betaFeedback

WorksComputer Science

RAPID: Reconstruction Accelerated by Polarization-Informed Depth for 2D Gaussian Splatting

Abstract

Efficient and accurate 3D surface reconstruction from multi-view images remains a fundamental challenge in Gaussian splatting. Many existing methods rely on sparse geometric initialization and iterative densification, resulting in substantial optimization overhead and difficulties in balancing rendering quality and geometric accuracy, particularly in reflective and weakly textured regions. To address these limitations, we propose RAPID, a geometry-first framework that uses multiview polarization cues to recover dense surface priors. We integrate surface normals into depth maps using sparse geometric anchors. These depth maps are then fused across views. Curvature-adaptive sampling and oriented Gaussian initialization then support efficient optimization. Experiments on the PISR dataset demonstrate an approximately 5×5\times speedup over PhyGaP, which is a state-of-the-art Gaussian-based reconstruction method. RAPID reduces the average optimization time from approximately 21 minutes (PhyGaP) to 4 minutes while maintaining competitive novel-view synthesis quality and mesh reconstruction accuracy. These results demonstrate the effectiveness of RAPID in accelerating Gaussian optimization from polarimetric multi-view inputs.

Figure 1. We propose Pol2DGS, the two panels on the left illustrate how our method balances geometric accuracy and training time. The center shows the information flow of the proposed method. The right compares our novel view synthesis result with that of PhyGaP [39], a state-of-the-art reconstruction method based on Gaussian splatting.

Introduction

Recovering accurate 3D surfaces from multiview images [12, 34, 40, 41] and reproducing object appearance from unseen viewpoints are fundamental problems in 3D vision. 2D Gaussian splatting (2DGS) [19] represents local surfaces using oriented disks distributed in 3D space and efficiently synthesizes novel views through rasterization. However, rendering speed after model construction and the time required to build the model are distinct measures of efficiency [10, 25]. Starting from a sparse point cloud still requires repeated adjustment of disk positions, orientations, and scales, along with densification to establish surface coverage [19, 23]. In reflective and weakly textured regions, color variations across views provide limited reliable geometric guidance for these adjustments [5, 7]. Reducing scene optimization time while preserving novel view synthesis quality and mesh reconstruction accuracy therefore remains a challenge in Gaussian surface reconstruction.

Polarization observations provide information about surface orientation, supporting geometric estimation beyond color observations [7, 9]. In neural implicit reconstruction, PANDORA [9] uses polarimetric inverse rendering to jointly recover object shape and separate diffuse and specular reflection components. NeRSP [16] combines Stokes vector rendering errors with multiview azimuth consistency to constrain the shape of reflective surfaces from sparse views. PISR [5] incorporates per-pixel perspective viewing rays into polarization constraints to directly constrain implicit surface normals. In explicit Gaussian reconstruction, Sparse2DGS [38] generates a dense initial point cloud from depth predicted by multiview stereo matching, then selectively updates primitives based on rendered geometry. PhyGaP [39] uses polarimetric deferred rendering to estimate geometry and material properties, supporting reflection decomposition and relighting. PolarGS [15] employs polarization information for depth completion and back-projects the completed depth to generate new Gaussians.

Inspired by these methods, we focus on geometric initialization. We investigate how polarization observations can determine the initial positions, orientations, and spatial extents of 2D Gaussians before training to reduce subsequent optimization time. Our research faces two core challenges:

The first challenge is recovering surface positions. Since polarization-derived normals specify tangent-plane orientations but not positions along viewing rays. Under perspective projection, normals constrain log-depth gradients, so integration recovers only relative depth [31]; we use sparse geometric anchors to fix its scale. Cross-view scale inconsistencies can place back-projected points on different depth layers despite accurate normals [3, 31], introducing position errors into Gaussian initialization and subsequent training. Depth anchoring and multiview surface alignment are therefore both necessary. A further challenge is the mismatch between depth sampling density and required primitive density [10, 26]. Assigning one Gaussian to each depth sample creates redundant coverage in low-curvature regions, while uniformly coarser sampling risks losing detail in curved regions. A single scale is also insufficient when curvature differs between the two principal directions. Initializing representative primitives from dense observations using curvature adaptive spacing and scales emerges as a viable option. During training, merging redundant primitives and subdividing inadequately represented regions constitutes a principled strategy [1, 18, 32], so that the primitive count reflects local surface requirements.

In light of these considerations, we propose Pol2DGS, a method that uses polarization-derived surface priors to accelerate 2D Gaussian optimization. Given calibrated multiview color images, polarization observations, and foreground masks, we first estimate polarization-derived normals and their confidence, and then we convert these normals into perspective log-depth gradients and recover depth for each view through least-squares integration with sparse geometric anchors. After multiview consistency checks, we fuse the depth maps using a truncated signed distance function (TSDF) [8] and extract an initial surface. We then fit a quadratic surface in local tangent-plane coordinates. The two principal curvatures determine the sampling spacing along their respective principal directions. Sampled positions define Gaussian centers, while surface normals and principal directions determine disk orientations. Local sampling spacing initializes the two tangential scales. Training thus begins with a set of oriented disks that already provides surface coverage. The primitive distribution is then refined according to observations and local geometry.

Our main contributions are as follows:

  • We integrate polarimetric normals with sparse geometric anchors, then use anisotropic, curvature-adaptive surface sampling to initialize Gaussian positions, orientations, and scales without depth sensors.

