betaFeedback

WorksComputer Science

DepthGS: Depth-Guided Gaussian Splatting with Adaptive Regularization for High-Quality Surface Reconstruction

Abstract

Recently, 3D Gaussian Splatting (3DGS) has emerged as an efficient representation for novel view synthesis, offering fast differentiable rendering and high visual fidelity, and has increasingly been extended to surface reconstruction. However, the relationship between depth information and reconstruction error remains insufficiently explored, limiting the principled design of geometric optimization strategies. To fill the above gap, we propose DepthGS, a depth-guided Gaussian splatting framework. We derive theoretical relations between depth discrepancies and reconstruction error under single-view and multi-view settings, and rigorously establish corresponding upper bounds on reconstruction error. The above analysis informs the design of adaptive regularization strategies in both single-view and multi-view optimization. More specifically, for single-view optimization, we propose a depth-difference-guided adaptive normal regularization strategy that modulates the normal constraint strength using a geometric reliability weight computed from local depth differences. For multi-view optimization, we introduce a cross-view depth consistency regularization strategy that promotes consistency between neighboring-view depth maps and the corresponding depth predictions obtained by transforming the current view's depth predictions into those neighboring views. Extensive experiments on DTU, Mip-NeRF360, and Tanks and Temples demonstrate that DepthGS achieves competitive performance in both surface reconstruction and novel view synthesis.

INTRODUCTION

Multi-view 3D reconstruction plays a crucial role in computer vision, with applications ranging from robotics [1, 2, 3] and autonomous driving [4, 5, 6] to augmented and virtual reality [7, 8, 9]. Neural radiance fields (NeRFs) [10, 11, 12] represent scenes using continuous radiance fields and achieve high-quality novel view synthesis and geometric reconstruction. However, their computationally expensive volume rendering limits real-time applications. In contrast, 3D Gaussian Splatting (3DGS) [13] represents scenes using explicit anisotropic Gaussians and enables efficient optimization and real-time high-fidelity rendering through differentiable α\alpha-blending.

Despite its advances in rendering speed and quality, 3DGS still faces challenges in accurate surface reconstruction. The inherent irregularity of Gaussian primitives makes it difficult to precisely fit real-world surfaces, which may lead to geometric artifacts [14].

Geometry-aware approaches [15, 16] improve Gaussian representations through additional geometric constraints, while recent methods incorporating depth and normal cues [17, 18, 19] demonstrate the effectiveness of such geometric information for surface reconstruction. Nevertheless, the reconstruction errors of the three-dimensional points obtained from unbiased depth and rendered distance through inverse projection have not been explicitly characterized with respect to the corresponding ground-truth surface points. Moreover, the cross-view consistency of these two depth representations remains underexplored. These limitations motivate us to derive reconstruction-error bounds and use the resulting depth discrepancies to guide adaptive geometric regularization.

In this paper, we propose DepthGS, a depth-guided Gaussian splatting framework with adaptive regularization for high-quality surface reconstruction. Through mathematical analysis, we investigate the relationship between depth information and reconstruction error and derive the corresponding upper bounds. Based on the single-view analysis, we propose a depth-difference-guided adaptive normal regularization strategy. For multi-view scenes, we propose a cross-view depth consistency regularization strategy to constrain depth discrepancies between neighboring views and improve multi-view geometric consistency. Our main contributions are summarized as follows:

  • Through mathematical analysis, we derive the relationships between depth discrepancies and reconstruction error under single-view and multi-view conditions, together with the upper bounds for reconstruction error, providing theoretical guidance for the subsequent optimization strategies.

  • Based on the single-view error analysis, we propose a depth-difference-guided adaptive normal regularization strategy that modulates the normal constraint strength ac

This work was supported by National Key R&D Program of China under Grant No. 2024YFB4710900, the National Natural Science Foundation of China under Grant No. U24A20265, the Shenzhen Science and Technology Program under Grant KQTD20221101093557010, and the Guangdong Science and Technology Program under Grant 2024B1212010002. Corresponding author: He Kong.

Xiao Ren, Yu Liu, and He Kong are with School of Automation and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen 518055, China; Emails: [12431359,12250044]@mail.sustech.edu.cn; [email protected]. Ning An and Jian Cheng are with the Research Institute of Mine Artificial Intelligence, China Coal Research Institute, and the State Key Laboratory of Intelligent Coal Mining and Strata Control, Beijing 100031, China; Emails: [email protected]; [email protected]. Xin Qiao is with the Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an 710049, China; Email: [email protected]. You-Fu Li is with the Department of Mechanical Engineering, City University of Hong Kong, Hong Kong SAR, China. Email: [email protected]. cording to the local depth difference. Specifically, weaker constraints are applied to regions with larger depth differences, whereas stronger constraints are applied to regions with smaller depth differences.

  • Based on the multi-view error analysis, we propose a cross-view depth consistency regularization strategy that reduces depth discrepancies between neighboring views, thereby improving multi-view geometric consistency.

  • Extensive experiments on Mip-NeRF360 [20], DTU [21] and Tanks and Temples [22] demonstrate that DepthGS achieves competitive performance in both novel view synthesis and surface reconstruction.

RELATED WORK

Novel View Synthesis

Novel view synthesis aims to render unseen views while preserving scene structure and visual details. Classical Structure-from-Motion and Multi-View Stereo methods recover scene geometry through multi-view correspondences and epipolar constraints [23]. Soft3D [24] introduces a differentiable density-field representation for view synthesis, while NeRF and its variants represent scenes using continuous radiance fields and achieve high-quality rendering [25], [26]. However, their volumetric formulations generally incur considerable computational costs.

3DGS provides an efficient alternative by representing scenes using explicit anisotropic Gaussians and differentiable rasterization [27], [28]. Subsequent approaches improve representation compactness and rendering efficiency [29]–[31], enhance structural integrity and large-scale rendering [32], [33], and extend Gaussian representations to dynamic scenes [34], [35]. Nevertheless, improvements in rendering quality and efficiency do not necessarily ensure accurate and view-consistent surface geometry.

Gaussian Splatting Surface Reconstruction

With the increasing use of Gaussian splatting in novel view synthesis, its potential for surface reconstruction has also been explored. 2DGS [36] replaces volumetric Gaussian ellipsoids with planar Gaussian disks to achieve more consistent geometric representations, while GOF [37] extracts surfaces from Gaussian opacity fields.

Recent approaches further improve reconstruction through geometry-oriented representations and surface constraints. GF-SGS [38] and GSDF [39] introduce geometry-aware fields and distance representations to regularize Gaussian optimization. GS-Pull [40] and GS-UDF [41] infer signed or unsigned distance functions from Gaussian primitives. QGS [42] models surfaces using second-order geometric primitives, while GSPrior [43] incorporates self-constrained geometric priors for high-fidelity surface reconstruction.

Methods incorporating depth and normal cues provide complementary geometric information. DN-Splatter [17] incorporates depth and normal priors for Gaussian splatting and meshing. VCR-GaS [19] couples depth and normal information through a view-consistent regularizer. PGSR [18] introduces unbiased-depth rendering and multi-view regularization for efficient and high-fidelity surface reconstruction.

These studies demonstrate the effectiveness of depth and normal cues for geometric regularization. However, existing Gaussian splatting methods incorporating depth and normal cues do not explicitly derive reconstruction-error bounds that involve the discrepancy between unbiased depth and rendered distance. Moreover, the multi-view consistency of these two depth representations across neighboring views remains under-explored. DepthGS addresses these limitations by deriving the relationships between depth discrepancies and reconstruction errors under single-view and multi-view conditions. Based on this analysis, we introduce a depth-difference-guided adaptive normal regularization strategy and a cross-view depth consistency regularization strategy.

Preliminaries on 3D Gaussian Splatting

Following 3DGS [13], a scene is represented by a set of anisotropic Gaussian primitives. The ii-th Gaussian is defined by its center μi\mu_i and covariance matrix Σi\Sigma_i:

G(μ;μi,Σi)=exp⁡[−12(μ−μi)TΣi−1(μ−μi)].(1)G(\mu;\mu_i,\Sigma_i)=\exp\left[-\frac{1}{2}(\mu-\mu_i)^{\mathsf{T}}\Sigma_i^{-1}(\mu-\mu_i)\right]. \tag*{(1)}

To ensure positive semidefiniteness, the covariance is parameterized as Σi=RiSiSiTRiT\Sigma_i=R_iS_iS_i^{\mathsf{T}}R_i^{\mathsf{T}}, where RiR_i and SiS_i denote the rotation and scaling matrices, respectively. Under a given camera view, the Gaussians are projected onto the image plane and composited in depth order through differentiable α\alpha-blending.

