation}
}
The contracted space is then uniformly partitioned into $n$ blocks (the specific number of blocks used will be discussed further in Sec.~\ref{experiments}.), resulting in a more balanced Gaussian partitioning.
After partitioning the Gaussians, our objective is to ensure that each block is sufficiently trained. In other words, the training data assigned to each block should be highly relevant to the region it represents, focusing on refining the details within the block. To achieve this, we select observations and retain only those that contribute significantly to the visible content of the corresponding block in the rendering results. Since SSIM loss effectively captures structural differences and is somewhat robust to brightness variations~\cite{mss}, we use it as the foundation for our data partition strategy. Specifically, for the $j$-th block, the global Gaussians contained within it are represented as:
\(
G_{Kj} = \{ G_k \mid b_{j, \text{min}} \leq \text{contract}(\hat{\mathbf{p}}_k) < b_{j, \text{max}}, k = 1, \dots, K_j \},
\)
where $b_{j, \text{min}}$ and $b_{j, \text{max}}$ define the spatial bounds of the $j$-th block, and $K_j$ is the number of Gaussians contained within the block.
%Whether the $i$-th pose $\tau_i$ is assigned to the $j$-th block is determined by the following condition:
The set of observations assigned to the \(j\)-th block is defined by the following formula:
\small{
\begin{equation}
\begin{aligned}
\mathbf{P}^1_j = \text{Mask}\left(\mathcal{L}_{\text{SSIM}} \left( I_{G_K} (\boldsymbol{\tau}), I_{G_K \setminus G_{Kj}} (\boldsymbol{\tau}) \right ) > \epsilon \right )
\odot \boldsymbol{\tau},
\end{aligned}
\label{13}
\end{equation}
}
where \( \text{Mask}(\cdot) \) generates an element-wise binary mask. Each element of the mask is set to 1 if it satisfies the condition inside the mask (i.e. the SSIM loss exceeds a threshold \( \epsilon \)), and 0 otherwise.
The term \( G_K \setminus G_{Kj} \) denotes the portion of the global set \( G_K \) excluding the block \( G_{Kj} \). \( \boldsymbol{\tau} \) is a matrix containing all camera poses, with each column \( \tau_i \) representing the \( i \)-th camera pose.
\( \odot \) is element-wise product operation.
And the resulting set \( \mathbf{P}^1_j \) represents the camera poses assigned to the \( j \)-th block.
%Where $G_K \setminus G_{Kj}$ denotes the portion of $G_K$ that does not include $G_{Kj}$. The $i$-th pose $\tau_i$ is assigned to the $j$-th block only when the SSIM loss $L_{\text{SSIM}}$ exceeds a threshold $\epsilon$.
However, this strategy does not account for the projection of the considered block, which may lead to artifacts at the edges of the block. To address this issue, we further include poses that fall within the boundaries of the considered block:
\begin{equation}
\mathbf{P}^2_j = \text{Mask}\left(b_{j, \text{min}} \leq \text{contract} (\hat{\mathbf{p}}_{\tau_i}) < b_{j, \text{max}} \right) \odot \boldsymbol{\tau}.
\label{14}
\end{equation}
where $\hat{\mathbf{p}}_{\tau_i}$ is the position under the world coordinate of pose $i$. The final assignment is:
% \begin{equation}
% B_2 (\tau_i, G_{Kj}) =
% \begin{cases}
% 1, & b_{j, \text{min}} \leq \text{contract} (\hat{\mathbf{p}}_{\tau_i}) < b_{j, \text{max}}, \\
% 0, & \text{otherwise}.
% \end{cases}
% \label{14}
% \end{equation}
% \begin{equation}
% \mathbf{P}^2_j = \text{Mask}\left(b_{j, \text{min}} \leq \text{contract} (\hat{\mathbf{p}}_{\tau_i}) < b_{j, \text{max}} \right) \odot \boldsymbol{\tau}.
% \label{14}
% \end{equation}
% where $\hat{\mathbf{p}}_{\tau_i}$ is the position under the world coordinate of pose $i$. The final assignment is:
% \begin{equation}
% \mathbf{P}_j(\boldsymbol{\tau}, G_{Kj}) = \mathrm{Merge}\bigl(\mathbf{P}^1_j, \mathbf{P}^2_j\bigr),
% \end{equation}
% where $\mathrm{Merge}$ denotes the concatenate operator that removes any duplicate elements, ensuring only one copy of each element is retained.
% \begin{figure*}[h]
% \centering
% \centering \includegraphics[width=1.0\textwidth]{fig3.png}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{%
% \textbf{\textcolor{red}{Qualitative Comparison on the Mip-NeRF 360 Dataset. }}%
% Three representative scenes demonstrate that our method more faithfully preserves fine-scale structures and achieves superior visual fidelity compared to 3DGS and Mip-Splatting.
% }
% \label{fig:4}
% \end{figure*}
\begin{equation}
\mathbf{P}_j(\boldsymbol{\tau}, G_{Kj}) = \mathrm{Merge}\bigl(\mathbf{P}^1_j, \mathbf{P}^2_j\bigr),
\end{equation}
where $\mathrm{Merge}$ denotes the concatenate operator that removes any duplicate elements, ensuring only one copy of each element is retained.
To prevent overfitting, we employ a binary search method~\cite{binary} to incrementally expand $b_{j, \text{min}}$ and $b_{j, \text{max}}$ until $K_j$ exceeds a predefined threshold.
Notably, this procedure is applied exclusively during the data partitioning phase for each block.
\noindent\textbf{Importance-Driven Gaussian Pruning (IDGP).}
After the Gaussian primitives and data division, we proceed to train each block in parallel in the original uncontracted space. Specifically, we first initialize each block using the coarse global Gaussian prior, and then fine‐tune each block using high‐resolution data as detailed in Sec.~\ref{3.3}. During block‐level optimization, we further accelerate convergence and reduce redundancy by applying a lightweight importance scoring and pruning strategy. Let $\mathcal{R}_b$ denote the set of all rays cast from the training views assigned to block $b$. For each Gaussian primitive $p_i$ in block $b$, we only consider its interactions with $\mathcal{R}_b$ and define the weighted hit count as
\small{
\begin{equation}
H_i \;=\;\sum_{r\in\mathcal{R}_b}\mathbf{1}(p_i \cap r)\,T_{i,r},
% \end{equation}
% \begin{equation}
\text{where } T_{i,r} =\prod_{\substack{p_k \cap r\\\mathrm{depth}(p_k)<\mathrm{depth}(p_i)}}\bigl(1 - \alpha_k\bigr).
\end{equation}
}
Here, $\mathbf{1}(p_i \cap r)=1$ if and only if ray $r$ intersects $p_i$, and $T_{i,r}$ accumulates the transmission up to $p_i$ by all closer primitives $p_k$. We then compute the raw volume of $p_i$ as
\small{
$
v_i \;=\;\prod_{d=1}^3 s_{i,d},
$}
where each \(s_{i,d}\) is the scale factor of $p_i$ along the $d$-th spatial axis, and apply logarithmic compression
$
\widetilde{v}_i \;=\;\ln\bigl(1 + v_i\bigr).
$
Finally, we assign each primitive an importance score with its opacity $\alpha_i$:
$
S_i \;=\;\alpha_i\;\widetilde{v}_i\;H_i.
$
After evaluating $\{S_i\}$ for all primitives in the block, we sort them in descending order and remove the lowest 20\%. The remaining Gaussians, now both globally informed by the coarse prior and locally pruned of low‐impact points, continue through block‐level fine‐tuning. Finally, we select the fine‐tuned Gaussians within each block and, guided by the global geometric prior, concatenate the blocks to obtain the fine‐tuned global Gaussian. Through this process, the previously coarse global Gaussians are significantly enhanced in areas where they lacked detail.
\subsection{Loss Function}
\label{3.3}
To optimize both the coarse and refined stages,the loss functions are defined as follows. First, we use the RGB loss $\mathcal{L}_{RGB}$ from 3DGS for the novel view synthesis task. To reconstruct scene surfaces, we enforce normal priors $\mathbf{N}$ predicted by a pretrained monocular deep neural network~\cite{BaeD24} to supervise the rendered normal map $\hat{\mathbf{N}}$ using $\mathbf{L1}$ and cosine losses:
\begin{equation}
\mathcal{L}_n = \| \hat{\mathbf{N}} - \mathbf{N} \|_1 + (1 - \hat{\mathbf{N}} \cdot \mathbf{N}).
\label{eq2}
\end{equation}
Additionally, to effectively update Gaussian positions, we utilize the predicted normal $\mathbf{N}$ from the pretrained model to supervise the D-Normal $\overline{\mathbf{N}}_d$. The D-Normal is derived from the rendered depth by computing the cross-product of horizontal and vertical finite differences from neighboring points:
\begin{equation}
\overline{\mathbf{N}}_d = \frac{\nabla_v \mathbf{d} \times \nabla_h \mathbf{d}}{\left|\nabla_v \mathbf{d} \times \nabla_h \mathbf{d}\right|},
\label{eq5}
\end{equation}
where $\mathbf{d}$ represents the 3D coordinates of a pixel obtained via back-projection from the depth map. We then apply the D-Normal regularization from ~\cite{vcR}:
\begin{equation}
\mathcal{L}_{dn} = w \cdot \left(\| \bar{\mathbf{N}}_d - \mathbf{N} \|_1 + (1 - \bar{\mathbf{N}}_d \cdot \mathbf{N})\right),
\label{eq8}
\end{equation}
where $w$ is a confidence term. The overall loss function integrates these components:
\begin{equation}
\mathcal{L}_{total} = \mathcal{L}_{RGB} + \lambda_1 \mathcal{L}_s + \lambda_2 \mathcal{L}_n + \lambda_3 \mathcal{L}_{dn},
\label{eq9}
\end{equation}
where $\lambda_1$, $\lambda_2$, and $\lambda_3$ balance the individual terms. The term $\mathcal{L}_s$ is introduced to simplify depth computation, as described in~\cite{vcR}.
\begin{figure*}[t]
\centering
\centering \includegraphics[width=1.0\textwidth]{fig3.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{%
\textbf{Qualitative Comparison on the Mip-NeRF 360 Dataset. }%
Three representative scenes demonstrate that our method more faithfully preserves fine-scale structures and achieves superior visual fidelity compared to 3DGS and Mip-Splatting.
}
\label{fig:4}
% \vspace{-0.2cm}
\end{figure*}
\definecolor{BurntOrange}{RGB}{204, 85, 0}
\begin{figure*}[t]
\centering
\centering \includegraphics[width=1.0\textwidth,height=0.2\textheight]{mesh.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Qualitative Comparison on TNT dataset.} Reconstructions from left to right—SuGar, NeuS, 2DGS, and VCR-Gaus—demonstrate that our method delivers more complete surface geometry, enhanced smoothness in planar regions, and superior preservation of fine structural details, thereby outperforming existing approaches in geometric fidelity. }
\label{fig:3}
\vspace{-0.2cm}
\end{figure*}
\section{Experiment}
\label{experiments}
\subsection{Experimental setups}
\noindent\textbf{Dataset and Metrics.}
To evaluate the effectiveness of our reconstruction method, we conduct experiments on two core tasks: novel view synthesis (NVS) and surface reconstruction, using multiple benchmark datasets. We first assess high-resolution NVS performance on Mip-NeRF360~\cite{mip}, followed by high-fidelity surface reconstruction on the Tanks and Temples (TNT)\cite{TNT} dataset. Additionally, we perform comparative experiments on the Replica\cite{replica} dataset to further validate our method. For a comprehensive evaluation, we employ standard metrics including SSIM, PSNR, LPIPS, and F1-score. Rendering efficiency is also assessed in terms of frames per second (FPS).