  • We replace the native 2DGS [19] primitive update strategy with insertion guided by surface geometry, deterministic splitting along major axes, merging subject to coverage constraints, and pruning based on opacity.

  • Our Pol2DGS reduces average measured runtime from roughly 21 to 4 minutes (5× speedup) while maintaining competitive novel-view synthesis quality and mesh reconstruction accuracy.

Related Work

Novel View Synthesis and Mesh Extraction

Neural radiance fields (NeRF) [27] synthesize novel views by integrating volume density and view-dependent radiance along camera rays. However, a density field can reproduce images without identifying a unique surface [36]. NeuS [36] addresses this ambiguity by representing the surface as the zero level set of a learned signed distance function. 3D Gaussian splatting (3DGS) [23] instead rasterizes anisotropic Gaussians for real-time rendering, but the primitives do not directly define a mesh.

Gaussian reconstruction methods differ in how they recover a surface from these primitives. SuGaR [13] aligns Gaussians with the surface and applies Poisson reconstruction [22] to sampled points. 2DGS [19] represents local surfaces as oriented disks and extracts a mesh by fusing rendered depth maps. Gaussian Opacity Fields (GOF) [43] extracts an opacity isosurface on an adaptive tetrahedral grid. MILo [14] incorporates differentiable mesh extraction into training, allowing discrepancies between mesh and Gaussian renderings to guide optimization. Sparse2DGS [38] exploits geometry at an earlier stage by initializing primitives from stereo depth maps and selectively updating them using rendered geometry. Inspired by this approach, we instead integrate polarization-derived normals and use curvature-adaptive sampling to initialize oriented primitives.

Polarization-Guided 3D Reconstruction

Polarization provides surface orientation cues in weakly textured regions, but orientation alone does not determine depth. The method in Ref. [7] employs polarization-derived azimuths to trace iso-depth contours and propagate reliable stereo depths. PANDORA [9] incorporates polarization into neural inverse rendering to recover shape and illumination while separating diffuse and specular radiance. NeRSP [16] combines supervision on rendered Stokes vectors with multi-view azimuth consistency for sparse-view reconstruction. PISR [5] directly constrains implicit surface normals using the polarization angle and the viewing ray at each pixel.

Gaussian methods exploit polarization at different stages. PolGS [17] uses measured Stokes vectors to supervise diffuse and specular reflection, while PhyGaP [39] uses polarimetric deferred rendering and self-occlusion-aware environment maps for reflection decomposition and relighting. PolarGS [15] incorporates polarization into PatchMatch [2] depth completion and back-projects the completed depths to add Gaussians. Our Pol2DGS uses polarization to establish surface coverage before Gaussian optimization. Sparse geometric anchors fix the depth scale during normal integration. On the fused surface, local normals and principal directions determine disk orientation, while principal curvatures guide sampling intervals and disk scales.

Method

Figure 2 presents the overview of Pol2DGS. The input, as illustrated, comprises multiview red-green-blue (RGB) images, angle of linear polarization (AoLP) and degree of linear polarization (DoLP) measurements, foreground and polarization validity masks, fixed camera calibration, and feature tracks.

Figure 2. Overview. Our method recovers surface normals from multi-view polarization observations, integrates them into dense depth maps using sparse depth anchors, and initializes oriented 2D Gaussians through TSDF surface fusion and curvature-adaptive sampling. Surface-constrained optimization and unified Insert, Split, Merge, and Prune updates then refine the representation for novel-view synthesis and mesh reconstruction.

Polarization-Guided Surface Recovery

As shown in Figure 3, polarization observations provide local directional cues about surfaces but exhibit two types of directional ambiguity. First, linear polarization defines an unoriented axis, making the observed angle of linear polarization (AoLP) ϕ\phi indistinguishable from ϕ+π\phi+ \pi. This gives rise to a π\pi ambiguity. Second, under standard reflection models, diffuse and specular reflection produce orthogonal polarization directions. An additional π/2\pi/2 ambiguity therefore arises when the reflection mechanism is unknown [28, 44]. Moreover, recovered normals cannot directly determine surface positions [31].

Four possible azimuth angles

Figure 3. Four possible azimuth angles. The figure illustrates the π/2\pi/2 and π\pi ambiguities.

For a valid observation, let n\mathbf{n} denote the unit surface normal in camera coordinates and r^\hat{\mathbf{r}} the unit ray from the camera toward the surface. To relate the 3D normal to the polarization observation, we predict the polarization azimuth from the plane of incidence defined by these two vectors [5]:

m=n×r^,α=atan2⁡(−my,mx).(1)\mathbf{m} = \mathbf{n} \times\hat{\mathbf{r}}, \qquad\alpha= \operatorname{atan2}(-m_y,m_x). \tag*{(1)}

where m\mathbf{m} is the normal to the plane of incidence, mx,mym_x,m_y are its two components, and α\alpha is the predicted polarization azimuth. The minus sign follows the coordinate conventions of the camera and AoLP images. Each surface normal is jointly constrained by its corresponding multiview observations.