At pixel pp, the rendered color and normal are computed as

C^(p)=∑i∈NTiαici,N(p)=∑i∈NTiαiRcTni,(2)\widehat{C}(p)=\sum_{i\in N}T_i\alpha_i c_i,\qquad N(p)=\sum_{i\in N}T_i\alpha_iR_c^{\mathsf{T}}n_i, \tag*{(2)}

where Ti=∏j=1i−1(1−αj)T_i=\prod_{j=1}^{i-1}(1-\alpha_j) denotes the accumulated transmittance, αi\alpha_i is the contribution weight of the ii-th Gaussian, cic_i denotes its view-dependent appearance, and NN is the ordered set of Gaussians contributing to pixel pp. Moreover, nin_i denotes the Gaussian normal, and RcTniR_c^{\mathsf{T}}n_i denotes the corresponding normal expressed in the camera coordinate system.

Following PGSR [18], the renderer further produces the rendered distance and unbiased depth:

DD(p)=∑i∈NTiαidi,Du(p)=DD(p)N(p)⋅v(p).(3)D_{\mathcal{D}}(p)=\sum_{i\in N}T_i\alpha_i d_i,\qquad D_u(p)=\frac{D_{\mathcal{D}}(p)}{N(p)\cdot v(p)}. \tag*{(3)}

Here, di=(RcTni)T(μi−Tc)d_i=(R_c^{\mathsf{T}}n_i)^{\mathsf{T}}(\mu_i-T_c) denotes the plane-to-camera distance of the ii-th Gaussian, and TcT_c is the camera center. For pixel p=(u,v)p=(u,v) with homogeneous coordinate p~=[u,v,1]T\widetilde{p}=[u,v,1]^{\mathsf{T}}, v(p)=K−1p~/∥K−1p~∥2v(p)=K^{-1}\widetilde{p}/\lVert K^{-1}\widetilde{p}\rVert_2 is the corresponding unit camera-ray direction, where KK denotes the camera intrinsic matrix.

METHODOLOGY

The framework proposed in this paper comprises two major components, which address geometric constraint optimization under single-view and multi-view conditions, respectively. The overall framework of our algorithm is depicted in Figure 1.

Figure 1. Overview of the proposed framework. The Gaussian representation is initialized from a sparse point cloud and used to render the distance and normal maps. Unbiased depth maps are then derived from these rendered quantities. The discrepancy between the unbiased depth and rendered distance is used to compute the depth-difference-based weight WΔDW_{\Delta D}, which modulates the single-view normal constraint Lsvgeo\mathcal{L}_{svgeo}. In the multi-view branch, coordinate transformations establish depth correspondences between neighboring views to construct the cross-view depth consistency constraint Lmv_D\mathcal{L}_{mv\_D}. The proposed constraints are jointly optimized with the baseline objectives to improve surface reconstruction and multi-view geometric consistency.

Single-View Geometric Constraint Optimization

Single-View Depth Difference-Reconstruction Error Relationship

1) Single-View Depth Difference-Reconstruction Error Relationship: As shown in Eq. (3), Gaussian splatting produces an unbiased depth DuD_u and a rendered distance DDD_{\mathcal{D}}. The corresponding three-dimensional points are obtained from DuD_u and DDD_{\mathcal{D}} through inverse projection along the same camera ray. Their Euclidean distance reflects the discrepancy between the two depth representations. As illustrated in Figure 2, large depth differences are often observed in regions with irregular Gaussian distributions.

Depth differences and corresponding Gaussian distributions

Figure 2. Depth differences (left) and corresponding Gaussian distributions (right); large discrepancies often coincide with irregular Gaussian layouts.

To formalize this observation, we analyze the reconstruction errors of the corresponding three-dimensional points obtained from DuD_u and DDD_{\mathcal{D}} with respect to the ground-truth surface point. We first derive the single-view relationship between the depth difference and reconstruction error and then extend the analysis to the multi-view case.

Lemma 1. In a single-view scene, the aggregate reconstruction error associated with the unbiased depth satisfies

Esv≜∫Ωw(p)(ϵusv(p))2 dp≤∫Ωw(p)(ϵDsv(p)+∣ΔD(p)∣)2 dp.(4)E_{sv}\triangleq\int_{\Omega}w(p)\left(\epsilon_u^{sv}(p)\right)^2\,dp \leq\int_{\Omega}w(p)\left(\epsilon_{\mathcal{D}}^{sv}(p)+\left|\Delta D(p)\right|\right)^2\,dp. \tag*{(4)}

where w(p)≥0w(p)\geq0, Ω\Omega denotes the valid image domain, and ϵusv(p)\epsilon_u^{sv}(p) and ϵDsv(p)\epsilon_{\mathcal{D}}^{sv}(p) denote the reconstruction errors associated with the unbiased depth and rendered distance, respectively.

Proof. The depth difference is given by

ΔD(p)=Du(p)−DD(p)=DD(p)N(p)⋅v(p)−DD(p).(5)\Delta D(p)=D_u(p)-D_{\mathcal{D}}(p)=\frac{D_{\mathcal{D}}(p)}{N(p)\cdot v(p)}-D_{\mathcal{D}}(p). \tag*{(5)}

The corresponding three-dimensional points obtained through inverse projection are denoted by XuX_u and XDX_{\mathcal{D}}, respectively:

Xu=π−1(p,Du(p))=DD(p)N(p)⋅v(p)v(p),(6)X_u=\pi^{-1}(p,D_u(p))=\frac{D_{\mathcal{D}}(p)}{N(p)\cdot v(p)}v(p), \tag*{(6)}
XD=π−1(p,DD(p))=DD(p)v(p).(7)X_{\mathcal{D}}=\pi^{-1}(p,D_{\mathcal{D}}(p))=D_{\mathcal{D}}(p)v(p). \tag*{(7)}

Here, v(p)v(p) is the unit camera-ray direction defined in Eq. (3). Let XtrueX_{\mathrm{true}} denote the corresponding ground-truth surface point. The reconstruction errors associated with the unbiased depth and rendered distance are defined as

ϵusv(p)=∥Xu−Xtrue∥2=∥π−1(p,Du(p))−Xtrue∥2,(8)\epsilon_u^{sv}(p)=\lVert X_u-X_{\mathrm{true}}\rVert_2=\lVert\pi^{-1}(p,D_u(p))-X_{\mathrm{true}}\rVert_2, \tag*{(8)}
ϵDsv(p)=∥XD−Xtrue∥2=∥π−1(p,DD(p))−Xtrue∥2.(9)\epsilon_{\mathcal{D}}^{sv}(p)=\lVert X_{\mathcal{D}}-X_{\mathrm{true}}\rVert_2=\lVert\pi^{-1}(p,D_{\mathcal{D}}(p))-X_{\mathrm{true}}\rVert_2. \tag*{(9)}

Application of the triangle inequality to Xu−Xtrue=(Xu−XD)+(XD−Xtrue)X_u-X_{\mathrm{true}}=(X_u-X_{\mathcal{D}})+(X_{\mathcal{D}}-X_{\mathrm{true}}), together with the above error definitions, yields

ϵusv(p)≤ϵDsv(p)+∥π−1(p,Du(p))−π−1(p,DD(p))∥2.(10)\epsilon_u^{sv}(p)\leq\epsilon_{\mathcal{D}}^{sv}(p)+\left\lVert\pi^{-1}(p,D_u(p))-\pi^{-1}(p,D_{\mathcal{D}}(p))\right\rVert_2. \tag*{(10)}

Using Eqs. 6–7 and ∥v(p)∥2=1\lVert v(p)\rVert_{2}=1, the second term on the right-hand side of Eq. 10 can be expressed as

∥π−1(p,Du(p))−π−1(p,DD(p))∥2=∣DD(p)∣∣1N(p)⋅v(p)−1∣=∣ΔD(p)∣.(11)\left\lVert\pi^{-1}(p,D_u(p))-\pi^{-1}(p,D_D(p))\right\rVert_{2} = \left|D_D(p)\right|\left|\frac{1}{\mathbf{N}(p)\cdot v(p)}-1\right| = \left|\Delta D(p)\right|. \tag*{(11)}

