PolarGS: Polarization-Enhanced 3D Reconstruction

Updated 7 December 2025

PolarGS is a reconstruction method that integrates polarization cues with 3D Gaussian Splatting to resolve ambiguities in reflective and textureless regions.
It employs DoLP and AoLP measurements to detect specular artifacts and constructs Color Refinement Maps for photometric correction.
Quantitative results demonstrate improved geometric accuracy, significantly reducing Chamfer Distance compared to RGB-only approaches.

PolarGS is a reconstruction method that integrates polarization cues into 3D Gaussian Splatting (3DGS) pipelines to resolve photometric ambiguities in multiview surface reconstruction, with particular emphasis on reflective and textureless regions. By leveraging the Degree and Angle of Linear Polarization (DoLP, AoLP) from polarization images, PolarGS detects and corrects specular artifacts, recovers geometry in low-texture areas, and achieves improved geometric accuracy over standard RGB-only approaches. The design is modular and applies polarization-guided photometric correction and Gaussian densification as enhancements to existing 3DGS frameworks (Guo et al., 30 Nov 2025).

1. Challenges in Multiview Surface Reconstruction

Photometric 3D reconstruction frameworks, especially 3D Gaussian Splatting, rely on multi-view RGB consistency to optimize scene geometry and appearance. Performance degrades in two key photometrically ambiguous scenarios:

Reflective Surfaces: Materials exhibiting strong specular reflection (e.g., metal, glass) generate highlights that are highly view-dependent and not aligned with intrinsic surface color. Purely RGB photometric losses incorrectly fit these highlights, leading to local geometry artifacts such as mesh ripples, spikes, and holes.
Textureless Surfaces: Homogeneous regions lacking discernible color or texture (e.g., smooth white plastic) provide insufficient correspondence cues for multi-view photometric terms, resulting in undersampled or flat geometry with missing details.

In both cases, the RGB photometric terms fail: specular areas break the Lambertian assumption, and textureless regions are degenerate under photometric matching (Guo et al., 30 Nov 2025).

2. Polarization-Guided Photometric Correction

PolarGS incorporates optics-based polarization measurements, providing physical cues for distinguishing specular and diffuse components, and encoding partial surface orientation even under uncontrolled lighting. The core photometric correction mechanism proceeds as follows:

2.1 Degree of Linear Polarization (DoLP) Computation

A polarization camera captures four intensity frames at 0°, 45°, 90°, and 135°. Stokes vector components are calculated:

$s_0 = I_0 + I_{90}$ (total intensity)
$s_1 = I_0 - I_{90}$
$s_2 = I_{45} - I_{135}$

The Degree of Linear Polarization is:

$\rho(x, y) = \frac{\sqrt{s_1^2 + s_2^2}}{s_0}, \quad \rho \in [0, 1]$

High $\rho$ values identify strongly polarized (typically specular) light, while low $\rho$ values denote diffuse reflection.

2.2 Reflective Region Localization

A binary specular mask $M_s(x, y)$ is generated by thresholding DoLP:

$M_s(x, y) = [\rho(x, y) > \tau_s]$

where typically $\tau_s \in [0.5, 0.6]$ . Overexposed pixels (where $s_0$ approaches saturation) are also masked, as photometric consistency is unreliable there.

To correct the appearance in masked regions, PolarGS derives Color Refinement Maps:

Diffuse Map ( $I_{\text{diff}}$ ): Attenuates specular spikes,

$I_{\text{diff}}(x, y) = \frac{ s_0(x, y) - [I_{\max}(x, y) - I_{\min}(x, y)] }{2 }$

Chromaticity Map ( $I_{\text{chro}}$ ): Addresses overexposure by re-chromatizing,

$I_{\text{chro}}(x, y) = \rho_d \cdot \frac{I_{\text{prop}}(x, y)}{\sum_{\lambda} I_{\text{prop}}(x, y) + \overline{I_{\text{diff}}}(x, y) }$

Here, $I_{\text{prop}}(x, y)$ refers to a propagated diffuse color from neighboring non-reflective pixels with similar polarimetric reference intensity (PRI). $\overline{I_{\text{diff}}}$ is the mean per-channel value of $I_{\text{diff}}$ .