To account for these ambiguities, we define a periodic angular difference and its penalty for the predicted azimuth α\alpha and observed AoLP ϕ\phi:

h(α,ϕ)=min⁡j∈Z∣α−ϕ−jπ2∣,Ck(α,ϕ)=[e−k(1−4h/π)−e−k1−e−k]2.(2)\begin{aligned} h(\alpha,\phi) &= \min_{j \in\mathbb{Z}} \left|\alpha-\phi-j\frac{\pi}{2}\right|,\\ C_k(\alpha,\phi) &= \left[\frac{e^{-k(1-4h/\pi)}-e^{-k}}{1-e^{-k}}\right]^2. \tag*{(2)} \end{aligned}

where the integer jj enumerates periodically equivalent azimuths, and h∈[0,π/4]h \in[0,\pi/4] is the minimum angular difference. The parameter k>0k > 0 controls the penalty shape, and CkC_k measures the inconsistency between the prediction and observation. We use the observation-weighted mean of CkC_k as the polarization data term. An orientation constraint suppresses back-facing normals, while a smoothness term between neighboring vertex normals reduces unstable local variation. The degree of linear polarization (DoLP) is used to reject unreliable observations and assign weights [35, 44, 45]. The normal variables shared across views and these constraints jointly reduce polarization direction ambiguities.

For each view, we obtain normalized image coordinates (x,y)(x,y) by subtracting the principal point (cx,cy)(c_x,c_y) from pixel coordinates (u,v)(u,v) and dividing by the corresponding focal lengths fx,fyf_x,f_y. Each surface point lies along its viewing ray at depth zz, measured along the camera’s optical axis, and we use its log depth ℓ\ell for subsequent integration.

The orthogonality of surface normals and tangent vectors gives the following log-depth gradients under perspective projection [31]:

gu=∂ℓ∂u=−nxfx(nxx+nyy+nz),gv=∂ℓ∂v=−nyfy(nxx+nyy+nz).(3)\begin{aligned} g_u &= \frac{\partial\ell}{\partial u} = -\frac{n_x}{f_x(n_xx+n_yy+n_z)},\\ g_v &= \frac{\partial\ell}{\partial v} = -\frac{n_y}{f_y(n_xx+n_yy+n_z)}. \tag*{(3)} \end{aligned}

where nx,ny,nzn_x,n_y,n_z are the components of the unit normal in this view, and gu,gvg_u,g_v describe log-depth changes along the two image coordinate directions. Pixels with near-zero denominators are excluded because the conversion is unstable there.

Matching these gradients alone cannot fix the depth of each independent region. We therefore introduce geometric anchors. Sparse anchors are obtained by retriangulating feature tracks from the training views and filtering them using reprojection and foreground support checks. For each independent region without sparse anchors, we add one reference depth anchor from the visual hull [24] to fix its remaining depth degree of freedom. Depth is recovered for each view by solving

ℓ^=arg⁡min⁡ℓ ∑(p,q)∈E(ℓq−ℓp−Δpq)2+λA∑a∈A(ℓa−log⁡Za)2.(4)\hat{\ell} = \arg\min_{\ell}\ \sum_{(p,q)\in E}(\ell_q-\ell_p-\Delta_{pq})^2+\lambda_A\sum_{a\in A}(\ell_a-\log Z_a)^2. \tag*{(4)}

where p,qp,q are adjacent valid pixels, and EE is the set of valid adjacency edges. The predicted log-depth difference Δpq\Delta_{pq} is computed from the average gradient at the endpoints and the actual pixel spacing. The set AA contains depth anchors, and ZaZ_a is the reference depth associated with anchor index aa. The parameter λA\lambda_A controls anchoring strength, and ℓ^\hat{\ell} is the optimal log depth.

To avoid propagating gradients across surface discontinuities, we remove adjacency edges where the visual-hull reference depth changes abruptly. We solve this least-squares problem with LSQR [29] and then exponentiate the solution to recover depth.

Finally, we fuse the per-view depths with equal weights into a TSDF [8], excluding observations outside the foreground and those with invalid depth. Multiview support checks and connected-component processing then produce a unified surface.

Surface-Adaptive Initialization of Oriented Gaussians

Primitives can cover larger areas in relatively flat regions, while curved regions require finer resolution. The two sides of an edge may also have different surface orientations. To accommodate these differences, we take the fused surface and its support information as input. We determine sampling positions from local curvature and edge structure, then use these samples to initialize the orientations and scales of 2D Gaussians.

In local tangent coordinates (ξ,η)(\xi, \eta), we fit the normal height ζ\zeta using a local polynomial approximation [4]:

ζ(ξ,η)=12aξ2+bξη+12cη2+dξ+eη+f,H=[abbc].(5)\zeta(\xi,\eta) = \frac{1}{2}a\xi^2 + b\xi\eta+ \frac{1}{2}c\eta^2 + d\xi+ e\eta+ f, \qquad H = \begin{bmatrix} a & b \\ b & c \end{bmatrix}. \tag*{(5)}

The fit uses distance and support weights with one robust reweighting step. Under a local small-slope approximation, the eigenvalues of HH approximate principal curvatures κ1,κ2\kappa_1,\kappa_2, and its eigenvectors define principal directions.