Substituting Eq. 11 into Eq. 10 yields

ϵusv(p)≤ϵDsv(p)+∣ΔD(p)∣.(12)\epsilon_u^{sv}(p)\leq\epsilon_D^{sv}(p)+\left|\Delta D(p)\right|. \tag*{(12)}

Since both sides of Eq. 12 are nonnegative, squaring the inequality, multiplying both sides by the nonnegative weight w(p)w(p), and integrating over the valid image domain Ω\Omega yield the aggregate reconstruction-error bound in Eq. 4. This completes the proof. □

Remark 1. Expanding the right-hand side of Eq. 4 gives

Esv≤∫Ωw(p)[(ϵDsv(p))2+2∣ϵDsv(p)∣∣ΔD(p)∣+∣ΔD(p)∣2] dp.(13)E_{sv}\leq\int_{\Omega}w(p)\left[\left(\epsilon_D^{sv}(p)\right)^2+2\left|\epsilon_D^{sv}(p)\right|\left|\Delta D(p)\right|+\left|\Delta D(p)\right|^2\right]\,dp. \tag*{(13)}

Thus, for a fixed ϵDsv(p)\epsilon_D^{sv}(p), reducing ∣ΔD(p)∣\left|\Delta D(p)\right| tightens the derived reconstruction-error bound, motivating the adaptive regularization below.

Depth-Difference-Weighted Adaptive Normal Regularization

Eqs. 11 and 12 show that ∣ΔD(p)∣\left|\Delta D(p)\right| equals the Euclidean distance between XuX_u and XDX_D and appears explicitly in the derived reconstruction-error bound. Thus, for a fixed or comparable ϵDsv(p)\epsilon_D^{sv}(p), a larger depth difference enlarges the right-hand side of the bound and indicates a larger discrepancy between the unbiased depth Du(p)D_u(p) and the rendered distance DD(p)D_D(p).

Guided by this relationship, we use ∣ΔD(p)∣\left|\Delta D(p)\right| to adaptively weight the normal constraint between the reference normal Nd(p)\mathbf{N}_d(p) computed from Du(p)D_u(p) and the rendered normal N(p)\mathbf{N}(p).

As illustrated in Figure 3, rather than directly minimizing ∣ΔD(p)∣\left|\Delta D(p)\right|, we treat it as a local geometric reliability measure and map it to the adaptive weight WΔD(p)\mathcal{W}_{\Delta D}(p). The weight is defined as

Depth-difference weights derived from rendered distance, unbiased depth, and their difference

Figure 3. Depth-difference weights are derived from the discrepancy between the rendered-distance map and the unbiased-depth map. The greater the depth difference, the smaller the weight.

WΔD(p)=1−∣ΔD(p)∣−ΔDmin⁡ΔDmax⁡−ΔDmin⁡+σ,(14)\mathcal{W}_{\Delta D}(p)=1-\frac{\left|\Delta D(p)\right|-\Delta D_{\min}}{\Delta D_{\max}-\Delta D_{\min}+\sigma}, \tag*{(14)}

where σ>0\sigma>0 prevents division by zero, and ΔDmin⁡/ΔDmax⁡\Delta D_{\min}/\Delta D_{\max} are the minimum/maximum of ∣ΔD(p)∣\left|\Delta D(p)\right| over valid pixels. A smaller depth difference produces a larger weight and hence stronger normal regularization, whereas a larger difference weakens the constraint to avoid propagating unreliable local geometry.

The single-view normal constraint is defined as

Lsvgeo=1∣ΩΔD∣∑p∈ΩΔDW(p)∥Nd(p)−N(p)∥1,(15)\mathcal{L}_{svgeo}=\frac{1}{\left|\Omega_{\mathit{\Delta D}}\right|}\sum_{p\in\Omega_{\mathit{\Delta D}}}\mathcal{W}(p)\left\lVert\mathbf{N}_d(p)-\mathbf{N}(p)\right\rVert_1, \tag*{(15)}

where ΩΔD\Omega_{\mathit{\Delta D}} denotes the valid pixel domain of the rendered-distance map, Nd\mathbf{N}_d represents the reference normals generated from unbiased depth, and N(p)\mathbf{N}(p) denotes the rendered normal. The complete weight is W(p)=WΔD(p)(∇I‾(p))2\mathcal{W}(p)=\mathcal{W}_{\Delta D}(p)\left(\overline{\nabla I}(p)\right)^2, where WΔD\mathcal{W}_{\Delta D} is the proposed depth-difference-based reliability weight, while (∇I‾(p))2\left(\overline{\nabla I}(p)\right)^2 is the edge-aware modulation inherited from PGSR [18]. As shown in Figure 4, the normal–ray consistency values obtained by PGSR [18] and the proposed method are compared across selected scenes.

Figure 4. Comparison of N(p)⋅v(p)\mathbf{N}(p)\cdot v(p) obtained by PGSR and the proposed method. The proposed method generally yields normal–ray consistency values closer to 1 in the evaluated scenes.

Multi-View Geometric Constraint Optimization

Multi-View Depth Difference-Reconstruction Error Relationship

Corresponding three-dimensional points observed from different views should satisfy rigid-transformation consistency. However, as shown in Figure 5, neighboring views may exhibit different depth discrepancies at corresponding surface locations. This motivates us to relate cross-view depth-difference residuals to multi-view reconstruction error.

Cross-view inconsistency of depth-difference maps at corresponding surface locations in neighboring views

Figure 5. Cross-view inconsistency of depth-difference maps at corresponding surface locations in neighboring views.

Lemma 2. In multi-view scenes, the aggregate reconstruction error associated with the unbiased-depth-based point clouds satisfies

Emv≜∑(i,j)∈P∫Ωijwij(p)(eijmv(p))2 dp≤2∑(i,j)∈P∫Ωijwij(p)[(ϵD,ijmv(p))2+rΔ,ij2(p)] dp,(16)E_{mv}\triangleq\sum_{(i,j)\in\mathcal{P}}\int_{\Omega_{ij}}w_{ij}(p)\left(e_{ij}^{mv}(p)\right)^2\,dp \leq2\sum_{(i,j)\in\mathcal{P}}\int_{\Omega_{ij}}w_{ij}(p)\left[\left(\epsilon_{D,ij}^{mv}(p)\right)^2+r_{\Delta,ij}^{2}(p)\right]\,dp, \tag*{(16)}

where P\mathcal{P} denotes the selected directed current/nearest-neighbor view pairs, RijR_{ij} is the rotation from view ii to view jj, wij(p)≥0w_{ij}(p)\geq0, and Ωij\Omega_{ij} is the valid correspondence domain.

Moreover, ϵijmv(p)\epsilon_{ij}^{mv}(p) and ϵD,ijmv(p)\epsilon_{D,ij}^{mv}(p) are the unbiased- and rendered-distance-based cross-view discrepancies, respectively, and rΔ,ij(p)=∥ΔDVj(p)vVj(p)−Rij(ΔDVi(p)vVi(p))∥2r_{\Delta,ij}(p)=\left\|\Delta D^{V_j}(p)v^{V_j}(p)-R_{ij}\left(\Delta D^{V_i}(p)v^{V_i}(p)\right)\right\|_2 is the corresponding depth-difference residual.