2.4 Reflective-Aware Photometric Loss

Only in reflective (high-DoLP or overexposed) regions, PolarGS imposes specialized corrections via:

$L_{\text{ref}} = L_s(C, I_{\text{diff}}) + L_o(C, I_{\text{chro}})$

where $L_s$ and $L_o$ are combinations of $L_1$ and D-SSIM losses on masked regions. The total color loss is:

$L_c = L_{\text{non}}(C, I_{\text{rgb}}) + \lambda_{\text{ref}} L_{\text{ref}}$

$L_{\text{non}}$ is the multi-view loss on non-reflective pixels, and $\lambda_{\text{ref}}$ balances the correction.

3. Pipeline Integration

The PolarGS correction is integrated in the Gaussian optimization loop as shown in the pseudocode below.

for iteration in range(max_iter):
    for view in training_views:
        # 1) Polarization preprocessing
        load I_0, I_45, I_90, I_135
        compute s0, s1, s2
        rho = sqrt(s1**2 + s2**2) / s0
        M_s = (rho > tau_s)
        M_o = (s0 near saturation)

        # 2) Build Color Refinement Maps
        I_pri = max(R, G, B) - min(R, G, B)  # over angles
        propagate diffuse color -> I_prop
        I_diff = (s0 - (I_max - I_min)) / 2
        I_chro = rho_d * I_prop / (sum(I_prop) + mean(I_diff))

        # 3) Rendering and losses
        C = render_rgb(gaussians, view)
        L_non = photometric_loss(C, I_rgb)
        L_ref = L_s(C, I_diff, mask=M_s) + L_o(C, I_chro, mask=M_o)
        L_c = L_non + lambda_ref * L_ref

        # 4) Update Gaussian parameters
        update_gaussians()

This approach ensures that Gaussian color and geometry optimization is supervised by both conventional photometric consistency and physically meaningful polarization cues.

4. Quantitative Results and Ablation Analysis

PolarGS demonstrates improvements in geometric reconstruction accuracy, particularly as measured by Chamfer Distance (CD). Representative results include:

Method	NeISF (CD ×10⁻³)	RMVP3D Lion	Average
GNeRP	3.20	5.83	4.52
NeRSP	3.33	5.12	4.23
NeISF	3.76	4.97	4.37
GaussianPro	3.75	5.67	4.71
GOF	2.36	7.95	5.16
PGSR	1.84	6.18	4.01
PolarGS	1.65	3.77	2.71

Ablation studies confirm that disabling the photometric correction module increases CD from 1.65 to 2.91, and removing polarization-driven densification increases it to 2.13 (Guo et al., 30 Nov 2025).

5. Dependencies, Limitations, and Materials Considerations

Polarization Hardware: PolarGS requires a multi-angle polarization camera. Standard RGB sensors cannot natively provide the required Stokes measurements.
Material Model Assumptions: The DoLP framework presumes a diffuse-specular dichotomy; materials with strong transmission, layered structures, or nearly perfect dielectrics can disrupt the segmentation.
Circular Polarization: The framework ignores the Stokes $s_3$ (circular) parameter, which could enhance discrimination for birefringent materials.
Extreme Cases: Very dark or highly saturated areas may yield noisy or ill-defined CRMs, resulting in localized artifacts.

6. Extensions and Future Directions

Potential research avenues include:

End-to-End CRM Learning: Direct learning of Color Refinement Maps from raw polarization stacks could further stabilize corrections.
Full Stokes Imaging: Inclusion of the circular polarization component ( $s_3$ ) to extend correction schemes beyond linear polarization.
Joint Optimization: Integrating lighting, BRDF, and full polarization cues in a unified neural-radiance/Gaussian framework (Guo et al., 30 Nov 2025).
Generalization to Other Modalities: A plausible implication is that similar polarization-guided correction schemes may benefit related tasks, such as HDR reconstruction (Ting et al., 2022), intrinsic image separation (Wen et al., 2021), and photometric calibration in X-ray and solar polarimetry contexts (Rankin et al., 2021, Jaeggli et al., 2022).

PolarGS constitutes a framework-agnostic optical augmentation to photometric multi-view 3D reconstruction, with direct physical segmentation of reflection modalities and a pipeline to propagate this guidance for improved color and geometry recovery in non-Lambertian and degenerate scenes.