The directional sampling spacings are

hj=clip⁡(8ϵ∣κj∣+δ11+(θ/θ0)2,hmin⁡,hmax⁡),(6)h_j = \operatorname{clip}\left(\sqrt{\frac{8\epsilon}{\lvert\kappa_j\rvert+\delta}\frac{1}{\sqrt{1+(\theta/\theta_0)^2}}},h_{\min},h_{\max}\right), \tag*{(6)}

where j∈{1,2}j \in\{1,2\}, ϵ\epsilon is the height-error tolerance, δ>0\delta> 0 stabilizes the denominator, and θ\theta is the root-mean-square angle of neighborhood normal variation. For fusion voxel size vv, we set ϵ=0.5v\epsilon= 0.5v, hmin⁡=vh_{\min} = v, hmax⁡=5vh_{\max} = 5v, and θ0=15∘\theta_0 = 15^\circ. In regular regions, candidates are allocated by triangle area and local spacing, followed by variable-radius Poisson-disk rejection [6] with scalar exclusion radius

r=h1h2.(7)r = \sqrt{h_1h_2}. \tag*{(7)}

where rr is a scalar exclusion radius, and h1,h2h_1,h_2 are the sampling spacings along the two tangential directions. Point exclusion is based on rr, while the directional scale differences are retained for subsequent Gaussian initialization.

We detect sharp edges by adjacent-face dihedral angles and partition the surface while accounting for boundary and nonmanifold edges. Samples on the two sides of each sharp edge are generated separately using their respective normals.

After transforming samples and their directions and spacings into training coordinates, we initialize

μ=p,R=[t1,t2,n],(s1,s2)=γ(h1,h2),(8)\mu= p,\qquad R = [t_1,t_2,n],\qquad(s_1,s_2) = \gamma(h_1,h_2), \tag*{(8)}

where pp is the sample position and (t1,t2,n)(t_1,t_2,n) is an orthonormal frame with surface normal nn. Tangents follow principal directions in regular regions; edge samples use tangent planes defined by their respective side normals. We set γ=0.8\gamma= 0.8 and opacity to 0.1, initialize zeroth-order spherical harmonics [11, 23, 42] from fused surface colors, and set higher orders to zero. Sample positions, reference normals, and support confidences serve as surface anchors for subsequent optimization; initial rotations and scales define reference ranges for geometry updates.

Geometry-Guided Primitive Adaptation

We first freeze positions, rotations, and scales, optimizing only spherical harmonic coefficients and opacity. We then enable geometry updates with a lower learning rate. After each update, center displacement, relative rotation, and scale changes are bounded by their respective allowed ranges to control deviations from the reference surface. We also remove the original Clone operation [19, 23] and manage the primitive set through Insert, Split, Merge, and Prune.

Insert targets regions with persistently low coverage. We recover patches through robust integration using polarization-derived normals and reliable boundary depths. After confirming support from at least two other views and the absence of conflicts, we sample these patches to generate new Gaussians. We first check their projected extents, then insert them temporarily and render the scene. The new primitives are retained if coverage increases in missing regions and background coverage, image error, and depth changes remain within their allowed limits. Otherwise, they are removed. We establish surface anchors from their patches, temporarily freeze their geometry, and apply a protection period for primitive updates.

Split refines the local representation provided by existing primitives. Candidates must have large average screen-space gradients over recent training iterations. They must also exhibit at least one of the following conditions: excessive projected coverage, substantial scale growth relative to initialization, large neighborhood normal variation, or a large deviation from the reference tangent plane. Observation-count and opacity criteria must also be satisfied. We construct two child Gaussians along the major axis in the tangent plane:

μ±=μ±0.45smax⁡tmax⁡,smax⁡,±=0.65smax⁡,o±=1−1−o.(9)\mu_{\pm} = \mu\pm0.45s_{\max}t_{\max},\qquad s_{\max,\pm} = 0.65s_{\max},\qquad o_{\pm} = 1-\sqrt{1-o}. \tag*{(9)}

where μ\mu and oo are the parent’s center and opacity, respectively. The quantity smax⁡s_{\max} is the larger of the two tangential scales, and tmax⁡t_{\max} is the corresponding unit tangent direction. The subscripts ++ and −- identify the two children. The children inherit the parent’s minor-axis scale and orientation and replace the parent. The opacity adjustment reduces changes in local coverage caused by splitting.

Merge combines redundant primitives that are spatially close, have aligned normal axes, and are approximately coplanar. We use opacity and coverage area to fit a new center and scales. The merged primitive retains the orientation and anchor of the parent with higher opacity. Its opacity is fitted to the combined coverage of both parents. To measure the local coverage loss caused by merging, we define

Aold(x)=a1(x)+a2(x)−a1(x)a2(x),A_{\mathrm{old}}(x) = a_1(x) + a_2(x) - a_1(x)a_2(x),
Ecov=∑xmax⁡{Aold(x)−Anew(x),0}∑xAold(x).(10)E_{\mathrm{cov}} = \frac{\sum_x \max\{A_{\mathrm{old}}(x) - A_{\mathrm{new}}(x), 0\}}{\sum_x A_{\mathrm{old}}(x)}. \tag*{(10)}

where xx is a sample location in the local tangent plane, and a1,a2a_1, a_2 are the opacity contributions of the two parents. The quantities AoldA_{\mathrm{old}} and AnewA_{\mathrm{new}} represent coverage before and after merging, respectively, and EcovE_{\mathrm{cov}} measures the relative loss of local coverage. A merge is accepted if this ratio does not exceed τcov\tau_{\mathrm{cov}} and the geometric changes remain within their allowed ranges. Each primitive can participate in at most one merge pair per update.