Proof. Let the rigid transformation from the current view VcV_c to its nearest-neighbor view VnV_n be TVcVn=(R,t)T_{V_c}^{V_n}=(R,t), where R∈SO(3)R \in SO(3) and t∈R3t \in\mathbb{R}^3. Let pc=(uc,vc)p_c=(u_c,v_c) and pn=(un,vn)p_n=(u_n,v_n) denote corresponding pixels in the two views, with unit camera-ray directions vVcv^{V_c} and vVnv^{V_n}, respectively. For clarity, the view-pair and pixel indices are omitted in the following derivation. Consistent with the single-view definition, the three-dimensional points obtained from the unbiased depth and rendered distance are

XuVc=DuVcvVc,XDVc=DDVcvVc,XuVn=DuVnvVn,XDVn=DDVnvVn.(17)X_u^{V_c}=D_u^{V_c}v^{V_c},\quad X_D^{V_c}=D_D^{V_c}v^{V_c}, \qquad X_u^{V_n}=D_u^{V_n}v^{V_n},\quad X_D^{V_n}=D_D^{V_n}v^{V_n}. \tag*{(17)}

Here, DDVcD_D^{V_c} and DDVnD_D^{V_n} represent the rendered distance maps under the current view and the nearest-neighbor view, respectively. To relate the exact vector residual rΔr_{\Delta} to a scalar cross-view depth-difference residual, we adopt a shared-target-ray construction, under which the transformed unbiased-depth and rendered-distance representations are projected onto the target ray vVnv^{V_n}. The corresponding projected depths are

D^uVn=vVnT(RDuVcvVc+t),(18)\hat{D}_u^{V_n}=v^{V_n\mathsf{T}}\left(RD_u^{V_c}v^{V_c}+t\right), \tag*{(18)}
D^DVn=vVnT(RDDVcvVc+t).(19)\hat{D}_D^{V_n}=v^{V_n\mathsf{T}}\left(RD_D^{V_c}v^{V_c}+t\right). \tag*{(19)}

To quantify multi-view consistency, the Euclidean distance between corresponding points after coordinate transformation is calculated. Accordingly, the cross-view reconstruction error associated with the unbiased-depth-based points is defined as

ϵmv=∥XuVn−XuVc→Vn∥2=∥DuVnvVn−(RDuVcvVc+t)∥2.(20)\epsilon^{mv}=\left\|X_u^{V_n}-X_u^{V_c\rightarrow V_n}\right\|_2 =\left\|D_u^{V_n}v^{V_n}-\left(RD_u^{V_c}v^{V_c}+t\right)\right\|_2. \tag*{(20)}

Define the depth differences as ΔDVc=DuVc−DDVc\Delta D^{V_c}=D_u^{V_c}-D_D^{V_c} and ΔDVn=DuVn−DDVn\Delta D^{V_n}=D_u^{V_n}-D_D^{V_n}. Accordingly, the unbiased depths can be expressed as DuVc=DDVc+ΔDVcD_u^{V_c}=D_D^{V_c}+\Delta D^{V_c} and DuVn=DDVn+ΔDVnD_u^{V_n}=D_D^{V_n}+\Delta D^{V_n}.

Substituting these expressions into Eq. (20) yields:

ϵmv=∥DDVnvVn−RDDVcvVc−t+(ΔDVnvVn−RΔDVcvVc)∥2.(21)\epsilon^{mv} = \left\| D_D^{V_n}v^{V_n}-RD_D^{V_c}v^{V_c}-t +\left(\Delta D^{V_n}v^{V_n}-R\Delta D^{V_c}v^{V_c}\right) \right\|_2. \tag*{(21)}

Applying the triangle inequality yields:

ϵmv≤ϵDmv+∥ΔDVnvVn−RΔDVcvVc∥2,(22)\epsilon^{mv}\leq\epsilon_D^{mv}+\left\|\Delta D^{V_n}v^{V_n}-R\Delta D^{V_c}v^{V_c}\right\|_2, \tag*{(22)}

where

ϵDmv≜∥XDVn−XDVc→Vn∥2=∥DDVnvVn−R(DDVcvVc)−t∥2\epsilon_D^{mv} \triangleq \left\|X_D^{V_n}-X_D^{V_c\rightarrow V_n}\right\|_2 = \left\|D_D^{V_n}v^{V_n}-R\left(D_D^{V_c}v^{V_c}\right)-t\right\|_2

denotes the Euclidean distance between the rendered-distance-based point in the neighboring view and the corresponding current-view point transformed into the neighboring-view coordinate system. For brevity, we denote the second term on the right-hand side of Eq. (22) by rΔ≜∥ΔDVnvVn−RΔDVcvVc∥2r_{\Delta}\triangleq\left\|\Delta D^{V_n}v^{V_n}-R\Delta D^{V_c}v^{V_c}\right\|_2.

By squaring Eq. (22), applying (a+b)2≤2a2+2b2(a+b)^2\leq2a^2+2b^2, multiplying by the nonnegative weight wij(p)w_{ij}(p), integrating over Ωij\Omega_{ij}, and summing over all valid view pairs, we directly obtain the aggregate bound in Eq. (16). Squaring rΔr_{\Delta} yields

rΔ2=(ΔDVn)2∥vVn∥22+(ΔDVc)2∥RvVc∥22−2ΔDVnΔDVc(vVn)TRvVc.(23)r_{\Delta}^2=(\Delta D^{V_n})^2\left\|v^{V_n}\right\|_2^2+(\Delta D^{V_c})^2\left\|Rv^{V_c}\right\|_2^2 -2\Delta D^{V_n}\Delta D^{V_c}(v^{V_n})^{\mathsf{T}}Rv^{V_c}. \tag*{(23)}

Since vVnv^{V_n} and vVcv^{V_c} are unit camera-ray directions and RR is an orthogonal rotation matrix, we have ∥vVn∥2=∥RvVc∥2=1\left\|v^{V_n}\right\|_2=\left\|Rv^{V_c}\right\|_2=1. Moreover, let (vVn)TRvVc=cos⁡θ(v^{V_n})^{\mathsf{T}}Rv^{V_c}=\cos\theta, where θ\theta denotes the angle between vVnv^{V_n} and RvVcRv^{V_c}. (23) then becomes

rΔ2=(ΔDVn)2+(ΔDVc)2−2ΔDVnΔDVccos⁡θ.(24)r_{\Delta}^2=(\Delta D^{V_n})^2+(\Delta D^{V_c})^2-2\Delta D^{V_n}\Delta D^{V_c}\cos\theta. \tag*{(24)}

Here, θ\theta is measured in radians. By completing the square, Eq. (24) can be decomposed as

rΔ2=(ΔDVn−ΔDVccos⁡θ)2+(ΔDVc)2sin⁡2θ.(25)r_{\Delta}^2=\left(\Delta D^{V_n}-\Delta D^{V_c}\cos\theta\right)^2+(\Delta D^{V_c})^2\sin^2\theta. \tag*{(25)}

From Eqs. (18) and (19), the translation terms cancel exactly when taking the difference between the two predicted depths. Using ΔDVc=DuVc−DDVc\Delta D^{V_c}=D_u^{V_c}-D_D^{V_c}, we obtain

D^uVn−D^DVn=(vVn)TR(DuVc−DDVc)vVc=ΔDVccos⁡θ.(26)\hat{D}_u^{V_n}-\hat{D}_D^{V_n} =(v^{V_n})^{\mathsf{T}}R\left(D_u^{V_c}-D_D^{V_c}\right)v^{V_c} =\Delta D^{V_c}\cos\theta. \tag*{(26)}

Therefore, under the shared-target-ray construction and using Eq. (26), the first term in Eq. (25) is exactly the squared scalar cross-view depth-difference residual. Specifically, we define

Δδ≜∣(ΔDVn−ΔDVc)−(D^uVn−D^DVn)∣=∣ΔDVn−ΔDVccos⁡θ∣.(27)\begin{aligned} \Delta\delta\triangleq \left|\left(\Delta D^{V_n}-\Delta D^{V_c}\right)-\left(\hat{D}_u^{V_n}-\hat{D}_D^{V_n}\right)\right| \\ &=\left|\Delta D^{V_n}-\Delta D^{V_c}\cos\theta\right|. \tag*{(27)} \end{aligned}

Accordingly, Eq. (25) can be equivalently rewritten as

rΔ2=(Δδ)2+(ΔDVc)2sin⁡2θ.r_{\Delta}^2=(\Delta\delta)^2+(\Delta D^{V_c})^2\sin^2\theta.

This completes the proof.

Remark 2. Equation 16 bounds the aggregate multi-view reconstruction error in terms of the rendered-distance-based error and the depth-difference-induced vector residual rΔr_{\Delta}. Moreover, Eqs. 25 and 27 give the exact relation

rΔ2=(Δδ)2+(ΔDVc)2sin⁡2θ,r_{\Delta}^2=(\Delta\delta)^2+(\Delta D^{V_c})^2\sin^2\theta,

where Δδ\Delta\delta denotes the scalar cross-view depth-difference residual. When the angle-dependent component ∣ΔDVcsin⁡θ∣\left|\Delta D V_c \sin\theta\right| is sufficiently small, the exact relationship in the preceding equation can be approximated by

rΔ≈Δδ.(28)r_{\Delta} \approx\Delta\delta. \tag*{(28)}

This approximation provides a scalar interpretation of rΔr_{\Delta} and is not required for the validity of the aggregate bound in Eq. (16). Thus, reducing cross-view depth-difference inconsistency reduces the corresponding scalar residual component and motivates the practical multi-view regularization introduced below.