%
% To validate the effectiveness of our reconstruction method, we conduct evaluations on two key sub-tasks: Novel View Synthesis and Surface Reconstruction on multiple datasets. We first assessed the high-resulution NVS on the Mip-NeRF360~\cite{mip}. Subsequently, we further validated the high-fidelity surface reconstruction quality of our method on the Tanks and Temples (TNT)~\cite{TNT} dataset. To comprehensively verify the effectiveness of our method, we also conducted comparative experiments on the Replica~\cite{replica} with other methods. To provide a thorough evaluation of the reconstruction quality across different methods, we adopted standard metrics, including SSIM, PSNR, LPIPS and F1-score. Additionally, to assess the rendering speed, we compared the frames per second (FPS) performance of each method.
%
\noindent\textbf{Implementation Details.}
We begin by following the 3DGS~\cite{3dgs} pipeline, performing 30{,}000 iterations at a low resolution (0.3K) to obtain a coarse global Gaussian prior. During this stage, we introduce our Importance-Driven Gaussian Pruning (IDGP) strategy, which scores the rendering contribution of each Gaussian primitive and prunes those with the lowest impact. This step prevents irrelevant viewpoints from being assigned to training blocks in subsequent stages, reducing unnecessary computational overhead. The resulting coarse prior serves as initialization for the refinement phase.
In the contraction stage, we define the central one-third of the scene as the internal region and the remainder as the external region. The contracted Gaussians are then divided into four spatial sub-blocks. For data assignment, we use an SSIM threshold of $\epsilon = 0.1$. Each sub-block is further trained for 30{,}000 iterations. Specifically, we apply IDGP at the 10{,}000th, 15{,}000th, and 25{,}000th iterations to prune low-impact Gaussians based on their interaction contributions with training rays. This dynamic pruning accelerates convergence and reduces computational redundancy.
To facilitate surface reconstruction, we adopt the depth-normal regularization method described in Sec.~\ref{3.3}. Specifically, we use the pretrained DSINE~\cite{BaeD24} model for outdoor scenes and the pretrained GeoWizard~\cite{GeoWizard} for indoor scenes to predict normal maps. The hyperparameters $\lambda_1, \lambda_2,$ and $\lambda_3$ are set to 1, 0.01, and 0.015, respectively. After rendering the depth maps, we perform truncated signed distance function (TSDF) fusion and process the results using Open3D~\cite{Open3D}.
Additional details are provided in the supplementary.
\begin{table}[t]
\centering
% \vspace{-0.5cm}
\caption{\textbf{Mip-NeRF 360 Full-Resolution Results.} The rendering quality comparison highlights the best and second-best results.}
\label{tab:mipnerf_transposed}
% 减小列间距，默认是6pt，这里减到3pt
\setlength{\tabcolsep}{5pt}
\begin{tabular}{@{}lccccccc@{}}
\toprule
& Mip-NeRF & Instant-NGP & zip-NeRF & 3DGS & 3DGS+EWA & Mip-Splatting & \textbf{Ours} \\
\midrule
PSNR $\uparrow$ & 24.10 & 24.27 & 20.27 & 19.59 & 22.81 & \textcolor{cyan}{26.22} & \textcolor{BurntOrange}{27.91} \\
SSIM $\uparrow$ & 0.706 & 0.698 & 0.559 & 0.619 & 0.643 & \textcolor{cyan}{0.765} & \textcolor{BurntOrange}{0.863} \\
LPIPS $\downarrow$ & 0.428 & 0.475 & 0.494 & 0.476 & 0.449 & \textcolor{cyan}{0.392} & \textcolor{BurntOrange}{0.342} \\
\bottomrule
\vspace{-0.4cm}
\end{tabular}
\label{tab1}
\end{table}
\begin{table*}[t]
\centering
\small
\caption{\textbf{Quantitative Results on the Tanks and Temples Dataset \cite{TNT}.} The best results are highlighted in \textbf{\textcolor{BurntOrange}{orange}}, while the second-best results are marked in \textbf{\textcolor{cyan}{blue}}. }
%Reconstructions are evaluated with the official evaluation scripts and we report F1-score, average optimization time and FPS. Ours outperforms all 3DGS-based surface reconstruction methods by a large margin and performs better than neural implicit methods by a minor margin while optimizing significantly faster.}
\setlength{\tabcolsep}{4pt} % 紧凑列间距
\resizebox{0.88\textwidth}{!}{ % 调整缩放比例
\begin{tabular}{@{}lccccccc@{}} % 列数从 9 列减少为 8 列
\hline
& \multicolumn{2}{c}{\textbf{NeuS-based}} & \multicolumn{5}{c}{\textbf{Gaussian-based}} \\ % NeuS-based 从3列改为2列
\cline{2-3} \cline{4-8} % 调整横线范围
\textbf{Scene} & NeuS & MonoSDF & SuGaR & 3DGS & 2DGS & VCR-GauS & Ours \\ % 删除 Geo-Neus
\hline
Barn & 0.29 & 0.49 & 0.14 & 0.13 & 0.36 & \textcolor{cyan}{0.62} & \textcolor{BurntOrange}{0.65} \\
Caterpillar & \textcolor{cyan}{0.29} & \textcolor{BurntOrange}{0.31} & 0.16 & 0.08 & 0.23 & 0.26 & \textcolor{cyan}{0.29} \\
Courthouse & 0.17 & 0.12 & 0.08 & 0.09 & 0.13 & \textcolor{cyan}{0.19} & \textcolor{BurntOrange}{0.23} \\
Ignatius & \textcolor{BurntOrange}{0.83} & \textcolor{cyan}{0.78} & 0.33 & 0.04 & 0.44 & 0.61 & 0.64 \\
Meetingroom & \textcolor{BurntOrange}{0.24} & \textcolor{cyan}{0.23} & 0.15 & 0.16 & 0.16 & 0.19 & \textcolor{BurntOrange}{0.24} \\
Truck & 0.45 & 0.42 & 0.26 & 0.18 & 0.16 & \textcolor{cyan}{0.52} & \textcolor{BurntOrange}{0.61} \\
\hline
Mean & 0.38 & 0.39 & 0.19 & 0.09 & 0.30 & \textcolor{cyan}{0.40} & \textcolor{BurntOrange}{0.45} \\
Time & $>$24h & $>$24h & $>$1h & \textcolor{BurntOrange}{14.3m} & \textcolor{cyan}{34.2m} & 53m & 2h \\
\hline
FPS & \multicolumn{2}{c}{$<$10} & -- & \textcolor{BurntOrange}{159} & 68 & 145 & \textcolor{cyan}{146} \\ % 合并列数同步调整
\hline
\end{tabular}
}
\label{tab2}
% \vspace{-0.1cm}
\end{table*}
% To validate the effectiveness of our reconstruction method, we conduct evaluations on two key sub-tasks: Novel View Synthesis and Surface Reconstruction.
% \begin{wraptable}{r}{0.5\textwidth}
% \vspace{-1em} % 控制表格与上一行文字的垂直距离，可调
% \centering
% \scriptsize
% \caption{\textbf{Results on Mip-NeRF 360.} Best and second-best results are highlighted.}
% \label{tab:mipnerf}
% \begin{tabular}{@{}lccc@{}}
% \toprule
% \textbf{Method} & PSNR $\uparrow$ & SSIM $\uparrow$ & LPIPS $\downarrow$ \\
% \midrule
% Mip-NeRF & 24.10 & 0.706 & 0.428 \\
% Instant-NGP & 24.27 & 0.698 & 0.475 \\
% zip-NeRF & 20.27 & 0.559 & 0.494 \\
% 3DGS & 19.59 & 0.619 & 0.476 \\
% 3DGS+EWA & 22.81 & 0.643 & 0.449 \\
% Mip-Splatting & \textcolor{cyan}{26.22} & \textcolor{cyan}{0.765} & \textcolor{cyan}{0.392} \\
% \textbf{Ours} & \textcolor{BurntOrange}{27.91} & \textcolor{BurntOrange}{0.863} & \textcolor{BurntOrange}{0.342} \\
% \bottomrule
% \end{tabular}
% \vspace{-1em} % 控制表格与下一段文字的垂直距离
% \end{wraptable}
\noindent\textbf{Novel View Synthesis.}
As shown in Tab.~\ref{tab1}, we compare our method with several existing approaches, including mip-NeRF~\cite{Mip-NeRF360}, Instant-NGP~\cite{InstantNGP}, zip-NeRF~\cite{zip-nrf}, 3DGS~\cite{3dgs}, 3DGS+EWA~\cite{EWV}, and Mip-Splatting~\cite{mip360-splatting}. At high resolutions, our method significantly outperforms all state-of-the-art techniques.
As shown in Fig.~\ref{fig:4}, our method produces high-fidelity imagery devoid of fine-scale texture distortions.
While 3DGS~\cite{3dgs} introduces noticeable erosion artifacts due to dilation operations, Mip-Splatting~\cite{mip360-splatting} shows improved performance, yet still exhibits evident texture distortions. In contrast, our method avoids such issues, producing images that are both aesthetically pleasing and closely aligned with the ground truth, demonstrating the effectiveness of our hierarchicall refined strategy.
% \caption{\textbf{Quantitative results on the Tanks and Temples Dataset \cite{TNT}.} The best results are highlighted in \textbf{\textcolor{BurntOrange}{orange}}, while the second-best results are marked in \textbf{\textcolor{cyan}{blue}}. Reconstructions are evaluated with the official evaluation scripts and we report F1-score, average optimization time and FPS. Ours outperforms all 3DGS-based surface reconstruction methods by a large margin and performs better than neural implicit methods by a minor margin while optimizing significantly faster.}
% \vspace{-0.3cm}
% \label{tab2}
% \end{table*}
% \vspace{-0.3cm}
% \subsection{Comparison}
%为了验证我们重建方法的有效性，我们主要在两个子任务上进行了验证：Novel view Synthesis 和 Surface Reconstruction。
% \begin{wraptable}{r}{0.7\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% \caption{\textbf{Quantitative Results on the ScanQA.}}
% \label{tab:scanqa}
% \renewcommand{\arraystretch}{0.5}
% \setlength{\tabcolsep}{6pt}
% \begin{tabular}{lccccc}
% \toprule
% \textbf{Method} & \textbf{CIDEr} & \textbf{METEOR} & \textbf{ROUGE} & \textbf{EM@1} & \textbf{EM@1-R} \\
% \midrule
% LLaVA-OV & 47.4 & 12.1 & 28.7 & 14.6 & 31.5 \\
% ChatSplat & \textbf{56.3} & \textbf{13.8} & \textbf{31.3} & \textbf{15.9} & \textbf{31.7} \\
% \bottomrule
% \end{tabular}
% \vspace{-0.42cm}
% \end{wraptable}
% To validate the effectiveness of our reconstruction method, we conduct evaluations on two key sub-tasks: Novel View Synthesis and Surface Reconstruction.
% \noindent\textbf{Novel View Synthesis.}
% As shown in Tab.~\ref{tab1}, we compare our method with several existing approaches, including mip-NeRF~\cite{Mip-NeRF360}, Instant-NGP~\cite{InstantNGP}, zip-NeRF~\cite{zip-nrf}, 3DGS~\cite{3dgs}, 3DGS+EWA~\cite{EWV}, and Mip-Splatting~\cite{mip360-splatting}.
% Our approach demonstrates comparable performance to these prior methods at one-eighth of the original resolution.
% Furthermore, at higher resolutions, our method significantly outperforms all state-of-the-art techniques.