Prune removes primitives whose opacity falls below a threshold while retaining at least one primitive.

The four operations jointly maintain a single Gaussian set while synchronously updating its parameters, surface anchors, and optimizer states. Meshes are extracted by fusing depth maps rendered from the optimized Gaussian model [8, 19].

Training Loss

The total training loss comprises five terms:

L=Lrgb+λ1Lnormal+λ2Lp2pl+λ3Lmask+λ4Lsmooth.(11)\mathcal{L} = \mathcal{L}_{\mathrm{rgb}} + \lambda_1\mathcal{L}_{\mathrm{normal}} + \lambda_2\mathcal{L}_{\mathrm{p2pl}} + \lambda_3\mathcal{L}_{\mathrm{mask}} + \lambda_4\mathcal{L}_{\mathrm{smooth}}. \tag*{(11)}

where Lrgb\mathcal{L}_{\mathrm{rgb}} is the foreground image reconstruction loss, Lnormal\mathcal{L}_{\mathrm{normal}} enforces consistency between rendered normals and depth-derived normals [19], and Lp2pl\mathcal{L}_{\mathrm{p2pl}} constrains the distance from Gaussian centers to their reference tangent planes. The mask loss Lmask\mathcal{L}_{\mathrm{mask}} suppresses opacity coverage outside the foreground silhouette, while Lsmooth\mathcal{L}_{\mathrm{smooth}} includes normal and depth smoothness constraints. The weights λ1,…,λ4\lambda_1,\ldots,\lambda_4 control the corresponding loss terms.

Experiments

Implementation and Dataset

We train for 15000 iterations, freezing geometry for the first 1500. We set λ1=0.5\lambda_1 = 0.5, λ2=0.3\lambda_2 = 0.3, λ3=0.1\lambda_3 = 0.1, and λ4=1\lambda_4 = 1; LsmoothL_{\mathrm{smooth}} combines normal and depth terms with weights 0.2 and 0.8, respectively. We select seven scenes from two datasets: vase, owl, and gnome from PANDORA [9], and BlackLoong, RedLoong, LayingRabbit, and StandingRabbit from PISR [5]. Both datasets contain real-world captures, and PISR additionally provides ground-truth meshes obtained using structured-light scanning.

Comparison

Mesh Reconstruction

As shown in Table 1, Pol2DGS achieves the shortest reported runtime (4.3min), yielding a 5× speedup over PhyGaP [39] while improving both CD and F-score on three of the four scenes. Compared with polarization-guided implicit methods [5, 9, 16], our method substantially reduces training time while maintaining competitive reconstruction quality. It is important to appreciate that, Pol2DGS further outperforms PhyGaP on Black Loong and Red Loong, which contain complex surface structures and fine-scale geometric details. These gains are consistent with the role of polarization-derived surface priors in guiding Gaussian placement and orientation before optimization. Curvature-adaptive sampling allocates samples according to local geometric variation, providing finer coverage in geometrically complex regions [20, 21, 30]. Further strong evidence that our design may help preserve fine surface structures within a limited optimization budget is shown in Figure 4. As shown, Pol2DGS avoids overfitting surface patterns as geometry and produces cleaner, more coherent surfaces than the other competing methods. The overall shape and key geometric structures are likewise preserved.

MethodB. LoongR. LoongL. RabbitS. RabbitTime
CD ↓F-Score ↑CD ↓F-Score ↑CD ↓F-Score ↑CD ↓F-Score ↑
Polarimetric Implicit
PANDORA [9]2.4939.34%2.8744.25%1.5055.27%1.5360.79%12h+
NeRSP [16]2.4241.35%2.3847.74%1.3468.66%0.8677.69%12h+
PISR [5]1.1954.50%1.6472.88%1.3675.92%0.8975.55%24.5m
Explicit
SuGaR [13]2.0134.85%2.8328.65%1.4646.55%1.7841.42%43.0m
2DGS [19]2.4626.54%2.4123.70%1.4340.43%2.4134.03%25.2m
GOF [43]4.1018.21%4.1519.52%2.9327.90%2.9130.03%27.2m
MILo [14]4.0815.13%2.1928.17%1.5240.92%1.7634.34%33.1m
Sparse2DGS [38]8.9921.20%7.8318.62%7.3222.63%7.4322.21%6.1m*
PhyGaP [39]2.2233.69%2.1540.74%1.5355.81%1.4158.22%21.3m
Pol2DGS (Ours)1.9540.46%1.9548.06%1.4356.86%1.6649.28%4.3m

Table 1. Quantitative results of different methods on mesh reconstruction. Red, orange, and yellow indicate the best, second-best, and third-best results, respectively.

Figure 4. Mesh reconstruction results of different methods. Meshes extracted from three real-world scenes in the PANDORA [9] dataset.