Multi-View Depth-Difference Geometric Constraints

Motivated by Eqs. (27) and (28), we construct a practical cross-view depth consistency regularizer for neighboring views. For s∈{u,D}s \in\{u,D\}, psVn(p)p_{s}^{V_n}(p) denotes the image-plane projection in VnV_n of the point DVcVc(p)vVc(p)D_{V_c}^{V_c}(p)v^{V_c}(p) transformed from VcV_c.

To determine valid cross-view reprojections, we define the joint validity mask as

M(p)≜I[puVn(p)∈[0,W−1)×[0,H−1)∧pDVn(p)∈[0,W−1)×[0,H−1)∧D^uVn(p)>τd∧D^DVn(p)>τd].(29)\begin{aligned} M(p) \triangleq\mathbb{I}\big[& p_{u}^{V_n}(p) \in[0,W-1) \times[0,H-1) \\ &\land p_{D}^{V_n}(p) \in[0,W-1) \times[0,H-1) \\ &\land\hat{D}_{u}^{V_n}(p) > \tau_d \land\hat{D}_{D}^{V_n}(p) > \tau_d \big]. \tag*{(29)} \end{aligned}

Here, I[⋅]\mathbb{I}[\cdot] denotes the indicator function; puVn(p)p_{u}^{V_n}(p) and pDVn(p)p_{D}^{V_n}(p) denote the projected coordinates associated with the unbiased depth and rendered distance, respectively; WW and HH denote the width and height of the neighboring-view depth maps; and τd\tau_d is the minimum valid-depth threshold. The valid reprojection domain is

Ω≜{p∈ΩDD∣M(p)=1}.(30)\Omega\triangleq\{p \in\Omega_{D_D} \mid M(p)=1\}. \tag*{(30)}

View pairs with Ω=∅\Omega= \varnothing are excluded from the multi-view loss computation.

To obtain differentiable depth values at noninteger projected coordinates, we employ bilinear interpolation:

Dπ(p)≜∑a=01∑b=01ωabD(⌊v⋆⌋+a,⌊u⋆⌋+b),p∈Ω,(31)D_{\pi}(p) \triangleq\sum_{a=0}^{1}\sum_{b=0}^{1}\omega_{ab}D(\lfloor v^{\star}\rfloor+a,\lfloor u^{\star}\rfloor+b),\quad p\in\Omega, \tag*{(31)}

where (u⋆,v⋆)(u^{\star},v^{\star}) denotes a projected pixel coordinate, Δu=u⋆−⌊u⋆⌋\Delta u=u^{\star}-\lfloor u^{\star}\rfloor, Δv=v⋆−⌊v⋆⌋\Delta v=v^{\star}-\lfloor v^{\star}\rfloor, and ωab=[(1−a)(1−Δv)+aΔv][(1−b)(1−Δu)+bΔu]\omega_{ab}=[(1-a)(1-\Delta v)+a\Delta v][(1-b)(1-\Delta u)+b\Delta u]. Applying Eq. (31) to DVnVn(p)D_{V_n}^{V_n}(p) at puVn(p)p_{u}^{V_n}(p) and to DDVnD_{D}^{V_n} at pDVn(p)p_{D}^{V_n}(p) yields Du,πVn(p)D_{u,\pi}^{V_n}(p) and DD,πVn(p)D_{D,\pi}^{V_n}(p), respectively.

The two cross-view depth discrepancies are defined as

δu(p)≜∣D^uVn(p)−Du,πVn(p)∣,(32)\delta_u(p) \triangleq\left|\hat{D}_{u}^{V_n}(p)-D_{u,\pi}^{V_n}(p)\right|, \tag*{(32)}
δD(p)≜∣D^DVn(p)−DD,πVn(p)∣.(33)\delta_D(p) \triangleq\left|\hat{D}_{D}^{V_n}(p)-D_{D,\pi}^{V_n}(p)\right|. \tag*{(33)}

Eqs. 32 and 33 measure the discrepancies between the current-view depth predictions transformed into the neighboring view and the corresponding depths sampled from the neighboring-view depth maps.

To connect these practical discrepancies with the scalar residual in Eq. (27), we define its sampled counterpart as

Δδπ(p)≜∣(Du,πVn(p)−DD,πVn(p))−(D^uVn(p)−D^DVn(p))∣.(34)\Delta\delta_{\pi}(p) \triangleq\left|\left(D_{u,\pi}^{V_n}(p)-D_{D,\pi}^{V_n}(p)\right)-\left(\hat{D}_{u}^{V_n}(p)-\hat{D}_{D}^{V_n}(p)\right)\right|. \tag*{(34)}

Applying the triangle inequality to Eq. (34) and using Eqs. 32 and 33 yield

Δδπ(p)≤δu(p)+δD(p).(35)\Delta\delta_{\pi}(p) \leq\delta_u(p)+\delta_D(p). \tag*{(35)}

Moreover, for any α∈(0,1)\alpha\in(0,1),

δu(p)+δD(p)≤αδu(p)+(1−α)δD(p)min⁡{α,1−α}.(36)\delta_u(p)+\delta_D(p)\leq\frac{\alpha\delta_u(p)+(1-\alpha)\delta_D(p)}{\min\{\alpha,1-\alpha\}}. \tag*{(36)}

Eqs. 35 and 36 establish that a positive weighted combination of the two depth discrepancies controls Δδπ(p)\Delta\delta_{\pi}(p) up to the constant factor 1/min⁡{α,1−α}1/\min\{\alpha,1-\alpha\}.

Accordingly, we define the cross-view depth consistency regularizer as

Lmv_D=1∣Ω∣∑p∈Ω[αδu(p)+(1−α)δD(p)],(37)\mathcal{L}_{mv\_D}=\frac{1}{|\Omega|}\sum_{p\in\Omega}\left[\alpha\delta_u(p)+(1-\alpha)\delta_D(p)\right], \tag*{(37)}

where α\alpha balances the two depth discrepancies and is set to α=0.8\alpha=0.8. Since α∈(0,1)\alpha\in(0,1), both terms are assigned positive weights. Thus, minimizing Lmv_D\mathcal{L}_{mv\_D} jointly penalizes the two reprojection discrepancies and serves as a practical surrogate for reducing the sampled cross-view depth-difference residual Δδπ(p)\Delta\delta_{\pi}(p).

Method243740556365698397105106110114118122MeanTime
NeRF [10]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49>12h
VolSDF [44]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86>12h
NeuS [45]1.001.370.930.431.100.650.571.481.090.830.521.200.350.490.540.84>12h
2DGS [36]0.480.820.340.431.350.900.811.271.240.650.701.250.410.660.500.798.5min
GOF [37]0.500.830.400.381.350.820.741.261.790.690.711.310.520.640.510.8333min
GSFGS [38]0.380.650.290.360.790.650.641.091.210.590.471.190.320.420.370.639.9min
GS-Pull [40]0.600.660.360.380.770.740.631.220.960.670.590.580.420.580.530.6422min
GS-UDF [41]0.620.670.430.420.830.860.721.201.030.680.610.630.430.560.520.6825min
PGSR [18]0.350.550.400.350.780.580.491.070.640.590.500.540.330.370.340.5317min
GSDF [39]0.580.930.460.371.300.770.731.581.280.760.581.220.370.510.510.8032min
QGS [42]0.420.640.360.380.860.640.501.130.970.610.480.690.340.410.360.5948min
GSPrior [43]0.320.550.300.310.740.540.481.060.690.570.450.580.290.350.330.5042min
GSPrior† [43]0.330.540.370.340.790.580.491.070.640.590.470.560.310.380.350.5249.9min
Ours0.350.550.370.340.770.560.491.040.640.580.470.500.300.360.330.5117.6min

Table 1. TABLE I Quantitative reconstruction accuracy on DTU measured by Chamfer distance (mm)↓. In the CD columns, red, orange, and yellow indicate the best, second-best, and third-best results, respectively. The superscript † denotes results reproduced by us.