% As shown in Fig.~\ref{fig:4}, our method produces high-fidelity imagery devoid of fine-scale texture distortions.
% While 3DGS~\cite{3dgs} introduces noticeable erosion artifacts due to dilation operations, Mip-Splatting~\cite{mip360-splatting} shows improved performance, yet still exhibits evident texture distortions. In contrast, our method avoids such issues, producing images that are both aesthetically pleasing and closely aligned with the ground truth, demonstrating the effectiveness of our hierarchicall refined strategy.
% \begin{table}[h]
% \centering
% \caption{\textbf{Quantitative assessment on Replica \cite{replica}.} \textbf{Bold} indicates the best.} % 上移至此
% \label{tab:3} % 标签紧随标题后
% \begin{tabular}{@{}llcc@{}}
% \toprule
% \textbf{Type} & \textbf{Method} & \textbf{F1-score} & \textbf{Time} \\
% \midrule
% \multirow{2}{*}{Implicit}
% & NeuS & 65.12 & \multirow{2}{*}{\(>\)10h} \\
% & MonoSDF & \textbf{81.64} & \\
% \cline{1-4}
% \multirow{4}{*}{Explicit}
% & 3DGS & 50.79 & \multirow{4}{*}{\(\leq\)2h} \\
% & SuGar & 63.20 & \\
% & 2DGS & 64.36 & \\
% & Ours & \textbf{74.87} & \\
% \bottomrule
% \end{tabular}
% \end{table}
% \begin{wraptable}{r}{0.55\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% \caption{\textbf{Ablation of the Data Division on the ``Ignatius'' scene from the TNT dataset.}
% SO/BO Ass. refer to SSIM-/Boundary-based Observation Assignment.}
% %SO Ass. is the abbreviation for SSIM-based Observation Assignment, and BO Ass. stands for Boundary-based Observation Assignment.}
% \label{tab4}
% \renewcommand{\arraystretch}{0.9}
% \setlength{\tabcolsep}{6pt}
% \begin{tabular}{lcc}
% \toprule[1pt]
% \textbf{} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% baseline & 21.87 & 0.55 \\
% w/o contraction & 22.73 & 0.54 \\
% w/o SO Ass. & 24.29 & 0.58 \\
% w/o BO Ass. & 22.96 & 0.60 \\
% Full & \textbf{25.24} & \textbf{0.64} \\
% \bottomrule[1pt]
% \end{tabular}
% \vspace{-0.5cm}
% \end{wraptable}
% \begin{wraptable}{r}{0.6\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% \caption[Short caption]{\textbf{Quantitative Results on Replica~\cite{replica}.}
% \textbf{Bold} indicates the best.} % 自动换行
% \label{tab:3}
% \renewcommand{\arraystretch}{0.9}
% \setlength{\tabcolsep}{4pt} % 缩小列间距
% \begin{tabular}{@{}llcc@{}} % 移除默认列边距
% \toprule
% \textbf{Type} & \textbf{Method} & \textbf{F1-score} & \textbf{Time} \\
% \midrule
% \multirow{2}{*}{Implicit}
% & NeuS & 65.12 & \multirow{2}{*}{$>$10h} \\
% & MonoSDF & \textbf{81.64} & \\
% \cmidrule{1-4}
% \multirow{4}{*}{Explicit}
% & 3DGS & 50.79 & \multirow{4}{*}{$\leq$2h} \\
% & SuGar & 63.20 & \\
% & 2DGS & 64.36 & \\
% & Ours & \textbf{74.87} & \\
% \bottomrule
% \end{tabular}
% \vspace{-0.0cm}
% \end{wraptable}
\begin{wraptable}{r}{0.42\textwidth}
\small
\vspace{-0.4cm}
\centering
\caption{\textbf{Comparsions on Replica~\cite{replica}.}} % \textbf{Bold} indicates the best.
\label{tab:3}
\renewcommand{\arraystretch}{0.9}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{llcc}
\toprule
\textbf{Type} & \textbf{Method} & \textbf{F1-score} & \textbf{Time} \\
\midrule
\multirow{2}{*}{Implicit}
& NeuS & 65.12 & \multirow{2}{*}{$>$10h} \\
& MonoSDF & \textbf{81.64} & \\
\cmidrule{1-4}
\multirow{4}{*}{Explicit}
& 3DGS & 50.79 & \multirow{4}{*}{$\leq$2h} \\
& SuGar & 63.20 & \\
& 2DGS & 64.36 & \\
& Ours & \textbf{74.87} & \\
\bottomrule
\end{tabular}
\vspace{-0.4cm}
\end{wraptable}
\noindent\textbf{Surface Reconstruction.}
Our method not only provides high-quality novel views synthesis but also enables precise 3D surface reconstruction.
As shown in Tab.~\ref{tab2}, our method outperforms both NeuS-based methods (e.g., NeuS~\cite{NeuS}, MonoSDF~\cite{MonoSDF}, and Geo-NeuS~\cite{Geo-Neus}) and Gaussian-based methods (e.g., 3DGS~\cite{3dgs}, SuGaR~\cite{SuGaR}, 2DGS~\cite{2Dgs}, and VCR-GauS~\cite{vcr-gaus}) on the TNT dataset. Compared to NeuS-based methods, our approach demonstrates significantly faster reconstruction speeds. While our method is slightly slower than some concurrent works, it achieves considerably better reconstruction quality, with a noticeable improvement over 2DGS (0.3 \vs 0.45). Furthermore, our method outperforms the recent state-of-the-art method in surface reconstruction, VCR-GauS, with a higher reconstruction quality (0.45 \vs 0.4). As shown in Fig.~\ref{fig:3}, our method is particularly effective at recovering finer geometric details. Additionally, our method has a substantial advantage in rendering speed, surpassing the rendering speed of 2DGS by more than double.
On the Replica dataset, as shown in Tab.~\ref{tab:3},
our method achieves comparable performance to MonoSDF~\cite{MonoSDF} while running significantly faster.
Moreover, in comparison to explicit reconstruction approaches, 3DGS~\cite{3dgs}, SuGaR~\cite{SuGaR}, and 2DGS~\cite{2Dgs}, our method delivers substantially higher F1-scores.
\subsection{Ablation Studies}
%我们在TNT数据集Ignatius上验证了不同的设计选择对重建质量的有效性，包括划分的块数以及数据划分策略。
% We validated the impact of different design choices on reconstruction quality using the Ignatius scene from the TNT dataset, specifically evaluating the data partitioning strategy and the number of blocks.
To validate the effectiveness of individual components in our method, we conducted a series of ablation experiments on the ``Stump'' scene from the Mip-NeRF 360 dataset and the ``Ignatius'' scene from the TNT dataset. Specifically, we evaluated the impacts of the following components: hierarchical block optimization strategy, Importance-Driven Gaussian Pruning (IDGP), and data partitioning strategy.
% 调整后的表格代码
% \begin{table}[htbp]
% \centering
% \captionsetup{font=small}
% % 左表格 (宽度48%，增加3%)
% \begin{minipage}[t]{0.48\textwidth} % 原0.45改为0.48
% \centering
% \caption{\textbf{Ablation of the Data Division on the ``Ignatius'' scene from the TNT dataset.} SO Ass. is the abbreviation for SSIM-based Observation Assignment, and BO Ass. stands for Boundary-base Observation Assignment.}
% \begin{tabular}{@{\extracolsep{4pt}}lcc@{}} % 列间距从2pt增加到4pt
% \toprule[1pt]
% \textbf{} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% baseline & 21.87 & 0.55 \\
% w/o contraction & 22.73 & 0.54 \\
% w/o SO Ass. & 24.29 & 0.58 \\
% w/o BO Ass. & 22.96 & 0.60 \\
% Full & \textbf{25.24} & \textbf{0.64} \\
% \bottomrule[1pt]
% \end{tabular}
% \label{tab4}
% \end{minipage}
% \hfill
% % 右表格 (宽度48%，增加3%)
% \begin{minipage}[t]{0.48\textwidth} % 原0.45改为0.48
% \centering
% \caption{\textbf{Ablation of the number of blocks on the ``Ignatius'' scene from the TNT dataset.}}
% \begin{tabular}{@{\extracolsep{4pt}}lcc@{}} % 列间距从2pt增加到4pt
% \toprule[1pt]
% \textbf{Blocks} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% 2 & 24.86 & 0.62 \\
% 4 & \textbf{25.24} & \textbf{0.64} \\
% 8 & 23.93 & 0.61 \\
% 16 & 22.56 & 0.57 \\
% \bottomrule[1pt]
% \end{tabular}
% \label{tab5}
% \end{minipage}
%如表5所示，我们分析了数据划分策略的影响，并将原始的高斯全局先验作为基准。表~\ref{tab5}中第一行和最后一行的结果表明，我们提出的方法在提高性能方面是有效的（0.55 \vs 0.64）。表~\ref{tab5}中的第二行结果进一步表明，在收缩后的空间中分配有用的相关数据对于提升重建质量是必需的。表~\ref{tab5}中的第三行则强调了等式(3)对数据划分的重要性，并且我们发现等式(4)对于避免块边缘出现伪影起到了积极作用。、
% \begin{wraptable}{r}{0.45\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% % \caption{\textbf{Ablation of the number of blocks on the ``Ignatius'' scene from the TNT dataset.}}
% \caption{\chl{\textbf{Ablation on number of blocks.}}}
% \label{tab5}
% \renewcommand{\arraystretch}{0.9}
% \setlength{\tabcolsep}{6pt}
% \begin{tabular}{lcc}
% \toprule[1pt]
% \textbf{Blocks} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% 2 & 24.86 & 0.62 \\
% 4 & \textbf{25.24} & \textbf{0.64} \\
% 8 & 23.93 & 0.61 \\
% 16 & 22.56 & 0.57 \\
% \bottomrule[1pt]
% \end{tabular}
% \vspace{-0.0cm}
% \end{wraptable}
% \begin{wraptable}{r}{0.55\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% \caption{\textbf{Ablation of the Data Division on the 'Ignatius' scene from the TNT dataset.} SO Ass. is the abbreviation for SSIM-based Observation Assignment, and BO Ass. stands for Boundary-based Observation Assignment.}
% \label{tab4}
% \renewcommand{\arraystretch}{0.9}
% \setlength{\tabcolsep}{6pt}
% \begin{tabular}{lcc}
% \toprule[1pt]
% \textbf{} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% baseline & 21.87 & 0.55 \\
% w/o contraction & 22.73 & 0.54 \\
% w/o SO Ass. & 24.29 & 0.58 \\
% w/o BO Ass. & 22.96 & 0.60 \\
% Full & \textbf{25.24} & \textbf{0.64} \\
% \bottomrule[1pt]
% \end{tabular}
% \vspace{-0.1cm}
% \end{wraptable}
% \begin{wraptable}{r}{0.55\textwidth}
\begin{table*}[h]
\small
% \vspace{-0.4cm}
\centering
% \caption{\textbf{Ablation of the Data Division on the 'Ignatius' scene from the TNT dataset.} “SO Ass.” denotes SSIM-based Observation Assignment, while “BO Ass.” refers to Boundary-based Observation Assignment.}
\caption{\textbf{Ablation on Data Division.} “SO Ass.” refers to SSIM-based assignment, while “BO Ass.” denotes boundary-based assignment. \textbf{Bold} indicates the best.}
\label{tab4}
\renewcommand{\arraystretch}{0.9}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{lccccc}
\toprule[1pt]
\multirow{2}{*}{\textbf{Method}} & \multicolumn{5}{c}{\textbf{Settings}} \\
\cmidrule{2-6}
& baseline & w/o contraction & w/o SO Ass. & w/o BO Ass. & Full \\
\midrule[0.8pt]
PSNR $\uparrow$ & 21.87 & 22.73 & 24.29 & 22.96 & \textbf{25.24} \\
F1 $\uparrow$ & 0.55 & 0.54 & 0.58 & 0.60 & \textbf{0.64} \\
\bottomrule[1pt]
\end{tabular}
\label{tab4}
% \vspace{-0.2cm}
\end{table*}
% \end{wraptable}
% \end{table}
\noindent\textbf{Ablation of the Data Division.}
As shown in Tab~\ref{tab4}, we analyzed the impact of the data partitioning strategy, using the original Gaussian global prior as the baseline. The results in the first and last rows of Table~\ref{tab5} demonstrate the effectiveness of our proposed method in improving performance (0.55 \vs 0.64). The second row in Table~\ref{tab5} further indicates that assigning relevant data in the contracted space is essential for enhancing reconstruction quality. The third row in Table~\ref{tab5} highlights the importance of starategy 1 (Eq.~\ref{13}) in data partitioning, and we also found that starategy 2 (Eq.~\ref{14}) plays a significant role in preventing artifacts at the edges of blocks.