Novel View Synthesis

Figure 5 presents a comparison of novel-view renderings across the evaluated methods. The competing Gaussian-based approaches exhibit varying degrees of rendering artifacts, while PhyGaP [39] produces overly smooth results in some regions. Pol2DGS better preserves fine appearance details and sharp image boundaries in the illustrated examples. As shown in Table 2, Pol2DGS achieves the best or second-best scores in most scene-specific PSNR and SSIM [37] comparisons, demonstrating its competitive image reconstruction quality. The performance differences can be attributed to the fact that, PhyGaP [39] enables relighting through polarimetric deferred rendering and self-occlusion-aware environment map construction. While these components improve novel-view

SceneMetricSuGaR [13]2DGS [19]GOF [43]MILo [14]PhyGaP [39]Pol2DGS (Ours)
B. LoongPSNR ↑19.7820.0319.8217.5621.0420.50
SSIM ↑0.5080.6280.5730.4410.7020.635
R. LoongPSNR ↑21.6822.7522.1518.2523.7623.59
SSIM ↑0.6550.7600.7080.6250.8110.775
L. RabbitPSNR ↑17.7924.3022.7116.1226.3626.45
SSIM ↑0.7590.8820.8410.7170.9200.873
S. RabbitPSNR ↑11.3921.8021.3615.1025.6725.43
SSIM ↑0.5360.8570.8100.6580.9030.849

Table 2. Quantitative results of different methods on novel view synthesis.

Figure 5. Qualitative results of different methods on novel view synthesis. We zoom in on BlackLoong’s forelimb and RedLoong’s body to assess geometric reconstruction accuracy in regions with complex details.

synthesis quality, they do not substantially improve mesh reconstruction quality and incur additional computational overhead. Pol2DGS focuses on using surface priors to guide Gaussian initialization and optimization. Beyond the numbers reported and visual comparisons, it is a delicate balance between geometric accuracy and optimization efficiency.

Ablations

We conduct ablation studies to evaluate the contributions of polarization-derived geometric priors and geometry-guided primitive adaptation. Table 3 shows that incorporating geometric priors substantially improves overall reconstruction performance, providing an effective starting point for Gaussian optimization. Building on this initialization, the primitive adaptation strategy further improves geometric accuracy and novel-view synthesis quality by adjusting the primitive distribution to local surface requirements. These findings highlight the complementary roles of the two components. The geometric priors guide the initial positions, orientations, and scales of the Gaussians, while the adaptation strategy refines their distribution during optimization to better represent surface structure and appearance.

2DGS [19]Polarimetric Surface PriorsInsertSplitMergePruneCD ↓PSNR ↑
(a)✓2.4620.03
(b)✓✓2.0920.51
(c)✓✓✓✓✓2.0120.37
(d)✓✓✓✓✓2.1420.48
(e)✓✓✓✓✓2.0020.37
(f)✓✓✓✓✓2.1020.47
(g)✓✓✓✓✓✓1.9720.50

Table 3. Ablation study of Pol2DGS on BlackLoong. Polarimetric surface priors improve both geometric accuracy and rendering quality over 2DGS [19].

We further examine whether polarization-derived normals should continue to supervise Gaussian optimization. We compare three experimental settings: (i) Gaussian optimization is performed with an additional polarization-derived normal constraint. (ii) Gaussian optimization is performed without this constraint, using the same initialization and retaining all other optimization terms. (iii) Depth maps obtained by integrating the estimated normals are directly fused without subsequent Gaussian optimization. Experimentally, it finds that removing the additional constraint yields more accurate surface normals in the evaluated scenes. Referring to Table 4, this clearly shows that polarization-derived normals provide useful geometric initialization but need not remain supervision targets throughout training. Subsequent refinement under photometric supervision and the geometric regularization can correct residual errors in the initialized surface.

B.LoongR.LoongL.RabbitS.Rabbit
w/ Pol. Normal Constraint27.351∘27.351^\circ22.925∘22.925^\circ19.638∘19.638^\circ22.368∘22.368^\circ
w/o Pol. Normal Constraint23.257∘23.257^\circ21.976∘21.976^\circ16.129∘16.129^\circ19.824∘19.824^\circ
Direct Depth Fusion26.608∘26.608^\circ25.376∘25.376^\circ21.833∘21.833^\circ28.066∘28.066^\circ

Table 4. Effect of polarization-derived normal supervision on normal accuracy. We report the mean angular error (MAE), which is lowest across all four scenes without the polarization-derived normal constraint.

Qualitative comparison of ground-truth and predicted normal maps under different supervision settings

Figure 6. Qualitative results of normal maps under different supervision settings. The normal map obtained without an additional polarization-derived normal constraint more closely matches the ground truth.

As a more fine-grained ablation, we compare, on the same evaluated scenes, three initialization from (i) dense depth maps integrated from polarization-derived normals, (ii) sparse COLMAP [33] points, and (iii) adaptively sampled surface points. As in Table 5, Dense initialization improves

SceneDense Pol. Init.Sparse COLMAP Init. [33]Sampled Pol. Init.
B. Loong, CD ↓\downarrow2.042.191.95
B. Loong, Time ↓\downarrow349s291s267s
R. Loong, CD ↓\downarrow2.062.291.94
R. Loong, Time ↓\downarrow331s277s269s
L. Rabbit, CD ↓\downarrow1.611.911.44
L. Rabbit, Time ↓\downarrow338s277s253s
S. Rabbit, CD ↓\downarrow1.672.191.66
S. Rabbit, Time ↓\downarrow330s282s257s

Table 5. Comparison of different initialization strategies. Sampled polarization-based initialization achieves the lowest CD and shortest runtime across all four scenes.

geometric accuracy over COLMAP but increases optimization time. Adaptive sampling achieves higher accuracy and shorter optimization time than the dense alternative. Similar to the earlier finding that no additional constraint is needed for surface normal recovery, increasing the number of initial samples does not necessarily improve reconstruction quality and can introduce additional computational overhead. Selecting representative samples according to local surface geometry provides a more effective starting point for accurate and efficient reconstruction.