Optimization

The proposed regularizers are jointly optimized with the image reconstruction loss Lrgb\mathcal{L}_{rgb}, flattened-Gaussian loss Ls\mathcal{L}_{s}, and multi-view photometric consistency loss Lmvgeom\mathcal{L}_{mvgeom}:

L=λ1Lrgb+λ2Ls+λ3Lmvgeom+λ4Lsvgeo+λ5Lmv_D.(38)\mathcal{L}=\lambda_{1}\mathcal{L}_{rgb}+\lambda_{2}\mathcal{L}_{s}+\lambda_{3}\mathcal{L}_{mvgeom}+\lambda_{4}\mathcal{L}_{svgeo}+\lambda_{5}\mathcal{L}_{mv\_D}. \tag*{(38)}

Here, λ1\lambda_{1}–λ5\lambda_{5} denote the corresponding loss weights; λ1\lambda_{1}–λ4\lambda_{4} retain the PGSR settings, while λ5\lambda_{5} controls the proposed cross-view regularizer.

EXPERIMENTS

Datasets and Implementation Details

We conduct experiments on three public benchmarks: Mip-NeRF360 [20], Tanks and Temples [22], and DTU [21]. Novel view synthesis is evaluated on Mip-NeRF360, while surface reconstruction is evaluated on DTU and Tanks and Temples. Rendering quality is measured using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and learned perceptual image patch similarity (LPIPS). Surface reconstruction is evaluated using Chamfer distance (CD) on DTU and the benchmark F1-score on Tanks and Temples.

All experiments are conducted on a single NVIDIA GeForce RTX 4090 GPU. Unless otherwise specified, the training strategy and hyperparameters follow those of PGSR [18], and each scene is optimized for 30,000 iterations. Following PGSR, the minimum valid-depth threshold is set to τd=0.1\tau_d = 0.1, and the loss weights λ1\lambda_1, λ2\lambda_2, λ3\lambda_3, and λ4\lambda_4 adopt the baseline settings. For the proposed components, the numerical stabilization constant in Eq. (14) is set to σ=10−6\sigma= 10^{-6}, and the cross-view depth consistency regularization weight is set to λ5=0.008\lambda_5 = 0.008. After optimization, depth maps rendered from the optimized Gaussian representation are fused into a truncated signed distance function (TSDF) volume, from which a surface mesh is extracted for reconstruction evaluation.

Surface Reconstruction Comparisons

Comparisons on DTU. We evaluate the surface reconstruction performance of the proposed method on the DTU dataset [21] and compare it with representative neural implicit methods and recent Gaussian-splatting-based reconstruction approaches, as listed in Table I. The quantitative results show that our method obtains a mean Chamfer distance of 0.51 mm, ranking second in terms of average reconstruction accuracy among the compared methods. Compared with PGSR, our method reduces the mean Chamfer distance from 0.53 mm to 0.51 mm under a similar experimental setting. Overall, these results indicate that our method achieves competitive reconstruction accuracy under the reported experimental setting.

Figure 6 presents qualitative comparisons of reconstructed meshes with enlarged local regions. The visual results suggest that our method can maintain relatively continuous surfaces and preserve local geometric details in challenging regions, which is consistent with the quantitative comparison.

Qualitative surface reconstruction comparison on the DTU dataset

Figure 6. Qualitative surface reconstruction comparison on the DTU dataset. Reconstructed meshes and zoom-in regions are shown to compare local surface details and geometric completeness across different methods.

Comparisons on Tanks and Temples. We further evaluate our method on the large-scale Tanks and Temples dataset to assess its reconstruction generalization beyond the controlled DTU setting. As shown in Table II, our method achieves a competitive mean F1-score and shows comparable performance to recent Gaussian-splatting-based reconstruction methods. Figure 7 presents qualitative comparisons of reconstructed meshes and enlarged local regions. The results suggest that our method can produce relatively continuous surfaces and preserve local geometric details in challenging large-scale scenes.

Qualitative surface reconstruction comparison on the Tanks and Temples dataset

Figure 7. Qualitative surface reconstruction comparison on the Tanks and Temples dataset. Reconstructed meshes and zoom-in regions are shown to compare local surface continuity and geometric details across different methods.

ClassMethodBarnCaterpillarCourthouseIgnatiusMeetingroomTruckMean
ImplicitNeuS [45]0.290.290.170.830.240.450.378
ImplicitNeuralangelo [46]0.700.360.280.890.320.480.505
ImplicitMonoSDF [47]0.490.310.120.780.230.420.392
Explicit3DGS [13]0.130.080.090.040.010.190.090
Explicit2DGS [36]0.360.230.130.440.160.260.263
ExplicitGOF [37]0.510.410.280.680.280.590.458
ExplicitPGSR [18]0.660.410.210.800.290.600.495
ExplicitGS-Pull [40]0.600.370.160.710.220.520.430
ExplicitQGS [42]0.550.400.280.810.310.640.500
ExplicitGSPrior [43]0.660.430.200.800.310.640.507
ExplicitOurs0.640.430.210.810.300.630.503

Table II. TABLE II. QUANTITATIVE F1-SCORE COMPARISON ON TANKS AND TEMPLES. BOLD INDICATES THE BEST RESULT WITHIN EACH CLASS.

Novel View Synthesis Comparisons

Table III and Figure 8 present the quantitative and qualitative novel view synthesis results on the Mip-NeRF360 dataset. We evaluate rendering quality using PSNR, SSIM, and LPIPS, and compare our method with representative and recent Gaussian-splatting-based approaches.

Qualitative comparison of rendered images from different methods on the Mip-NeRF360 dataset

Figure 8. Qualitative comparison on the Mip-NeRF360 dataset. Rendered images from different methods are shown with enlarged local regions marked by red boxes. The results suggest that our method preserves fine image details and local structures in the selected scenes.

MethodOutdoor scenesIndoor scenesOverall average
SSIM↑\uparrowPSNR↑\uparrowLPIPS↓\downarrowSSIM↑\uparrowPSNR↑\uparrowLPIPS↓\downarrowSSIM↑\uparrowPSNR↑\uparrowLPIPS↓\downarrow
3DGS [13]0.74225.070.2320.93131.180.1640.83728.130.198
2DGS [36]0.70624.210.2840.91230.150.2120.80927.180.248
GOF [37]0.74624.810.2080.91730.400.1890.83227.610.199
GFSGS [38]0.73424.400.2240.91629.930.1940.82527.170.209
GS-UDF [41]0.69823.840.2920.90629.160.2190.80226.500.255
PGSR [18]0.75224.740.2030.92930.160.1590.84127.450.181
QGS [42]0.70623.320.2420.91930.450.1840.81326.890.213
GSPrior [43]0.75424.810.2000.93330.560.1550.84427.690.178
GSPrior†[43]0.75524.770.2010.93130.390.1590.84327.580.180
Ours0.74924.750.1990.93330.680.1510.84127.720.175

Table III. Novel-view-synthesis results on Mip-NeRF360. Red, orange, and yellow indicate the best, second-best, and third-best results, respectively; † denotes reproduced results.

As shown in Table III, our method achieves the lowest LPIPS across outdoor scenes, indoor scenes, and the overall average, indicating favorable perceptual rendering quality. It also obtains the second-best overall PSNR and competitive SSIM compared with recent methods. These results suggest

that the proposed geometric regularization maintains competitive novel view synthesis quality while improving perceptual rendering metrics. Figure 8 further provides qualitative comparisons with enlarged local regions, where our method preserves fine image details and local structures in the selected scenes.

Optimization Dynamics

We record the training- and test-view PSNR at 7k, 10k, 15k, 20k, 25k, and 30k iterations and compare DepthGS with PGSR in Figure 9. The first column shows the PSNR evolution, the second reports Δ\DeltaPSNR = PSNROurs_{\mathrm{Ours}} − PSNRPGSR_{\mathrm{PGSR}}, and the third compares the average training- and test-view PSNR. DepthGS generally achieves higher PSNR at most recorded iterations, with faster improvement during 7k–15k iterations in several scenes. The maximum training/test gains reach 0.38/0.37 dB in counter and 0.36/0.33 dB in bonsai.

Figure 9. Optimization dynamics compared with PGSR. (a) Training- and test-view PSNR curves; (b) per-iteration Δ\DeltaPSNR; (c) average training- and test-view PSNR.

The average training/test gains are 0.27/0.28 dB in counter, 0.26/0.23 dB in bonsai, and 0.18/0.08 dB in garden. In the more challenging bicycle scene, the training PSNR increases by 0.04 dB, whereas the test PSNR decreases by 0.05 dB.