\begin{wraptable}{r}{0.43\textwidth}
\small
\vspace{-0.4cm}
\centering
\caption{\textbf{Ablation on number of blocks.}}
\label{tab5_transposed}
\renewcommand{\arraystretch}{0.9}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{lcccc}
\toprule[1pt]
& \textbf{2} & \textbf{4} & \textbf{8} & \textbf{16} \\
\midrule[0.8pt]
\textbf{PSNR} $\uparrow$ & 24.86 & \textbf{25.24} & 23.93 & 22.56 \\
\textbf{F1} $\uparrow$ & 0.62 & \textbf{0.64} & 0.61 & 0.57 \\
\bottomrule[1pt]
\end{tabular}
\label{tab5}
\vspace{-0.3cm}
\end{wraptable}
\noindent\textbf{Ablation of the Number of Blocks.}
As shown in Tab.~\ref{tab5}, we investigate how the number of blocks affects reconstruction performance by splitting the coarse global Gaussian into 2, 4, 8, or 16 blocks. Our results indicate that too few blocks can cause conflicts between local and global optima, resulting in insufficient refinement of fine details, whereas too many blocks may lead to imbalanced data distribution and local overfitting. Consequently, we select four blocks for our experiments on the TNT dataset.
% \vspace{-0.3cm}
\begin{table}[h]
\centering
\small
\caption{\textbf{Ablation Studies on the ``Stump'' Scene of the Mip-NeRF 360 Dataset\cite{Mip-NeRF360}.} }
\label{tab:6}
\begin{tabular}{@{}lccccc@{}}
\toprule
\textbf{Method} & \textbf{Model Size(MB)} & \textbf{GPU Memory(G)} & \textbf{PSNR} & \textbf{SSIM} & \textbf{LPIPS} \\
\midrule
Baseline & 348.62 & 23 & 26.39 & 0.804 & 0.291 \\
Baseline w/o IDPG & 601.04 & 28 & 26.41 & 0.791 & 0.288 \\
\bottomrule
\label{tab6}
\end{tabular}
\vspace{-0.2cm}
\end{table}
\noindent\textbf{Ablation of Importance-Driven Gaussian Pruning.}
As shown in Tab.~\ref{tab6}, to validate the effectiveness of the proposed Importance-Driven Gaussian Pruning (IDGP) strategy, we conducted an ablation study on the ``stump'' scene of the Mip-NeRF 360 dataset.
Specifically, we compare the full method with the IDGP mechanism (Baseline) against a control variant in which IDGP is disabled (Baseline w/o IDGP).
%
% Specifically, we compared the full method incorporating IDGP (Baseline) against a control variant without the IDGP mechanism (Baseline/IDGP).
%
As shown in Table.~\ref{tab6}, IDGP is able to selectively prune redundant Gaussians during training without degrading rendering quality, thereby achieving significant improvements in both model structure and computational efficiency.
% }
% \begin{wraptable}{r}{0.45\textwidth}
% \small
% \vspace{-0.4cm}
% \centering
% \caption{\textbf{Ablation of the number of blocks on the ``Ignatius'' scene from the TNT dataset.}}
% \label{tab5}
% \renewcommand{\arraystretch}{0.9}
% \setlength{\tabcolsep}{6pt}
% \begin{tabular}{lcc}
% \toprule[1pt]
% \textbf{Blocks} & \textbf{PSNR} $\uparrow$ & \textbf{F1} $\uparrow$ \\
% \midrule[0.8pt]
% 2 & 24.86 & 0.62 \\
% 4 & \textbf{25.24} & \textbf{0.64} \\
% 8 & 23.93 & 0.61 \\
% 16 & 22.56 & 0.57 \\
% \bottomrule[1pt]
% \end{tabular}
% \vspace{-0.4cm}
% \end{wraptable}
% \noindent\textbf{Ablation of Importance-Driven Gaussian Pruning.}
% %
% % \textcolor{red}{
% }
\section{Conclusion}
In this work, we propose HRGS, a memory-efficient coarse-to-fine framework that uses low-resolution global Gaussians to guide the refinement of high-resolution local Gaussians, enabling high-resolution scene reconstruction under limited GPU memory.
Our novel partitioning strategy for Gaussian primitives and data effectively facilitates block-wise optimization, significantly alleviating the high memory overhead typical of traditional 3DGS in high-resolution 3D reconstruction. Despite reduced memory requirements, our method achieves superior reconstruction quality and demonstrates state-of-the-art performance on two key sub-tasks of 3D reconstruction.
As a result, this work establishes a new baseline for high-resolution 3D reconstruction, setting an important precedent for future research in the field.
\renewcommand\thefigure{\Alph{figure}}
\setcounter{figure}{0}
\renewcommand\thetable{\Alph{table}}
\setcounter{table}{0}
\renewcommand{\thesection}{\Alph{section}}
\setcounter{section}{0}
\setcounter{page}{1}
% \maketitlesupplementary
% \usepackage{multirow} % 补上缺失的multirow包
% \usepackage{siunitx}
% \usepackage{adjustbox}
\definecolor{myorange}{RGB}{230,145,56}
\newcolumntype{Y}{>{\raggedleft\arraybackslash}X}
\newcommand{\ccell}[4]{#1/#2/#3/#4} % 合并cell数据显示
\definecolor{myorange}{RGB}{230,145,56}
\section{Supplementary Material}
\label{sec:supplementary}
% % \title{Supplementary Material for
% % HRGS: Hierarchical Gaussian Splatting for Memory-Efficient High-Resolution 3D Reconstruction}
% % \author{
% % \parbox{\textwidth}{\centering
% % First Author \\
% % Institution1 \\
% % \texttt{firstauthor@i1.org}
% % }
% % \and
% % \parbox{\textwidth}{\centering
% % Second Author \\
% % Institution2 \\
% % \texttt{secondauthor@i2.org}
% % }
% % }
% \maketitle
% \input{secs/appendix}
% 参考文献配置
% \bibliographystyle{plainnat}
% \bibliography{reference} % 对应 reference.bib 文件
% \begin{figure*}[h!]
% \centering
% \centering \includegraphics[width=1.0\textwidth]{fig3.png}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{%
% \textbf{Qualitative Comparison on the Mip-NeRF 360 Dataset.} Comparisons are conducted on the full-resolution dataset, showing that our method more faithfully preserves fine-scale structures and achieves superior visual fidelity compared to 3DGS and Mip-Splatting.}
% \label{fig:1}
% % \vspace{-0.2cm}
% \end{figure*}
\begin{figure*}[ht!]
\centering
\centering \includegraphics[width=1.0\textwidth]{fig3.png}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{%
\textbf{Qualitative comparison on the Mip-NeRF 360 Dataset.} Comparisons are conducted on the full-resolution dataset, showing that our method more faithfully preserves fine-scale structures and achieves superior visual fidelity compared to 3DGS and Mip-Splatting.}
\label{fig:1}
% \vspace{-0.2cm}
\end{figure*}
\section{Implementation Details}
%In most experiments, we utilize PyTorch 2.0.1 with CUDA 11.7, preserving hyperparameter settings identical to those of 3DGS~\cite{3dgs}. For outdoor scenes in the TNT dataset, we integrate decoupled appearance modeling~\cite{vast} to address exposure artifacts. To holistically validate method robustness,We conduct supplementary experiments including: (1) detailed visual comparisons of rendering details under 1× 2×, 4×, and 8× downsampling ratios in the Mip-NeRF 360 bicycle scene (Fig.~\ref{fig:4}), and (2) stepwise degradation analysis through resolution reduction to 1/8 of the original scale (downsampled regime) with quantitative metrics summarized in Table~\ref{tab1}
All experiments are conducted using machines equipped with NVIDIA A5000 GPUs, with PyTorch 2.0.1 and CUDA 11.7 as the software environment. For comparison methods that exceed the memory capacity of the A5000 setup, we employ NVIDIA A800 GPUs to ensure reliable execution and fair evaluation. Unless otherwise specified, we adopt the same hyperparameter settings as 3DGS~\cite{3dgs}. For outdoor scenes in the TNT dataset~\cite{TNT}, we incorporate decoupled appearance modeling~\cite{vast} to mitigate exposure-related artifacts.
%To further enhance reconstruction quality, we adopt the pruning strategy from LightGaussian~\cite{lightgaussian} to remove outlier Gaussians.
All models are trained and evaluated on the same data splits as used in 2DGS~\cite{2Dgs}, across the TNT~\cite{TNT} and Mip-NeRF360~\cite{Mip-NeRF360} benchmarks.
\section{Additional Ablation Study}
In this section, we further investigate the contributions of our proposed modules. We begin by analyzing the block-wise training strategy in Sec.\ref{sec:block_training}. Next, we examine the impact of the Data Division and IDGP strategies in Sec.\ref{sec:data_division} and Sec.\ref{sec:idgp}, respectively. While these ablations were previously conducted in the main paper, they were limited to a single scene. Here, we extend the evaluation to the full dataset, providing a more comprehensive and robust analysis. Finally, in Sec.\ref{sec:res_scale}, we assess the effectiveness of our method across various resolution scales.
\subsection{ Block-Wise Training for High-Resolution Reconstruction}
\label{sec:block_training}
%In our methodology, we adopt a block-wise training strategy to enhance computational efficiency. To validate the effectiveness of this approach in high-resolution scenarios, we conducted comparative experiments with CityGS~\cite{Citygs}, which also employs block-wise partitioning. To ensure fairness in the experiments, we used the block partitioning method proposed by the CityGS authors, dividing the scenes into 4 blocks consistent with our method, and conducted experiments on the full-resolution Mip-NeRF 360 dataset. Evaluation metrics included reconstruction quality indicators such as PSNR, SSIM, and LPIPS, as well as efficiency metrics such as GPU memory consumption, per-block model size, and final aggregated model scale.
% \zhd{
% We employ a block-wise training strategy to improve computational efficiency, and evaluate its effectiveness in high-resolution scenarios through comparisons with CityGS~\cite{Citygs}, which also utilizes block partitioning.
We first validate our proposed block-wise training strategy in high-resolution scenarios by comparing it against the block partitioning scheme adopted in CityGS~\cite{Citygs}.
For fair comparison, we implemented the four-block division scheme following the methodology from the CityGS and conducted experiments on the full-resolution Mip-NeRF 360 dataset.
% we adopt the same four-block division scheme and conduct experiments on the full-resolution Mip-NeRF 360 dataset.
%Evaluation metrics include PSNR, SSIM, LPIPS for reconstruction quality, and GPU memory usage, per-block model size, and aggregated model scale for efficiency.