Conclusion

Pol2DGS accelerates 2D Gaussian surface reconstruction through geometric priors from polarization and primitive updates guided by geometry. Polarimetric normals, sparse geometric anchors, and anisotropic sampling adapted to local curvature initialize Gaussian positions, orientations, and scales without depth sensors. Experimentally, this initialization improves geometric accuracy and reduces optimization time. Insertion guided by surface geometry, deterministic splitting along the major axis, merging under coverage constraints, and pruning based on opacity further refine the representation. Ablations demonstrate that these operations jointly improve surface accuracy while largely preserving image quality. Across the evaluated scenes and under our timing protocol, optimization averages 4 minutes and 22 seconds. This yields a speedup of approximately 5 to 10 times over the mainstream methods while maintaining competitive mesh accuracy and novel view synthesis quality.

LLM Usage Statement

Large Language Models (LLMs) were used to aid in the writing and polishing of the manuscript. Specifically, we used an LLM to assist in refining the language, improving readability, and ensuring clarity in various sections of the paper. The model helped with tasks such as sentence rephrasing, grammar checking, and enhancing the overall flow of the text. We also used the LLM to assist in identifying code vulnerabilities, improving code style, and ensuring the stability of program execution. The model helped with tasks such as logic correction, bug fixing, and optimizing the overall execution performance of the code.

It is important to note that the LLM was not involved in the ideation, research methodology, or experimental design. All research concepts, ideas, and analyses were developed and conducted by the authors. The contributions of the LLM were solely focused on improving the linguistic quality of the paper, with no involvement in the scientific content or data analysis.