WΔD\mathcal{W}_{\Delta D}Lmv_D\mathcal{L}_{mv\_D}PSNR↑\uparrowSSIM↑\uparrowLPIPS↓\downarrow
××27.450.8410.181
✓×27.750.8420.175
×✓27.510.8420.180
✓✓27.730.8410.175

Table 4. ABLATION STUDY ON THE MIP-NERF360 DATASET. BEST RESULTS ARE IN BOLD.

This degradation may result from complex illumination, wide viewpoint variations, and fragmented geometry, which reduce the reliability of normal and depth estimates. Overall, the results indicate favorable but scene-dependent optimization behavior rather than universally improved convergence.

Ablation Studies

Ablation settings and results. We evaluate the depth-difference-guided weight WΔD\mathcal{W}_{\Delta D} and the cross-view depth consistency regularizer Lmv_D\mathcal{L}_{mv\_D}. The image-gradient factor ∇I‾\nabla\overline{I} is kept unchanged whenever the single-view normal constraint is used, so the ablation isolates the contribution of WΔD\mathcal{W}_{\Delta D}. In Tables IV and V, “×” and “✓” indicate the exclusion and inclusion of each component, respectively.

Table IV shows that WΔD\mathcal{W}_{\Delta D} primarily improves image-level rendering metrics, while the full model maintains comparable rendering quality. Table V shows that the proposed components reduce CD from 0.53 to 0.51 mm, with average and peak memory usages of 6.77 and 7.50 GB for the full model. The complementary effect of Lmv_D\mathcal{L}_{mv\_D} is further evaluated using EMVC below.

Qualitative ablation analysis. Figure 10 compares the rendered RGB images, unbiased depth maps, and normal maps produced by PGSR and the three ablation variants. Introducing

Novel-view-synthesis ablation results showing rendered RGB, unbiased depth, and normal maps for PGSR and ablation variants

Figure 10. Novel-view-synthesis ablation results. The rows show RGB, unbiased depth, and normal maps; the columns show PGSR, the two individual components, and the full model.

either Lmv_D\mathcal{L}_{mv\_D} or WΔD\mathcal{W}_{\Delta D} improves local boundary preservation and geometric coherence in the highlighted regions. Their combination achieves the highest PSNR of 31.19 dB for the illustrated view and provides a favorable overall result across the three representations.

MethodWΔD\mathcal{W}_{\Delta D}Lmv_D\mathcal{L}_{mv\_D}CD↓\downarrowAvg. Mem.↓\downarrowPeak Mem.↓\downarrow
PGSR××0.536.417.12
+ Lmv_D\mathcal{L}_{mv\_D}×✓0.526.897.98
+ WΔD\mathcal{W}_{\Delta D}✓×0.516.747.43
Ours✓✓0.516.777.50

Table 5. ABLATION STUDY ON THE DTU DATASET (CD: MM; MEMORY: GB).

Figure 11 shows that the proposed components generally produce clearer local contours and better-defined geometric structures. These results are consistent with their intended roles: WΔD\mathcal{W}_{\Delta D} adapts the local normal constraint according to geometric reliability, while Lmv_D\mathcal{L}_{mv\_D} reduces cross-view depth discrepancies.

3D reconstruction ablation study comparing PGSR and ablation variants

Figure 11. 3D reconstruction ablation study. The red dashed box indicates the magnified region. The comparison of local structures within the highlighted region illustrates the effects of different components on local contour preservation and geometric details.

Analysis of cross-view depth consistency regularization. Using the sampled cross-view depth-difference residual defined in Eq. (34), we define the multi-view consistency error for each valid view pair as

EMVC=1∣Ω∣∑p∈Ω∣(Du,πVn(p)−DD,πVn(p))−(D^uVn(p)−D^DVn(p))∣(39)E_{\mathrm{MVC}}=\frac{1}{\lvert\Omega\rvert}\sum_{p\in\Omega}\left\lvert\left(D_{u,\pi}^{\mathcal{V}_n}(p)-D_{D,\pi}^{\mathcal{V}_n}(p)\right)-\left(\widehat{D}_{u}^{\mathcal{V}_n}(p)-\widehat{D}_{D}^{\mathcal{V}_n}(p)\right)\right\rvert \tag*{(39)}

A lower EMVCE_{\mathrm{MVC}} indicates better cross-view depth-difference consistency. View pairs with Ω=∅\Omega=\varnothing are excluded from the evaluation.

As shown in Figure 12, the averaged EMVCE_{\mathrm{MVC}} curves on DTU and Mip-NeRF360 are generally lower with Lmv_D\mathcal{L}_{mv\_D} than without it, indicating that the proposed cross-view depth consistency regularization helps reduce cross-view depth-difference inconsistency. Since Lmv_D\mathcal{L}_{mv\_D} mainly targets cross-view geometric consistency, EMVCE_{\mathrm{MVC}} provides a direct diagnostic measure for evaluating its intended effect.

Figure 12. Mean EMVCE_{\mathrm{MVC}} on DTU and Mip-NeRF360. Shaded regions denote standard errors across scenes.

CONCLUSION

In this paper, we have proposed DepthGS, a depth-guided Gaussian splatting framework with adaptive regularization. More specifically, we have derived theoretical relations between depth information and reconstruction error in both single-view and multi-view settings and established corresponding upper bounds on the reconstruction error. We have further proposed depth-driven optimization frameworks for single-view and multi-view settings that use depth-derived geometric cues to regularize the optimization and improve geometric consistency. Experimental results on multiple public datasets, including DTU, Mip-NeRF360, and Tanks and Temples, demonstrate competitive reconstruction and rendering performance in both quantitative metrics and qualitative comparisons. Future research will focus on addressing challenges in complex outdoor environments, especially under strong illumination variations, noisy geometry, and unstructured scene layouts, to further improve the robustness of the method in real-world applications.