% As shown in Tab.~\ref{tab:D}, our method achieves superior rendering quality.
% Additionally, Tab.~\ref{tab:cell} illustrates its advantages in reducing memory overhead and improving model scalability.
% Our method demonstrates a consistent reduction in total model size on the Mip-NeRF 360 dataset.
% For instance, 587.05 MB vs. 1126.4 MB for the bicycle scene, and 160.10 MB vs. 278.52 MB for bonsai, with similarly substantial improvements observed at the block level (see Tab.~\ref{tab:cell}).
% %Benefiting from the Importance-Driven Gaussian Pruning (IDGP), our per-block models are 40–60\% smaller than those of CityGS, effectively balancing compactness and geometric fidelity (e.g., 187.22 MB vs. 512.63 MB in bicycle-cell0).
% \chl{In addition}, as illustrated in Fig.~\ref{fig:2}, our method achieves notably lower GPU memory consumption compared to 3DGS, Mip-Splatting, and CityGS, demonstrating the efficacy of structured partitioning and pruning in jointly optimizing computational efficiency and model compactness.
% }
%The experimental results demonstrate that our method achieves superior rendering quality.~\ref{tab:D} while significantly reducing memory overhead and optimizing model scalability. In terms of model scale, in the bicycle scene of the Mip-NeRF 360 dataset, the total model size of CityGS was 1126.4 MB, while ours was only 794.23 MB; in the bonsai scene, CityGS was 278.52 MB, and ours was 283.29 MB, with significantly smaller model sizes in multiple blocks (~\ref{tab:cell}). Thanks to the Importance-Driven Gaussian Pruning (IDGP) strategy, our block partitioning scheme achieved a 40-60\% reduction in per-block model size compared to CityGS. For example, in the cell0 block of the bicycle scene, our model size (567.05 MB) was only about 88\% of CityGS (512.63 MB), effectively controlling the overall model scale while preserving geometric fidelity.
%In terms of GPU memory consumption, our method's average memory footprint on the full-resolution dataset was significantly lower than that of 3DGS, Mip-Splatting, and CityGS (Fig.~\ref{fig:2}). Overall, this strategy achieves a balance between quality, efficiency, and model scale in high-resolution reconstruction through the combination of structured block partitioning and intelligent pruning.
\begin{figure}[ht!]
\centering
\centering \includegraphics[width=1.0\textwidth,height=0.4\textheight]{sub_fig.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Comparison of 3DGS, Mip-Splatting, CityGS and Ours at full resolution on the Mip-NeRF 360 dataset.} }
\label{fig:3}
\end{figure}
As shown in Tab.~\ref{tab:D}, our method achieves superior rendering quality compared to CityGS. As illustrated in Fig.~\ref{fig:3}, our approach demonstrates exceptional detail preservation in intricate geometric features and textural nuances. Additionally, Tab.~\ref{tab:cell} highlights the advantages of our block partitioning strategy over CityGS in enhancing model scalability. Across the Mip-NeRF 360 dataset, our method consistently reduces the total model size. For example, in the bicycle scene, the model size is 587.05 MB compared to the baseline's 1126.4 MB, with comparable improvements observed at the block level Tab.~\ref{tab:cell}. Furthermore, as shown in Fig.~\ref{fig:2}, our method substantially reduces GPU memory consumption compared to 3DGS, Mip-Splatting, and CityGS, underscoring its superior optimization of computational efficiency over CityGS despite employing similar block partitioning strategies.
\begin{table}[t]
\centering
\caption{\textbf{Mip-NeRF 360 Full-Resolution Results.} \textbf{Bold} indicates the best.}
\vspace{2pt}
\footnotesize
\setlength{\tabcolsep}{3.5pt}
\begin{tabular}{@{}lcccccccc@{}}
\toprule
& Mip-NeRF & Instant-NGP & zip-NeRF & 3DGS & 3DGS+EWA & CityGS & Mip-Splatting & \textbf{Ours} \\
\midrule
PSNR $\uparrow$ & 24.10 & 24.27 & 20.27 & 19.59 & 22.81 & 24.71 & {26.22} & {\textbf{27.91}} \\
SSIM $\uparrow$ & 0.706 & 0.698 & 0.559 & 0.619 & 0.643 & 0.702 & {0.765} & {\textbf{0.863}} \\
LPIPS $\downarrow$ & 0.428 & 0.475 & 0.494 & 0.476 & 0.449 & 0.401 & {0.392} & {\textbf{0.342}} \\
\bottomrule
\label{tab:D}
\end{tabular}
\vspace{-0.3cm}
\end{table}
\begin{table}[htbp]
\caption{\textbf{Comparison of model sizes (MB) across CityGS and Ours.} \textbf{Bold} indicates the best.}
\centering
\scriptsize
\begin{tabular}{c|ccccc|ccccc}
\hline
\multirow{2}{*}{Scene} & \multicolumn{5}{c|}{\textbf{CityGS (MB)}} & \multicolumn{5}{c}{\textbf{Ours (MB)}} \\
& Total & cell0 & cell1 & cell2 & cell3 & Total & cell0 & cell1 & cell2 & cell3 \\
\hline
bicycle & {1126.4} & 512.63 & 480.45 & 70.92 & 63.11 & \textbf{587.05} & 187.22 & 238.32 & 100.76 & 61.75 \\
bonsai & {278.52} & 49.97 & 43.83 & 117.54 & 67.18 & \textbf{160.10} & 49.81 & 47.16 & 48.18 & 34.01 \\
counter & {215.85} & 21.38 & 22.26 & 78.73 & 93.49 & \textbf{144.58} & 23.06 & 36.63 & 36.36 & 45.54 \\
flowers & {738.18} & 269.72 & 276.41 & 96.18 & 93.87 & \textbf{399.71} & 163.86 & 133.96 & 65.55 & 63.33 \\
garden & {1196.80} & 269.72 & 276.41 & 96.18 & 93.87 & \textbf{360.52} & 114.76 & 86.78 & 87.60 & 79.39 \\
kitchen & {324.08} & 30.26 & 26.46 & 157.15 & 110.21 & \textbf{170.94} & 38.41 & 35.78 & 76.55 & 47.52 \\
stump & {873.41} & 399.37 & 302.45 & 171.59 & 191.51 & \textbf{368.62} & 103.86 & 103.32 &80.75 & 80.69 \\
treehill& {745.24} & 273.96 & 236.07 & 136.24 & 98.97 & \textbf{385.84} & 109.68 & 149.32 & 66.55 & 60.29 \\
room & {301.02} & 33.58 & 94.33 & 74.08 & 106.03 & \textbf{221.97} & 316.31 & 33.84 & 75.64 & 76.19 \\
\hline
\end{tabular}
\label{tab:cell}
\end{table}
\begin{figure}[htbp]
\centering
\centering \includegraphics[width=0.7\textwidth,height=0.3\textheight]{gpu_memory_comparison.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Average GPU memory consumption of 3DGS, Mip-Splatting, CityGS, and our method on the Full-Resolution Mip-NeRF 360 dataset.} }
\label{fig:2}
\end{figure}
\subsection{ Ablation of the Data Division.}
\label{sec:data_division}
% As shown in Table.~\ref{tab2}, we conducted chunk-based ablation experiments on the entire TNT dataset to evaluate the impact of data-partitioning strategies, using the original Gaussian global prior as the baseline.
% The study systematically assessed the effects of progressively removing strategies such as contraction, SSIM-based assignment (SO Ass.), and boundary-based assignment (BO Ass.) on 3D reconstruction quality, with PSNR and F1 score serving as evaluation metrics.
% %The experimental settings included five groups: baseline (without any optimization strategies), w/o contraction (without contraction strategy), w/o SO Ass. (without SSIM-based assignment strategy), w/o BO Ass. (without boundary-based assignment strategy), and Full (with complete strategies).
% Results showed that the baseline achieved a PSNR of 22.19 and an F1 score of 0.36. Removing the contraction strategy (w/o contraction) increased PSNR to 23.63 and F1 to 0.38. Further removing SO Ass. (w/o SO Ass.) improved PSNR to 24.89 and F1 to 0.41. However, removing BO Ass. (w/o BO Ass.) decreased PSNR to 23.96 and F1 to 0.43. The Full group with complete strategies achieved the optimal performance, reaching a PSNR of 25.02 and an F1 score of 0.45.
As shown in Table.~\ref{tab2}, we conducted chunk-based ablation experiments on the entire TNT dataset to evaluate the impact of data-partitioning strategies, using the original Gaussian global prior as the baseline.
We systematically investigate the contribution of individual components, including contraction, SSIM-based assignment (SO Ass.) and boundary-based assignment (BO Ass.).
The results in the first and last columns of Tab.~\ref{tab2} demonstrate the effectiveness of our proposed method in improving performance (0.36 vs. 0.45).
The second column in Tab.~\ref{tab2} further indicates that assigning relevant data in the contracted space is essential for enhancing reconstruction quality.
The third column in Tab.~\ref{tab2} highlights the importance of SO Ass. in data partitioning, and we also found that BO Ass. plays a significant role in preventing artifacts at the edges of blocks.
\begin{table*}[h]
\small
% \vspace{-0.4cm}
\centering
% \caption{\textbf{Ablation of the Data Division on the 'Ignatius' scene from the TNT dataset.} “SO Ass.” denotes SSIM-based Observation Assignment, while “BO Ass.” refers to Boundary-based Observation Assignment.}
\caption{\textbf{Ablation on Data Division (performed on the full TNT dataset).} “SO Ass.” refers to SSIM-based assignment, while “BO Ass.” denotes boundary-based assignment. \textbf{Bold} indicates the best.}
\label{tab4}
\renewcommand{\arraystretch}{0.9}
\setlength{\tabcolsep}{6pt}
\begin{tabular}{lccccc}
\toprule[1pt]
\multirow{2}{*}{\textbf{Method}} & \multicolumn{5}{c}{\textbf{Settings}} \\
\cmidrule{2-6}
& baseline & w/o contraction & w/o SO Ass. & w/o BO Ass. & Full \\
\midrule[0.8pt]
PSNR $\uparrow$ & 22.19 & 23.63 & 24.89 & 23.96 & \textbf{25.02} \\
F1 $\uparrow$ & 0.36 & 0.38 & 0.41 & 0.43 & \textbf{0.45} \\
\bottomrule[1pt]
\end{tabular}
\label{tab2}
% \vspace{-0.2cm}
\end{table*}
\begin{figure*}[ht]
\centering
\centering \includegraphics[width=1.0\textwidth,height=0.55\textheight]{bicycle.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Qualitative comparsion on Mip-NeRF 360 for different resolution scales.} }
\label{fig:4}
\end{figure*}
\subsection{Ablation of Importance-DrivenGaussian Pruning.}
\label{sec:idgp}
%To validate the effectiveness of the Importance-Driven Gaussian Pruning (IDGP) strategy in high-resolution reconstruction, we conducted ablation experiments on the full-resolution Mip-NeRF 360 dataset, comparing the baseline model with a variant excluding IDGP (denoted as "Baseline w/o IDPG").
%The evaluation focuses on model size, GPU memory consumption, and reconstruction quality metrics (PSNR, SSIM, LPIPS) .
%As shown in Table.~\ref{tab6} , removing IDGP leads to a significant increase in model size from 313.72 MB (Baseline) to 621.04 MB (Baseline w/o IDPG), accompanied by a 42\% rise in GPU memory consumption (from 19 GB to 27 GB). Despite a marginal PSNR improvement (28.02 vs. 27.91), the SSIM decreases from 0.863 to 0.821, indicating a trade-off between model scale and structural fidelity. This demonstrates that IDGP effectively reduces redundant Gaussians, optimizing model efficiency while maintaining comparable reconstruction quality, thereby confirming its critical role in memory-efficient high-resolution 3D reconstruction on the Mip-NeRF 360 dataset.
To evaluate the impact of Importance-Driven Gaussian Pruning (IDGP), we performed ablation studies on the full-resolution Mip-NeRF 360 dataset by comparing the proposed method with a variant that excludes IDGP.
As shown in Tab.~\ref{tab6}, removing IDGP results in a substantial increase in model size from 313.72 MB to 621.04 MB, along with a 42\% rise in GPU memory usage (19 GB to 27 GB). Although PSNR slightly improves (28.02 vs. 27.91), the SSIM drops notably (0.821 vs. 0.863). These results underscore the effectiveness of IDGP in eliminating redundant Gaussians, significantly enhancing model efficiency while preserving high-quality reconstruction, thus highlighting its importance for memory-efficient 3D reconstruction in high-resolution settings.
\begin{table}[h]
\centering
\small
% \caption{\textbf{Ablation Studies on the Scene of the Full-Resolution Mip-NeRF 360 Dataset\cite{Mip-NeRF360}.} }
\caption{{\textbf{Ablation study on IDPG across the entire full resolution Mip-NeRF 360 Dataset\cite{Mip-NeRF360}.}} }
\label{tab:3}
\begin{tabular}{@{}lccccc@{}}
\toprule
\textbf{Method} & \textbf{Model Size(MB)} & \textbf{GPU Memory(G)} & \textbf{PSNR} & \textbf{SSIM} & \textbf{LPIPS} \\
\midrule
w/o IDPG & 621.04 & 27 & 28.02 & 0.821 & 0.308 \\
Ours (Full) & 313.72 & 19 & 27.91 & 0.863 & 0.342 \\
\bottomrule
\label{tab6}
\end{tabular}
\vspace{-0.2cm}
\end{table}
\subsection{Robustness Validation Across Resolution Scales}
\label{sec:res_scale}
% To comprehensively validate the robustness of our method across different resolution scales, we designed supplementary experiments integrating qualitative visual analysis and quantitative metric evaluation.
% %Specifically, these experiments aim to assess whether our method exhibits greater robustness than other approaches as resolution decreases, and to demonstrate the consistency of its structural fidelity under varying computational constraints.
% First, we conducted detailed visual comparisons of rendering details in the Mip-NeRF 360 bicycle scene at 2×, 4×, and 8× downsampling ratios Fig.~\ref{fig:4}.
% At each resolution scale, we visualized fine-grained structures such as bicycle spokes, tire textures, and background foliage to evaluate how our method preserves geometric and textural details as input resolution declines.
% The results show that even at an 8× downsampling ratio (1/8 of the original resolution), our method maintains discernible edge clarity and minimal aliasing, whereas comparative methods exhibit noticeable blurring and structural degradation.
% This highlights the method’s resilience in handling low-resolution inputs without compromising visual fidelity.
% Additionally, we performed a stepwise degradation analysis by systematically reducing the input resolution to 1/8 of the original scale and quantified performance using PSNR, SSIM, and LPIPS metrics Tab.~\ref{tab1}.
% Quantitative results show that our method achieves a PSNR of 29.39 at 1/2 resolution, 29.84 at 1/4 resolution, and 29.84 at 1/8 resolution, outperforming baselines such as 3DGS and Mip-Splatting across all scales.
% Notably, SSIM values remain above 0.884 even at the lowest resolution, indicating strong preservation of structural similarity.
% In contrast, competing methods show more pronounced declines in SSIM and increases in LPIPS (perceptual error), underscoring our method’s superior robustness to resolution variations.
% Collectively, these experiments demonstrate that our method maintains high-fidelity 3D reconstruction across a broad range of resolution scales, from full resolution to highly downsampled regimes, validating its robustness and adaptability to diverse computational and data input scenarios.
To thoroughly evaluate the robustness of our method across varying resolution scales, we performed detailed visual comparisons of rendering quality on the Mip-NeRF 360 dataset at multiple resolution levels.
As shown in Tab.~\ref{tab1}, we compare our method with several existing approaches, including mip-NeRF~\cite{Mip-NeRF360}, Instant-NGP~\cite{InstantNGP}, zip-NeRF~\cite{zip-nrf}, 3DGS~\cite{3dgs}, 3DGS+EWA~\cite{EWV}, and Mip-Splatting~\cite{mip360-splatting}.
Our approach demonstrates comparable performance to these prior methods at one-eighth of the original resolution.
Furthermore, at higher resolutions, our method significantly outperforms all state-of-the-art techniques.
As shown in Fig.~\ref{fig:4}, our method produces high-fidelity imagery devoid of fine-scale texture distortions.
While 3DGS~\cite{3dgs} introduces noticeable erosion artifacts due to dilation operations, Mip-Splatting~\cite{mip360-splatting} shows improved performance, yet still exhibits evident texture distortions. In contrast, our method avoids such issues, producing images that are both aesthetically pleasing and closely aligned with the ground truth, demonstrating the effectiveness of our hierarchicall refined strategy.
% \begin{table}[htbp]
% \centering
% \scriptsize
% \begin{tabular}{c|ccccc|c|ccccc}
% \hline
% \multirow{2}{*}{Scene} & \multicolumn{5}{c|}{\textbf{CityGS (MB)}} & \textbf{3DGS (MB)} & \multicolumn{5}{c}{\textbf{Ours (MB)}} \\
% & Total & cell0 & cell1 & cell2 & cell3 & Total & Total & cell0 & cell1 & cell2 & cell3 \\
% \hline
% bicycle & 1126.4 & 512.63 & 480.45 & 70.92 & 63.11 & \textcolor{cyan}{844.23} & \textcolor{orange}{587.05} & 187.22 & 238.32 & 100.76 & 61.75 \\
% bonsai & \textcolor{cyan}{278.52} & 49.97 & 43.83 & 117.54 & 67.18 & 343.29 & \textcolor{orange}{160.10} & 49.81 & 47.16 & 48.18 & 34.01 \\
% counter & \textcolor{cyan}{215.85} & 21.38 & 22.26 & 78.73 & 93.49 & 274.12 & \textcolor{orange}{144.58} & 23.06 & 36.63 & 36.36 & 45.54 \\
% flowers & 738.18 & 269.72 & 276.41 & 96.18 & 93.87 & \textcolor{cyan}{599.21} & \textcolor{orange}{399.71} & 163.86 & 133.96 & 65.55 & 63.33 \\
% garden & \textcolor{cyan}{1196.80} & 269.72 & 276.41 & 96.18 & 93.87 & 1259.52 & \textcolor{orange}{360.52} & 114.76 & 86.78 & 87.60 & 79.39 \\
% kitchen & \textcolor{cyan}{324.08} & 30.26 & 26.46 & 157.15 & 110.21 & 424.16 & \textcolor{orange}{170.94} & 38.41 & 35.78 & 76.55 & 47.52 \\
% stump & 873.41 & 399.37 & 302.45 & 171.59 & 191.51 & \textcolor{cyan}{609.14} & \textcolor{orange}{368.62} & 103.86 & 103.32 &80.75 & 80.69 \\
% treehill& 745.24 & 273.96 & 236.07 & 136.24 & 98.97 & \textcolor{cyan}{515.48} & \textcolor{orange}{385.84} & 109.68 & 149.32 & 66.55 & 60.29 \\
% room & \textcolor{cyan}{301.02} & 33.58 & 94.33 & 74.08 & 106.03 & 388.54 & \textcolor{orange}{221.97} &316.31 & 33.84 & 75.64 & 76.19 \\
% \hline
% \end{tabular}
% \caption{Comparison of model sizes (MB) across CityGS, 3DGS, and Ours. The smallest total model size for each scene is highlighted in orange; the second smallest is highlighted in cyan.}
% \label{tab:cell}
% \end{table}
\begin{table*}[!ht]
\caption{\textbf{Quantitative results on Mip-NeRF 360 (Downscaled Resolutions).} The best results are highlighted in \textbf{\textcolor{BurntOrange}{orange}}, while the second-best results are marked in \textbf{\textcolor{cyan}{blue}}. Our method demonstrates consistent superiority across different resolution scales.} % 标题移动到表格环境开头
\vspace{-0.3cm}
\label{tab1}
\centering % 保持居中
\vspace{0.2cm} % 可根据需要微调标题与表格间距
\scalebox{0.85}{
\begin{tabular}{lccc|ccc|ccc}
\hline
\textbf{} & \multicolumn{3}{c|}{\textbf{1/2x Res.}} & \multicolumn{3}{c|}{\textbf{1/4x Res.}} & \multicolumn{3}{c}{\textbf{1/8x Res.}} \\
\textbf{} & PSNR $\uparrow$ & SSIM $\uparrow$ & LPIPS $\downarrow$ & PSNR $\uparrow$ & SSIM $\uparrow$ & LPIPS $\downarrow$ & PSNR $\uparrow$ & SSIM $\uparrow$ & LPIPS $\downarrow$\\ \hline
mip-NeRF & 24.16 & 0.670 & 0.370 & 25.18 & 0.727 & 0.260 & 29.26 & 0.860 & 0.122 \\
Instant-NGP & 24.27 & 0.626 & 0.445 & 24.76 & 0.639 & 0.367 & 26.79 & 0.746 & 0.239 \\
zip-NeRF & 20.87 & 0.565 & 0.421 & 23.27 & 0.696 & 0.257 & \textbf{\textcolor{cyan}{29.66}} & 0.875 & \textbf{\textcolor{BurntOrange}{0.097}} \\
3DGS & 20.71 & 0.619 & 0.394 & 23.05 & 0.740 & 0.243 & 29.19 & 0.880 & \textbf{\textcolor{cyan}{0.107}} \\
3DGS+EWA & 23.70 & 0.667 & 0.369 & 25.90 & 0.775 & 0.236 & 29.30 & 0.880 & 0.114 \\
Mip-Splatting & \textbf{\textcolor{cyan}{26.47}} & \textbf{\textcolor{cyan}{0.754}} & \textbf{\textcolor{cyan}{0.305}} & \textbf{\textcolor{cyan}{27.39}} & \textbf{\textcolor{cyan}{0.808}} & \textbf{\textcolor{cyan}{0.205}} & 29.39 & \textbf{\textcolor{cyan}{0.884}} & 0.108 \\
Ours & \textbf{\textcolor{BurntOrange}{28.42}} & \textbf{\textcolor{BurntOrange}{0.868}} & \textbf{\textcolor{BurntOrange}{0.238}} & \textbf{\textcolor{BurntOrange}{29.30}} & \textbf{\textcolor{BurntOrange}{0.870}} & \textbf{\textcolor{BurntOrange}{0.146}} & \textbf{\textcolor{BurntOrange}{29.84}} & \textbf{\textcolor{BurntOrange}{0.893}} & 0.112 \\ \hline
\end{tabular}
}
\vspace{-0.3cm}
\end{table*}
\FloatBarrier
\begin{figure}[ht]
\centering
\centering \includegraphics[width=1.0\textwidth,height=0.46\textheight]{Fig5.pdf}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Qualitative results on the Mip-NeRF360 dataset.} Our method reconstructs surfaces with fine geometry details and produces high-fidelity renderings on Mip-NeRF360 dataset. }
\label{fig:a1}
\end{figure}
\section{Additional Qualitative Results}
% Fig.\ref{fig:a1} presents the rendering (top) and reconstruction (bottom) results on the Mip-NeRF360\cite{Mip-NeRF360} dataset. Additionally, the rendering results for the TNT~\cite{TNT} and Replica~\cite{replica} datasets are also shown in Fig.~\ref{fig:a2}. From these results, it is evident that our method achieves high-quality 3D reconstruction performance.
As shown in Fig.\ref{fig:1}, our method preserves fine-scale structures more accurately and achieves higher visual fidelity than 3DGS\cite{3dgs} and Mip-Splatting~\cite{mip360-splatting} across three representative scenes.
Fig.~\ref{fig:a1} displays the rendering (top) and surface reconstruction (bottom) results on the Mip-NeRF360~\cite{Mip-NeRF360} dataset.
Additional rendering results on the TNT~\cite{TNT} and Replica~\cite{replica} datasets are provided in Fig.~\ref{fig:a2}. Collectively, these visual comparisons substantiate our method's capability for high-quality 3D reconstruction while maintaining critical geometric details.
% \begin{figure*}[h]
% \centering
% \centering \includegraphics[width=1.0\textwidth,height=0.65\textheight]{bicycle.pdf}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{\textbf{Qualitative comparsion on Mip-NeRF 360 for different resolution scales.} }
% \label{fig:4}
% \end{figure*}
% \begin{figure}[t]
% \centering
% \centering \includegraphics[width=1.0\textwidth,height=0.5\textheight]{Fig5.pdf}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{\textbf{Qualitative results on the Mip-NeRF360 dataset.} Our method reconstructs surfaces with fine geometry details and produces high-fidelity renderings on Mip-NeRF360 dataset. }
% \label{fig:a1}
% \end{figure}
% \begin{figure}[t]
% \centering
% \centering \includegraphics[width=1.0\textwidth,height=0.5\textheight]{Fig6.png}
% % \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% % \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
% \caption{\textbf{Qualitative rendering results on TNT and Replica dataset.} }
% \label{fig:a2}
% \end{figure}
% \begin{table}[t]
% \centering
% \caption{\textbf{Mip-NeRF 360 Full-Resolution Results (Simplified)}. Best and second-best results are highlighted.}
% \label{tab:mipnerf_simple}
% \setlength{\tabcolsep}{4pt}
% \begin{tabular}{@{}lccccc@{}}
% \toprule
% & 3DGS & CityGS & Mip-Splat & \textbf{Ours} \\
% \midrule
% PSNR $\uparrow$ & 19.59 & 24.71 & \textcolor{cyan}{26.22} & \textcolor{BurntOrange}{27.91} \\
% SSIM $\uparrow$ & 0.619 & 0.702 & \textcolor{cyan}{0.765} & \textcolor{BurntOrange}{0.863} \\
% LPIPS $\downarrow$ & 0.476 & 0.401 & \textcolor{cyan}{0.392} & \textcolor{BurntOrange}{0.342} \\
% \bottomrule
% \end{tabular}
% \label{tab:3}
% \end{table}
\FloatBarrier
\begin{figure}[ht]
\centering
\centering \includegraphics[width=1.0\textwidth,height=0.5\textheight]{Fig6.png}
% \fbox{\rule{0pt}{2in} \rule{0.9\linewidth}{0pt}}
% \includegraphics[width=0.8\linewidth, color=gray]{egfigure.eps}
\caption{\textbf{Qualitative rendering results on TNT and Replica dataset.} }
\label{fig:a2}
\end{figure}
\section{Limitations}
%One limitation of our method is that while it achieves high-resolution scene reconstruction under limited GPU resources, it incurs an increase in training time. Consequently, although our approach is resource-efficient, its extended training duration may hinder rapid prototyping or deployment in time-sensitive applications. Future work should therefore explore strategies to accelerate training without sacrificing reconstruction quality.
Although our method can achieve high-resolution scene reconstruction under limited GPU resources, extending the block-based framework to dynamic scenes introduces new challenges. In particular, ensuring temporal consistency across frames and accurately modeling motion across spatial partitions remain open problems.
Addressing these challenges will be a key focus of our future work, with the goal of enabling temporally coherent and spatially consistent reconstruction in dynamic environments.
\section{Broader Social Impacts}
\label{sec:impact}
% This work introduces Hierarchical Gaussian Splatting (HRGS), a method designed to perform super-resolution image reconstruction under constrained computational budgets. HRGS attains high-fidelity 3D reconstruction while markedly decreasing GPU memory consumption and overall model size, thereby broadening the accessibility and practical reach of advanced 3D vision techniques. Such gains in resource efficiency facilitate deployment on low-power platforms—including mobile and embedded devices—opening avenues for more cost-effective applications in education, cultural-heritage preservation, smart-city visualization, and virtual or augmented-reality systems.
% At the same time, we recognize that HRGS could be exploited for unauthorized spatial reconstruction or other privacy-sensitive use cases. Consequently, its dissemination and deployment should be guided by explicit ethical principles and regulatory safeguards to ensure that the technology serves the public interest and minimizes the risk of misuse. Future research should therefore prioritize privacy-preserving mechanisms, data-security protocols, and equitable access frameworks to promote the responsible and sustainable adoption of HRGS in society.
This work presents Hierarchical Gaussian Splatting (HRGS), a super-resolution reconstruction framework tailored for computationally constrained environments. HRGS achieves high-fidelity 3D reconstruction while significantly reducing GPU memory usage and model size, thereby enhancing the accessibility of advanced 3D vision techniques. Its efficiency enables deployment on low-power platforms such as mobile and embedded devices, with promising applications in education, cultural heritage preservation, smart city visualization, and immersive virtual or augmented reality.
However, we acknowledge the potential for misuse in privacy-sensitive contexts, such as unauthorized spatial reconstruction. To mitigate such risks, the deployment of HRGS should be governed by clear ethical guidelines and regulatory oversight.
\begin{thebibliography}{10}\itemsep=-1pt
\bibitem{YAN}
Deep multi-view learning methods: A review.
\newblock {\em Neurocomputing}, 448:106--129, 2021.
\bibitem{bae2024dsine}
Gwangbin Bae and Andrew~J. Davison.
\newblock Rethinking inductive biases for surface normal estimation.
\newblock In {\em CVPR}, 2024.
\bibitem{BaeD24}
Gwangbin Bae and Andrew~J. Davison.
\newblock Rethinking inductive biases for surface normal estimation.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition 2024, Seattle, WA, USA, June 16-22, 2024}, pages 9535--9545, 2024.
\bibitem{re_matching1}
Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan~B. Goldman.
\newblock Patchmatch: a randomized correspondence algorithm for structural image editing.
\newblock {\em {ACM} Trans. Graph.}, 28(3):24, 2009.
\bibitem{Mip-NeRF360}
Jonathan~T. Barron, Ben Mildenhall, Dor Verbin, Pratul~P. Srinivasan, and Peter Hedman.
\newblock Mip-nerf 360: Unbounded anti-aliased neural radiance fields.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, {CVPR} 2022, New Orleans, LA, USA, June 18-24, 2022}, pages 5460--5469. {IEEE}, 2022.
\bibitem{mip}
Jonathan~T Barron, Ben Mildenhall, Dor Verbin, Pratul~P Srinivasan, and Peter Hedman.
\newblock Mip-nerf 360: Unbounded anti-aliased neural radiance fields.
\newblock In {\em Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, pages 5470--5479, 2022.
\bibitem{zip-nrf}
Jonathan~T. Barron, Ben Mildenhall, Dor Verbin, Pratul~P. Srinivasan, and Peter Hedman.
\newblock Zip-nerf: Anti-aliased grid-based neural radiance fields.
\newblock In {\em {IEEE/CVF} International Conference on Computer Vision, {ICCV} 2023, Paris, France, October 1-6, 2023}, pages 19640--19648. {IEEE}, 2023.
\bibitem{re_MVS1}
Michael Bleyer, Christoph Rhemann, and Carsten Rother.
\newblock Patchmatch stereo - stereo matching with slanted support windows.
\newblock In {\em British Machine Vision Conference, {BMVC} 2011, Dundee, UK, August 29 - September 2, 2011. Proceedings}, pages 1--11. {BMVA} Press, 2011.
\bibitem{re_depth1}
Michael Bleyer, Christoph Rhemann, and Carsten Rother.
\newblock Patchmatch stereo - stereo matching with slanted support windows.
\newblock In Jesse Hoey, Stephen~J. McKenna, and Emanuele Trucco, editors, {\em British Machine Vision Conference, {BMVC} 2011, Dundee, UK, August 29 - September 2, 2011. Proceedings}, pages 1--11. {BMVA} Press, 2011.
\bibitem{re_MVS2}
Adrian Broadhurst, Tom Drummond, and Roberto Cipolla.
\newblock A probabilistic framework for space carving.
\newblock In {\em Proceedings of the Eighth International Conference On Computer Vision (ICCV-01), Vancouver, British Columbia, Canada, July 7-14, 2001 - Volume 1}, pages 388--393. {IEEE} Computer Society, 2001.
\bibitem{tensorf}
Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su.
\newblock Tensorf: Tensorial radiance fields.
\newblock In {\em Computer Vision - {ECCV} 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part {XXXII}}, volume 13692, pages 333--350. Springer, 2022.
\bibitem{neusg}
Hanlin Chen, Chen Li, and Gim~Hee Lee.
\newblock Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance.
\newblock {\em CoRR}, abs/2312.00846, 2023.
\bibitem{vcR}
Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim~Hee Lee.
\newblock Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction.
\newblock In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub~M. Tomczak, and Cheng Zhang, editors, {\em Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, Vancouver, BC, Canada, December 10 - 15, 2024}, 2024.
\bibitem{vcr-gaus}
Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim~Hee Lee.
\newblock Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction.
\newblock In {\em Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024}, 2024.
\bibitem{learining}
Zhiqin Chen and Hao Zhang.
\newblock Learning implicit fields for generative shape modeling.
\newblock In {\em {IEEE} Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 16-20, 2019}, pages 5939--5948. Computer Vision Foundation / {IEEE}, 2019.
\bibitem{volume}
Robert~A. Drebin, Loren Carpenter, and Pat Hanrahan.
\newblock Volume rendering.
\newblock In {\em Proceedings of the 15th Annual Conference on Computer Graphics and Interactive Techniques}, SIGGRAPH '88, page 65–74, 1988.
\bibitem{plenoxels}
Sara Fridovich{-}Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa.
\newblock Plenoxels: Radiance fields without neural networks.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, {CVPR} 2022, New Orleans, LA, USA, June 18-24, 2022}, pages 5491--5500. {IEEE}, 2022.
\bibitem{Geo-Neus}
Qiancheng Fu, Qingshan Xu, Yew~Soon Ong, and Wenbing Tao.
\newblock Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction.
\newblock In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, {\em Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, New Orleans, LA, USA, November 28 - December 9, 2022}, 2022.
\bibitem{GeoWizard}
Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long.
\newblock Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image.
\newblock In Ales Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G{\"{u}}l Varol, editors, {\em Computer Vision - {ECCV} 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part {XXII}}, volume 15080 of {\em Lecture Notes in Computer Science}, pages 241--258. Springer, 2024.
\bibitem{re_point}
Yasutaka Furukawa and Jean Ponce.
\newblock Accurate, dense, and robust multiview stereopsis.
\newblock {\em {IEEE} Trans. Pattern Anal. Mach. Intell.}, 32(8):1362--1376, 2010.
\bibitem{lum}
Steven~J. Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael~F. Cohen.
\newblock The lumigraph.
\newblock In {\em Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, {SIGGRAPH} 1996, New Orleans, LA, USA, August 4-9, 1996}, pages 43--54. {ACM}, 1996.
\bibitem{SuGaR}
Antoine Gu{\'{e}}don and Vincent Lepetit.
\newblock Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 16-22, 2024}, pages 5354--5363. {IEEE}, 2024.
\bibitem{2Dgs}
Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao.
\newblock 2d gaussian splatting for geometrically accurate radiance fields.
\newblock In Andres Burbano, Denis Zorin, and Wojciech Jarosz, editors, {\em {ACM} {SIGGRAPH} 2024 Conference Papers, Denver, CO, USA, 27 July 2024- 1 August 2024}, page~32. {ACM}, 2024.
\bibitem{Tri}
Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu.
\newblock Tri-perspective view for vision-based 3d semantic occupancy prediction.
\newblock In {\em Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages 9223--9232, June 2023.
\bibitem{re_fit}
Michael Kazhdan and Hugues Hoppe.
\newblock Distributed poisson surface reconstruction.
\newblock {\em Comput. Graph. Forum}, 42(6), 2023.
\bibitem{3dgs}
Bernhard Kerbl, Georgios Kopanas, Thomas Leimk{\"{u}}hler, and George Drettakis.
\newblock 3d gaussian splatting for real-time radiance field rendering.
\newblock {\em {ACM} Trans. Graph.}, 42(4):139:1--139:14, 2023.
\bibitem{TNT}
Arno Knapitsch, Jaesik Park, Qian{-}Yi Zhou, and Vladlen Koltun.
\newblock Tanks and temples: benchmarking large-scale scene reconstruction.
\newblock {\em {ACM} Trans. Graph.}, 36(4):78:1--78:13, 2017.
\bibitem{re_MVS3}
Kiriakos~N. Kutulakos and Steven~M. Seitz.
\newblock A theory of shape by space carving.
\newblock {\em Int. J. Comput. Vis.}, 38(3):199--218, 2000.
\bibitem{eff}
Marc Levoy.
\newblock Efficient ray tracing of volume data.
\newblock {\em {ACM} Trans. Graph.}, 9(3):245--261, 1990.
\bibitem{light}
Marc Levoy and Pat Hanrahan.
\newblock Light field rendering.
\newblock In {\em Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques}, SIGGRAPH '96, page 31–42, 1996.
\bibitem{re_point2}
Maxime Lhuillier and Long Quan.
\newblock A quasi-dense approach to surface reconstruction from uncalibrated images.
\newblock {\em {IEEE} Trans. Pattern Anal. Mach. Intell.}, 27(3):418--433, 2005.
\bibitem{binary}
Anthony Lin.
\newblock Binary search algorithm.
\newblock {\em WikiJournal of Science}, 2(1):1--13, 2019.
\bibitem{vast}
Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang.
\newblock Vastgaussian: Vast 3d gaussians for large scene reconstruction.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, Seattle, WA, USA, June 16-22, 2024}, pages 5166--5175. {IEEE}, 2024.
\bibitem{Neural}
Lingjie Liu, Jiatao Gu, Kyaw~Zaw Lin, Tat{-}Seng Chua, and Christian Theobalt.
\newblock Neural sparse voxel fields.
\newblock In {\em Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual}, 2020.
\bibitem{Citygs}
Yang Liu, Chuanchen Luo, Lue Fan, Naiyan Wang, Junran Peng, and Zhaoxiang Zhang.
\newblock Citygaussian: Real-time high-quality large-scale scene rendering with gaussians.
\newblock In {\em Computer Vision - {ECCV} 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part {XVI}}, pages 265--282, 2024.
\bibitem{nerf}
Ben Mildenhall, Pratul~P. Srinivasan, Matthew Tancik, Jonathan~T. Barron, Ravi Ramamoorthi, and Ren Ng.
\newblock Nerf: Representing scenes as neural radiance fields for view synthesis.
\newblock In {\em Computer Vision - {ECCV} 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part {I}}, 2020.
\bibitem{re_nerf}
Ben Mildenhall, Pratul~P. Srinivasan, Matthew Tancik, Jonathan~T. Barron, Ravi Ramamoorthi, and Ren Ng.
\newblock Nerf: representing scenes as neural radiance fields for view synthesis.
\newblock {\em Commun. {ACM}}, 65(1):99--106, 2022.
\bibitem{Instant_neural}
Thomas M{\"{u}}ller, Alex Evans, Christoph Schied, and Alexander Keller.
\newblock Instant neural graphics primitives with a multiresolution hash encoding.
\newblock {\em {ACM} Trans. Graph.}, 41(4):102:1--102:15, 2022.
\bibitem{Instant}
Thomas M{\"{u}}ller, Alex Evans, Christoph Schied, and Alexander Keller.
\newblock Instant neural graphics primitives with a multiresolution hash encoding.
\newblock {\em {ACM} Trans. Graph.}, 41(4):102:1--102:15, 2022.
\bibitem{InstantNGP}
Thomas M{\"{u}}ller, Alex Evans, Christoph Schied, and Alexander Keller.
\newblock Instant neural graphics primitives with a multiresolution hash encoding.
\newblock {\em {ACM} Trans. Graph.}, 41(4):102:1--102:15, 2022.
\bibitem{deep}
Jeong~Joon Park, Peter~R. Florence, Julian Straub, Richard~A. Newcombe, and Steven Lovegrove.
\newblock Deepsdf: Learning continuous signed distance functions for shape representation.
\newblock In {\em {IEEE} Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, June 16-20, 2019}, pages 165--174. Computer Vision Foundation / {IEEE}, 2019.
\bibitem{merf}
Christian Reiser, Richard Szeliski, Dor Verbin, Pratul~P. Srinivasan, Ben Mildenhall, Andreas Geiger, Jonathan~T. Barron, and Peter Hedman.
\newblock {MERF:} memory-efficient radiance fields for real-time view synthesis in unbounded scenes.
\newblock {\em {ACM} Trans. Graph.}, 42(4):89:1--89:12, 2023.
\bibitem{sfm}
Johannes~L. Schonberger and Jan-Michael Frahm.
\newblock Structure-from-motion revisited.
\newblock In {\em Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition}, June 2016.
\bibitem{schoenberger2016sfm}
Johannes~Lutz Sch\"{o}nberger and Jan-Michael Frahm.
\newblock Structure-from-motion revisited.
\newblock In {\em Conference on Computer Vision and Pattern Recognition (CVPR)}, 2016.
\bibitem{re_MVS4}
Johannes~L. Sch{\"{o}}nberger, Enliang Zheng, Jan{-}Michael Frahm, and Marc Pollefeys.
\newblock Pixelwise view selection for unstructured multi-view stereo.
\newblock In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, {\em Computer Vision - {ECCV} 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part {III}}, volume 9907 of {\em Lecture Notes in Computer Science}, pages 501--518. Springer, 2016.
\bibitem{re_depth2}
Johannes~L. Sch{\"{o}}nberger, Enliang Zheng, Jan{-}Michael Frahm, and Marc Pollefeys.
\newblock Pixelwise view selection for unstructured multi-view stereo.
\newblock In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, {\em Computer Vision - {ECCV} 2016 - 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part {III}}, volume 9907 of {\em Lecture Notes in Computer Science}, pages 501--518. Springer, 2016.
\bibitem{schoenberger2016mvs}
Johannes~Lutz Sch\"{o}nberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm.
\newblock Pixelwise view selection for unstructured multi-view stereo.
\newblock In {\em European Conference on Computer Vision (ECCV)}, 2016.
\bibitem{re_MVS6}
Steven~M. Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski.
\newblock A comparison and evaluation of multi-view stereo reconstruction algorithms.
\newblock In {\em 2006 {IEEE} Computer Society Conference on Computer Vision and Pattern Recognition {(CVPR} 2006), 17-22 June 2006, New York, NY, {USA}}, pages 519--528. {IEEE} Computer Society, 2006.
\bibitem{re_MVS5}
Steven~M. Seitz and Charles~R. Dyer.
\newblock Photorealistic scene reconstruction by voxel coloring.
\newblock {\em Int. J. Comput. Vis.}, 35(2):151--173, 1999.
\bibitem{replica}
Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob~J. Engel, Raul Mur{-}Artal, Carl~Yuheng Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke~M. Strasdat, Renzo {De Nardi}, Michael Goesele, Steven Lovegrove, and Richard~A. Newcombe.
\newblock The replica dataset: {A} digital replica of indoor spaces.
\newblock {\em CoRR}, abs/1906.05797, 2019.
\bibitem{dir}
Cheng Sun, Min Sun, and Hwann{-}Tzong Chen.
\newblock Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, New Orleans, LA, USA, June 18-24, 2022}, pages 5449--5459. {IEEE}.
\bibitem{turkulainen2024dnsplatter}
Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala.
\newblock Dn-splatter: Depth and normal priors for gaussian splatting and meshing.
\newblock In {\em CVPR}, 2024.
\bibitem{NeuS}
Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang.
\newblock Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction.
\newblock In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann~N. Dauphin, Percy Liang, and Jennifer~Wortman Vaughan, editors, {\em Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, December 6-14, 2021, virtual}, pages 27171--27183, 2021.
\bibitem{mss}
Z. Wang, E.P. Simoncelli, and A.C. Bovik.
\newblock Multiscale structural similarity for image quality assessment.
\newblock In {\em The Thrity-Seventh Asilomar Conference on Signals, Systems \& Computers, 2003}, volume~2, pages 1398--1402 Vol.2, 2003.
\bibitem{scanerf}
Xiuchao Wu, Jiamin Xu, Xin Zhang, Hujun Bao, Qixing Huang, Yujun Shen, James Tompkin, and Weiwei Xu.
\newblock Scanerf: Scalable bundle-adjusting neural radiance fields for large-scale scene rendering.
\newblock {\em {ACM} Trans. Graph.}, 42(6):261:1--261:18, 2023.
\bibitem{bake}
Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul~P. Srinivasan, Richard Szeliski, Jonathan~T. Barron, and Ben Mildenhall.
\newblock Bakedsdf: Meshing neural sdfs for real-time view synthesis.
\newblock In {\em {ACM} {SIGGRAPH} 2023 Conference Proceedings, {SIGGRAPH} 2023, Los Angeles, CA, USA, August 6-10, 2023}, pages 46:1--46:9. {ACM}, 2023.
\bibitem{yifan2019differentiable}
Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung.
\newblock Differentiable surface splatting for point-based geometry processing.
\newblock {\em ACM Transactions on Graphics}, 38(6):1--14, 2019.
\bibitem{yin2019enforcing}
Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan.
\newblock Enforcing geometric constraints of virtual normal for depth prediction.
\newblock In {\em ICCV}, pages 5684--5693, 2019.
\bibitem{mip360-splatting}
Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger.
\newblock Mip-splatting: Alias-free 3d gaussian splatting.
\newblock In {\em {IEEE/CVF} Conference on Computer Vision and Pattern Recognition, {CVPR} 2024, Seattle, WA, USA, June 16-22, 2024}, pages 19447--19456. {IEEE}, 2024.
\bibitem{MonoSDF}
Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger.
\newblock Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction.
\newblock In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh, editors, {\em Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, New Orleans, LA, USA, November 28 - December 9, 2022}, 2022.
\bibitem{Open3D}
Qian{-}Yi Zhou, Jaesik Park, and Vladlen Koltun.
\newblock Open3d: {A} modern library for 3d data processing.
\newblock {\em CoRR}, abs/1801.09847, 2018.
\bibitem{EWV}
Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus~H. Gross.
\newblock {EWA} volume splatting.
\newblock In Thomas Ertl, Kenneth~I. Joy, and Amitabh Varshney, editors, {\em 12th {IEEE} Visualization Conference, {IEEE} Vis 2001, San Diego, CA, USA, October 24-26, 2001, Proceedings}, pages 29--36. {IEEE} Computer Society, 2001.
\end{thebibliography}