References

  1. [1]Yinlong Bai, Hongxin Zhang, Sheng Zhong, Junkai Niu, Hai Li, Yijia He, and Yi Zhou. MemGS: Memory-efficient Gaussian splatting for real-time SLAM. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 11097–11103, 2025. 2arxiv.org/abs/2509.13536
  2. [2]Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. PatchMatch: A randomized correspondence algorithm for structural image editing. ACM Transactions on Graphics, 28(3):1–11, 2009. 3
  3. [3]Xu Cao, Boxin Shi, Fumio Okura, and Yasuyuki Matsushita. Normal integration via inverse plane fitting with minimum point-to-plane distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2382–2391, 2021. 2DOI
  4. [4]Frédéric Cazals and Marc Pouget. Estimating differential quantities using polynomial fitting of osculating jets. Computer Aided Geometric Design, 22(2):121–146, 2005. 5
  5. [5]Guangcheng Chen, Yicheng He, Li He, and Hong Zhang. PISR: Polarimetric neural implicit surface reconstruction for textureless and specular objects. In Computer Vision – ECCV 2024, pages 205–222. Springer, 2024. 2, 3, 4, 6, 7arxiv.org/abs/2409.14331
  6. [6]Massimiliano Corsini, Paolo Cignoni, and Roberto Scopigno. Efficient and flexible sampling with blue noise properties of triangular meshes. IEEE Transactions on Visualization and Computer Graphics, 18(6):914–924, 2012. 5DOI
  7. [7]Zhaoqiang Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric multi-view stereo. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1558–1567, 2017. 2, 3
  8. [8]Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques, pages 303–312, 1996. 2, 4, 6
  9. [9]Akshat Dave, Yongyi Zhao, and Ashok Veeraraghavan. PANDORA: Polarization-aided neural decomposition of radiance. In Computer Vision – ECCV 2022, pages 538–556. Springer, 2022. 2, 3, 6, 7arxiv.org/abs/2203.13458
  10. [10]Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of Gaussians. In Computer Vision – ECCV 2024, pages 165–181. Springer, 2024. 2arxiv.org/abs/2403.14166
  11. [11]Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 5
  12. [12]Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(8):1362–1376, 2010. 2DOI
  13. [13]Antoine Guédon and Vincent Lepetit. SuGaR: Surface-aligned Gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 3, 6, 7arxiv.org/abs/2311.12775
  14. [14]Antoine Guédon, Diego Gomez, Nissim Maruani, Bingchen Gong, George Drettakis, and Maks Ovsjanikov. MILO: Mesh-in-the-loop Gaussian splatting for detailed and efficient surface reconstruction. ACM Transactions on Graphics, 44(6):1–15, 2025. 3, 6, 7
  15. [15]Bo Guo, Sijia Wen, Yifan Zhao, Jia Li, and Zhiming Zheng. PolarGS: Polarimetric cues for ambiguity-free Gaussian splatting with accurate geometry recovery. IEEE Transactions on Image Processing, 35:5439–5451, 2026. 2, 3arxiv.org/abs/2512.00794
  16. [16]Yufei Han, Heng Guo, Koki Fukai, Hiroaki Santo, Boxin Shi, Fumio Okura, Zhanyu Ma, and Yunpeng Jia. NeRSP: Neural 3D reconstruction for reflective objects with sparse polarized images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11821–11830, 2024. 2, 3, 6, 7arxiv.org/abs/2406.07111
  17. [17]Yufei Han, Bowen Tie, Heng Guo, Youwei Lyu, Si Li, Boxin Shi, Yunpeng Jia, and Zhanyu Ma. PolGS: Polarimetric Gaussian splatting for fast reflective surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 28073–28082, 2025. 3arxiv.org/abs/2509.19726
  18. [18]Shoukang Hu, Tao Hu, and Ziwei Liu. GauHuman: Articulated Gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418–20431, 2024. 2arxiv.org/abs/2312.02973
  19. [19]Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D Gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers, pages 1–11. Association for Computing Machinery, 2024. 2, 3, 5, 6, 7DOI
  20. [20]Hui Huang, Dan Li, Hao Zhang, Uri Ascher, and Daniel Cohen-Or. Consolidation of unorganized point clouds for surface reconstruction. ACM Transactions on Graphics, 28 (5):176:1–176:7, 2009. 7
  21. [21]Hui Huang, Shih Hao Wu, Minglun Gong, Daniel Cohen-Or, Uri Ascher, and Hao Zhang. Edge-aware point set resampling. ACM Transactions on Graphics, 32(1):9:1–9:12, 2013. 7DOI
  22. [22]Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the Symposium on Geometry Processing, pages 61–70, 2006. 3DOI
  23. [23]Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):1–14, 2023. 2, 3, 5arxiv.org/abs/2308.04079
  24. [24]Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(2):150–162, 1994. 4DOI
  25. [25]Yueh-Cheng Liu, Lukas Höllein, Matthias Nießner, and Angela Dai. QuickSplat: Fast 3D surface reconstruction via learned Gaussian initialization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 27851–27861, 2025. 2arxiv.org/abs/2505.05591
  26. [26]Chuanjiang Luo, Xiaoyin Ge, and Yusu Wang. Uniformization and density adaptation for point cloud data via graph laplacian. Computer Graphics Forum, 37(1):325–337, 2018. 2DOI
  27. [27]Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Computer Vision – ECCV 2020, pages 405–421. Springer, 2020. 2arxiv.org/abs/2003.08934
  28. [28]Daisuke Miyazaki, Takuya Shigetomi, Masashi Baba, Ryo Furukawa, Shin­saku Hiura, and Naoki Asada. Polarization-based surface normal estimation of black specular objects from multiple viewpoints. In Proceedings of the Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, pages 104–111, 2012. 4DOI
  29. [29]Christopher C. Paige and Michael A. Saunders. LSQR: An algorithm for sparse linear equations and sparse least squares. ACM Transactions on Mathematical Software, 8(1):43–71, 1982. 4DOI
  30. [30]Mark Pauly, Markus Gross, and Leif Kobbelt. Efficient simplification of point-sampled surfaces. In Proceedings of IEEE Visualization, pages 163–170, 2002. 7
  31. [31]Yvain Quéau, Jean-Denis Durou, and Jean-François Aujol. Normal integration: A survey. Journal of Mathematical Imaging and Vision, 60(4):576–593, 2018. 2, 4arxiv.org/abs/1709.05940
  32. [32]Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in Gaussian splatting. In Computer Vision – ECCV 2024, pages 347–362. Springer, 2024. 2arxiv.org/abs/2404.06109
  33. [33]Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104–4113, 2016. 8DOI
  34. [34]Johannes Lutz Schönberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision – ECCV 2016, pages 501–518. Springer, 2016. 2DOI
  35. [35]Derui Shan, Qian Qiao, Hao Lu, Tao Du, and Peng Lu. PolarGuide-GSDR: 3D Gaussian splatting driven by polarization priors and deferred reflection for real-world reflective scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26000–26009, 2026. 4arxiv.org/abs/2512.02664
  36. [36]Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. NeuS: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems, pages 27171–27183, 2021. 2arxiv.org/abs/2106.10689
  37. [37]Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600–612, 2004. 7pubmed.ncbi.nlm.nih.gov/15376593
  38. [38]Jiang Wu, Rui Li, Yu Zhu, Rong Guo, Jinqiu Sun, and Yanning Zhang. Sparse2DGS: Geometry-prioritized Gaussian splatting for surface reconstruction from sparse views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11307–11316, 2025. 2, 3, 6arxiv.org/abs/2504.20378
  39. [39]Jiale Wu, Xiaoyang Bai, Zongqi He, Weiwei Xu, and Yifan Peng. PhyGaP: Physically-grounded Gaussians with polarization cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7278–7288, 2026. 1, 2, 3, 6, 7arxiv.org/abs/2603.14001
  40. [40]Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multi-view stereo. In Computer Vision – ECCV 2018, pages 785–801. Springer, 2018. 2arxiv.org/abs/1804.02505
  41. [41]Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems, pages 4805–4815, 2021. 2arxiv.org/abs/2106.12052
  42. [42]Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752–5761, 2021. 5arxiv.org/abs/2103.14024
  43. [43]Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics, 43(6):1–13, 2024. 3, 6, 7arxiv.org/abs/2404.10772
  44. [44]Jinyu Zhao, Yusuke Monno, and Masatoshi Okutomi. Polarimetric multi-view inverse rendering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8798–8812, 2023. 4
  45. [45]Jinyu Zhao, Jumpei Oishi, Yusuke Monno, and Masatoshi Okutomi. Polarimetric PatchMatch multi-view stereo. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3476–3484, 2024. 4

Paper details

vlnc:W_9gnkv7cn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

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