References

  1. [1]M. Li, W. Chen, N. Cheng, J. Xu, D. Li, and H. Wang, “Garad-slam: 3d gaussian splatting for real-time anti dynamic slam,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 11 047–11 053.arxiv.org/abs/2502.03228
  2. [2]M. Li, Y. Zhou, H. Zhou, X. Hu, F. Roemer, H. Wang, and A. Osman, “Dy3dgs-slam: Monocular 3d gaussian splatting slam for dynamic environments,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 14 572–14 578.arxiv.org/abs/2506.05965
  3. [3]J. Wakulicz, H. Kong, and S. Sukkari eh, “Active information acquisition under arbitrary unknown disturbances,” in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 8429–8435.arxiv.org/abs/2109.09079
  4. [4]S. Eiffert, N. D. Wallace, H. Kong, N. Pirmarzda sht i, and S. Sukkari eh, “Resource and response aware path planning for long-term autonomy of ground robots in agriculture,” Field Robotics, vol. 2, pp. 1–33, 2022.arxiv.org/abs/2105.10690
  5. [5]J. Wang, Y. He, D. Su, K. Nakadai, J. W u, S. Huang, Y. Li, and H. Kong, “Slam-based joint calibration of multiple asynchronous microphone arrays and sound source localization,” IEEE Transactions on Robotics, 2024.arxiv.org/abs/2405.19813
  6. [6]M. Khan, H. Fazlali, D. Sharma, T. Cao, D. Bai, Y. Ren, and B. Liu, “Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction,” in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 8315–8321.arxiv.org/abs/2407.02598
  7. [7]Z. Liang, H. Li, K. Jia, K. Guo, and Q. Zhang, “Gus-ir: Gaussian splatting with unified shading for inverse rendering,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 10, pp. 8364–8378, 2025.DOI
  8. [8]P. Wang, L. Zhao, Y. Zhang, S. Zhao, and P. Liu, “Mba-slam: Motion blur aware dense visual slam with radiance fields representation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 12, pp. 11 168–11 186, 2025.DOI
  9. [9]Z.-X. Yin, P.-Y. Jiao, J. Qiu, M.-M. Cheng, and B. Ren, “Ms-nerf: Multi-space neural radiance fields,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 5, pp. 3766–3783, 2025.DOI
  10. [10]B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng, “NeRF: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021.arxiv.org/abs/2003.08934
  11. [11]W. Hu, C. Tian, L. Wen, and H. Ding, “Td-nerf: Transfer learning and diffusion regulation-based nerf for scene perception,” IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–12, 2025.DOI
  12. [12]Y. Han, T. Yu, X. Yu, D. Xu, B. Zheng, Z. Dai, C. Yang, Y. Wang, and Q. Dai, “Super-nerf: View-consistent detail generation for nerf super-resolution,” IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 9, pp. 6053–6066, 2025.DOI
  13. [13]B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3D gaussian splatting for real-time radiance field rendering,” ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023.
  14. [14]Y. Cheng, X. Li, X. Ren, J. Wang, H. Jia, D. Xu, W. Zhu, B. Ni, and Y. Yan, “Topo4d++: Realistic physically based 4d head capture with topology-preserving gaussian splatting and expression priors,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–17, 2025.DOI
  15. [15]Y. Li, C. Lyu, Y. Di, G. Zhai, G. H. Lee, and F. Tombari, “Geogaussian: Geometry-aware gaussian splatting for scene rendering,” in European Conference on Computer Vision, 2024, pp. 441–457.arxiv.org/abs/2403.11324
  16. [16]K. Cheng, X. Long, K. Yang, Y. Yao, W. Yin, Y. Ma, W. Wang, and X. Chen, “Gaussianpro: 3D gaussian splatting with progressive propagation,” in Forty-first International Conference on Machine Learning, 2024.arxiv.org/abs/2402.14650
  17. [17]M. Turkulainen, X. Ren, I. Melekhov, O. Seiskari, E. Rahtu, and J. Kannala, “DNG-Splat: Depth and normal priors for gaussian splatting and meshing,” arXiv preprint arXiv:2403.17822, 2024.
  18. [18]D. Chen, H. Li, W. Ye, Y. Wang, W. Xie, S. Zhai, N. Wang, H. Liu, H. Bao, and G. Zhang, “PGSR: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction,” IEEE Transactions on Visualization and Computer Graphics, 2024.DOI
  19. [19]H. Chen, F. Wei, C. Li, T. Huang, Y. Wang, and G. H. Lee, “VCR-GauS: View consistent depth-normal regularizer for gaussian surface reconstruction,” Advances in Neural Information Processing Systems, vol. 37, pp. 139 725–139 750, 2025.arxiv.org/abs/2406.05774
  20. [20]J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5470–5479.arxiv.org/abs/2111.12077
  21. [21]R. Jensen, A. Dahl, G. Vogiatzis, E. Tola, and H. Aanaes, “Large scale multi-view stereopsis evaluation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2014, pp. 406–413.DOI
  22. [22]A. Knapitsch, J. Park, Q.-Y. Zhou, and V. Koltun, “Tanks and temples: Benchmarking large-scale scene reconstruction,” ACM Transactions on Graphics (ToG), vol. 36, no. 4, pp. 1–13, 2017.
  23. [23]J. L. Schonberger and J.-M. Frahm, “Structure-from-motion revisited,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016, pp. 4104–4113.DOI
  24. [24]E. Penner and L. Zhang, “Soft 3d reconstruction for view synthesis,” ACM Transactions on Graphics (ToG), vol. 36, no. 6, pp. 1–11, 2017.DOI
  25. [25]N. Deng, Z. He, J. Ye, B. Duinkharin, P. Chakravuth, X. Yang, and Q. Sun, “Fov-nerf: Foveated neural radiance fields for virtual reality,” IEEE Transactions on Visualization and Computer Graphics, vol. 28, no. 11, pp. 3854–3864, 2022.DOI
  26. [26]X. Huang, Q. Zhang, Y. Feng, H. Li, and Q. Wang, “Ltm-nerf: Embedding 3d local tone mapping in hdr neural radiance field,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10944–10959, 2024.DOI
  27. [27]D. Zhu, Z. Wang, and X. Jiang, “An efficient geometric points-guided generalized exponential splatting,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 20 546–20 558, 2025.DOI
  28. [28]Z. Liu, J. Su, G. Cai, Y. Chen, B. Zeng, and Z. Wang, “Georgs: Geometric regularization for real-time novel view synthesis from sparse inputs,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 12, pp. 13 113–13 126, 2024.DOI
  29. [29]J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park, “Compact 3d gaussian representation for radiance field,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 719–21 728.arxiv.org/abs/2311.13681
  30. [31]S. S. Mallick, R. Goel, B. Kerbl, M. Steinberger, F. V. Carrasco, and F. De La Torre, “Taming 3dgs: High-quality radiance fields with limited resources,” in SIGGRAPH Asia 2024 Conference Papers, 2024, pp. 1–11.arxiv.org/abs/2406.15643
  31. [32]T. Shen, S. Liu, J. Feng, Z. Ma, and N. An, “Topology-aware 3d gaussian splatting: Leveraging persistent homology for optimized structural integrity,” in Proceedings of the 39th AAAI Conference on Artificial Intelligence, 2024.arxiv.org/abs/2412.16619
  32. [33]G. Feng, S. Chen, R. Fu, Z. Liao, Y. Wang, T. Liu, B. Hu, L. Xu, Z. Pei, H. Li et al., “Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 652–26 662.arxiv.org/abs/2408.07967
  33. [34]Z. Guo, W. Zhou, L. Li, M. Wang, and H. Li, “Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 4, pp. 3119–3133, 2025.arxiv.org/abs/2403.11447
  34. [35]R. Fan, J. Wu, X. Shi, L. Zhao, Q. Ma, and L. Wang, “Fov-gs: Foveated 3d gaussian splatting for dynamic scenes,” IEEE Transactions on Visualization and Computer Graphics, vol. 31, no. 5, pp. 2975–2985, 2025.DOI
  35. [36]B. Huang, Z. Yu, A. Chen, A. Geiger, and S. Gao, “2d gaussian splatting for geometrically accurate radiance fields,” in ACM SIGGRAPH 2024 conference papers, 2024, pp. 1–11.DOI
  36. [37]Z. Yu, T. Sattler, and A. Geiger, “Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes,” ACM Transactions on Graphics (ToG), vol. 43, no. 6, pp. 1–13, 2024.arxiv.org/abs/2404.10772
  37. [38]K. Jiang, V. Sivaram, C. Peng, and R. Ramamoorthi, “Geometry field splatting with gaussian surfels,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 5752–5762.arxiv.org/abs/2411.17067
  38. [39]M. Yu, T. Lu, L. Xu, L. Jiang, Y. Xiangli, and B. Dai, “Gsdf: 3dgs meets sdf for improved neural rendering and reconstruction,” Advances in Neural Information Processing Systems, vol. 37, pp. 129 507–129 530, 2024.DOI
  39. [40]W. Zhang, Y.-S. Liu, and Z. Han, “Neural signed distance function inference through splatting 3d gaussians pulled on zero-level set,” in Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2410.14189
  40. [41]S. Li, Y.-S. Liu, and Z. Han, “Gaussianudf: Inferring unsigned distance functions through 3d gaussian splatting,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 113–27 123.arxiv.org/abs/2503.19458
  41. [42]Z. Zhang, B. Huang, H. Jiang, L. Zhou, X. Xiang, and S. Shen, “Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 28 260–28 270.arxiv.org/abs/2411.16392
  42. [43]T. Noda, Y.-S. Liu, and Z. Han, “3d gaussian splatting with self-constrained priors for high fidelity surface reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 26 041–26 051.arxiv.org/abs/2603.19682
  43. [44]L. Yariv, J. Gu, Y. Kasten, and Y. Lipman, “Volume rendering of neural implicit surfaces,” Advances in Neural Information Processing Systems, vol. 34, pp. 4805–4815, 2021.arxiv.org/abs/2106.12052
  44. [45]P. Wang, L. Liu, Y. Liu, C. Theobalt, T. Komura, and W. Wang, “Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction,” arXiv preprint arXiv:2106.10689, 2021.arxiv.org/abs/2106.10689
  45. [46]Z. Li, T. Müller, A. Evans, R. T. Taylor, M. Ummenhofer, M.-Y. Liu, and C.-H. Lin, “Neuralangelo: High-fidelity neural surface reconstruction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 8456–8465.arxiv.org/abs/2306.03092
  46. [47]Z. Yu, S. Peng, M. Niemeyer, T. Sattler, and A. Geiger, “Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction,” Advances in Neural Information Processing Systems, vol. 35, pp. 25 018–25 032, 2022.arxiv.org/abs/2206.00665

Paper details

vlnc:W_3qye6ah3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents