LighthouseGS: 3D Splatting for Indoor Views

• LighthouseGS is a 3D Gaussian Splatting-based indoor view synthesis framework that leverages plane scaffold assembly and mobile priors for precise novel view rendering.
• It integrates rough ARKit poses and monocular depth estimates to overcome challenges from narrow baselines and textureless indoor scenes.
• The method uses stable Gaussian pruning, residual pose refinement, and photometric corrections to enhance rendering quality and support AR object placement.

Searching arXiv for the specified LighthouseGS paper and closely related work to ground the article in current sources. LighthouseGS is a 3D Gaussian Splatting (3DGS)-based indoor novel view synthesis framework designed for panorama-style mobile captures in which a handheld camera follows rotation-dominant motion with very narrow baselines. It targets a regime where conventional Structure-from-Motion and standard 3DGS initialization are unreliable, especially in textureless indoor scenes, by combining rough geometric priors from mobile device poses and monocular depth estimation with explicit planar structure, a new initialization method called plane scaffold assembly, a stable pruning strategy, and geometric and photometric corrections (Han et al., 8 Jul 2025).

1. Problem setting and scope

LighthouseGS addresses high-fidelity novel view synthesis for casual, panorama-style indoor captures acquired with a single handheld mobile device. In the intended capture mode, users rotate the device around a roughly fixed center in a “lighthouse sweep,” so the motion is rotation-dominant with very narrow baselines. This acquisition pattern is convenient for mobile users, but it produces limited parallax, ill-conditioned pose estimation, and sparse or unstable 3D point initialization when conventional pipelines depend on SfM or COLMAP (Han et al., 8 Jul 2025).

The framework is specifically motivated by indoor scenes, where large textureless planes such as walls and ceilings, as well as repetitive patterns, make geometry estimation ambiguous and cause floaters or holes when optimization is driven mainly by photometric gradients. The system therefore injects geometric structure and rough priors directly into the 3DGS pipeline. Its inputs are RGB frames $\{I_t\}$ captured with pano-style rotation, mobile device poses $\{\Pi_t\}$ from ARKit/IMU, and per-frame monocular depth $D_t$ and normals $N_t$ estimated by pretrained models Depth Anything V2 and DSINE. Camera intrinsics $K$ are computed from the device, and the world frame is inherited from ARKit (Han et al., 8 Jul 2025).

This scope distinguishes LighthouseGS from generic 3DGS pipelines. Its contribution is not merely a different regularizer or a pose-refinement add-on; the method is organized around the assumption that indoor environments are often plane-dominant and that rough mobile priors, while noisy, are sufficiently informative to replace SfM when combined with plane-wise alignment. A plausible implication is that LighthouseGS is most effective when capture follows the panoramic motion pattern it was designed for, rather than arbitrary free-form trajectories.

2. Plane scaffold assembly and structure-aware initialization

The central initialization mechanism is plane scaffold assembly, whose goal is to produce a globally and locally consistent set of 3D points $X_T$ with attached plane normals $N_T$. The process begins by back-projecting the monocular depth of the first frame with pose $\Pi_0$ to form an initial global set $X_0$, and by storing normals from $N_0$. For each subsequent frame $\{\Pi_t\}$0, LighthouseGS first performs image-wise global alignment by projecting accumulated global points $\{\Pi_t\}$1 onto frame $\{\Pi_t\}$2, denoted $\{\Pi_t\}$3, and fitting a per-frame scale and shift,

$\{\Pi_t\}$4

by minimizing

$\{\Pi_t\}$5

This resolves global scale and shift but not local misalignment (Han et al., 8 Jul 2025).

LighthouseGS then applies plane-wise local alignment. Plane segments are obtained using “efficient mean-shift” clustering over depth and normals, yielding masks $\{\Pi_t\}$6 for piece-wise planar segments. Each plane segment $\{\Pi_t\}$7 is parameterized as

$\{\Pi_t\}$8

with unit normal $\{\Pi_t\}$9 and offset $D_t$0. For each plane, the method refines scale and shift starting from $D_t$1:

$D_t$2

and solves

$D_t$3

The plane-adjusted depths are then back-projected to 3D and merged into the global scaffold:

$D_t$4

This sequential global-to-local strategy replaces SfM for initializing 3D Gaussians and is intended to preserve dense coverage in textureless regions (Han et al., 8 Jul 2025).

When pixels are lifted by a plane, the method uses standard ray-plane intersection. For pose $D_t$5 with camera center $D_t$6, pixel $D_t$7, direction $D_t$8, and plane $D_t$9, LighthouseGS solves

$N_t$0

with the numerical stability check $N_t$1 (Han et al., 8 Jul 2025).

The scaffold is then converted into a Gaussian initialization. Means $N_t$2 are initialized from downsampled scaffold points $N_t$3. Covariances $N_t$4 start from kNN-based isotropic scales and are compressed along the plane normal direction to make Gaussians thin and surface-aligned. With an orthonormal basis $N_t$5 for the plane, the initialization is

$N_t$6

with $N_t$7. Colors are represented with spherical harmonics, with the DC term initialized from average observed color and higher orders near zero (Han et al., 8 Jul 2025).

3. 3DGS formulation, stable pruning, and correction modules

Once initialized, LighthouseGS optimizes a standard 3DGS scene representation, but with plane-aware stabilization and mobile-specific correction modules. For Gaussian $N_t$8 with mean $N_t$9 and covariance $K$0,

$K$1

The covariance is parameterized as $K$2, where $K$3 is a quaternion-derived rotation and $K$4 contains diagonal scales. After projection through the camera Jacobian, the 2D covariance is

$K$5

and the screen-space weight at pixel $K$6 is

$K$7

Pixel colors are then produced by front-to-back alpha compositing,

$K$8

with view-dependent color $K$9 from SH coefficients (Han et al., 8 Jul 2025).

A major departure from standard adaptive density control is the stable pruning strategy. Standard 3DGS ADC tends to prune large or transparent Gaussians and may create holes in textureless regions; later cloning or splitting can then generate floaters near the camera. LighthouseGS instead preserves “high-confidence” Gaussians with large opacity among ADC’s pruning candidates, using the rule “keep Gaussian $X_T$0 if $X_T$1, even if it is oversized.” The rationale given in the paper is that, in smooth textureless planes, a small number of larger opaque Gaussians can faithfully approximate surfaces and stabilize optimization (Han et al., 8 Jul 2025).

The framework also incorporates residual pose refinement and photometric correction. Camera poses from ARKit are represented as $X_T$2 and corrected by learning a residual transform $X_T$3, composed as

$X_T$4

This residual refinement is optimized through photometric rendering error. To address auto-exposure and white-balance drift, the method applies a per-frame affine tone model,

$X_T$5

where $X_T$6 is a per-channel gain and $X_T$7 is a per-channel bias (Han et al., 8 Jul 2025).

Geometry regularization is likewise structure-aware. LighthouseGS introduces normal alignment,

$X_T$8

flatness,

$X_T$9

normal TV smoothness,

$N_T$0

and depth-to-normal consistency,

$N_T$1

where $N_T$2 and $N_T$3 are rendered normal and depth, and $N_T$4 is the prior normal from DSINE (Han et al., 8 Jul 2025).

4. Optimization objectives and implementation regime

The final training objective is

$N_T$5

with photometric term

$N_T$6

using $N_T$7 and $N_T$8, and geometric regularization

$N_T$9

with $\Pi_0$0 and $\Pi_0$1 (Han et al., 8 Jul 2025).

Training uses the gsplat rasterizer and Adam optimizer for 30k total iterations. Densification is enabled for the first 15k iterations, with ADC run every 100 steps. The gradient thresholds for ADC are 0.0008 on real data and 0.0004 on synthetic data. LighthouseGS also uses AbsGS accumulation to reduce gradient collisions and recover fine texture. Intrinsics are fixed from ARKit during optimization, and no extra gradient clipping or coarse-to-fine schedule is required beyond the thresholds described above (Han et al., 8 Jul 2025).

At test time, LighthouseGS can further refine unseen views by freezing Gaussians and optimizing test-view poses and tone parameters for 20k iterations. This point is methodologically significant because it indicates that the framework treats residual geometric and photometric inconsistency as an expected property of panorama-style mobile captures rather than as a preprocessing nuisance. A plausible implication is that LighthouseGS is best understood as a joint representation-and-correction system, not only as a 3DGS initializer.

5. Datasets, empirical results, and downstream uses

The evaluation uses both real and synthetic indoor scenes. The real-world dataset contains five scenes—Meeting room, Dressing room, Pantry, Lounge, and Conference room—captured at up to approximately 100 frames per scene with resolution $\Pi_0$2 using iPhone ARKit via NeRFCapture, with auto-exposure on and rotation-dominant pano sweep motion. The synthetic benchmark contains five scenes—Playroom, Bedroom, Primary bedroom, Living room, and Studio—rendered in Blender Cycles at $\Pi_0$3, with simulated camera radius 20–30 cm, injected ARKit-like drift, and auto-exposure variations (Han et al., 8 Jul 2025).

Baselines include 3DGS, DNGaussian, Scaffold-GS, and GeoGaussian, all initialized with LighthouseGS’s plane scaffold plus ARKit poses because COLMAP fails under pano motion. The paper reports that LighthouseGS achieves the best performance in all scenes. On real data, improvements versus the second best are roughly $\Pi_0$4 PSNR, $\Pi_0$5 SSIM, and $\Pi_0$6 LPIPS. On synthetic data, the corresponding gains are $\Pi_0$7 PSNR, $\Pi_0$8 SSIM, and $\Pi_0$9 LPIPS. A concrete example is the Conference room scene on real data, where 3DGS† obtains $X_0$0 and LighthouseGS obtains $X_0$1 for PSNR/SSIM/LPIPS (Han et al., 8 Jul 2025).

The ablation studies isolate the effect of individual modules. On the meeting room scene, the baseline (COLMAP-initialized) yields $X_0$2; adding residual pose gives $X_0$3; adding color correction gives $X_0$4; adding stable optimization gives $X_0$5; and the full model gives $X_0$6. In a separate initialization ablation, Spherical SfM yields $X_0$7, ARKit+COLMAP yields $X_0$8, and the plane scaffold yields $X_0$9. The paper identifies stable optimization as critical for clean, hole-free textureless surfaces, and the plane scaffold as a decisive factor under pano motion (Han et al., 8 Jul 2025).

Qualitatively, LighthouseGS is reported to reduce floaters and holes on textureless walls and ceilings, sharpen object details such as slippers in the dressing room, and minimize color inconsistency around windows with changing exposure. The reconstructed Gaussians can be rendered with a sphere-based rasterizer to produce wide-FOV panoramic imagery from novel viewpoints. The method also supports AR object placement because accurate rendered depth, alpha, and per-pixel normals enable physically plausible insertion of virtual objects, with virtual objects resting on surfaces with correct occlusions (Han et al., 8 Jul 2025).

6. Limitations, interpretation, and terminological disambiguation

The method has several explicit limitations. Plane detection errors or poorly segmented planes can bias depth alignment and initialization. Very large pose drift or extreme auto-exposure changes may exceed the correction capacity. Non-planar clutter or scenes that are not plane-dominant may reduce gains. The framework also depends on reasonable monocular depth and normal estimates, so strong errors in these priors degrade the scaffold. Computationally, densification and plane segmentation add overhead, although experiments are conducted on a single RTX 4090 (Han et al., 8 Jul 2025).

These limitations clarify a common misconception: LighthouseGS is not a general solution to all mobile 3DGS capture regimes, but a framework tuned to indoor scenes with many planes, panorama-style motion, and rough but usable mobile priors. The practical guidance reported with the method is consistent with this interpretation: the camera should be kept roughly centered and at steady height, the sweep should be slow, and some overlap between views is beneficial for global-to-local depth alignment (Han et al., 8 Jul 2025).

The name also requires disambiguation. The 2020 paper "Lighthouse: Predicting Lighting Volumes for Spatially-Coherent Illumination" introduced a multiscale volumetric lighting field for spatially-varying illumination and explicitly predates 3D Gaussian Splatting; its description of a principled integration with GS labels that integration “LighthouseGS,” but this is a separate lighting-plus-GS construction rather than the 2025 indoor panorama-style mobile capture framework (Srinivasan et al., 2020). The term also appears in a different, geospatial sense as “Lighthouse Geospatial Search” in the shoreline-distance system LIGHTHOUSE (Beukema et al., 23 Jun 2025), and the low-light enhancement paper "LL-GaussianMap: Zero-shot Low-Light Image Enhancement via 2D Gaussian Splatting Guided Gain Maps" refers to a “LighthouseGS-style system” only as an external design context rather than as the indoor novel view synthesis method itself (Chen et al., 22 Jan 2026).

Within the arXiv literature, the primary referent of LighthouseGS is therefore the 2025 framework for indoor structure-aware 3D Gaussian Splatting under panorama-style mobile captures. Its technical identity is defined by plane scaffold assembly, opacity-aware stable pruning, residual pose refinement, per-frame affine color correction, and plane-aware geometric regularization, all organized around the failure modes of rotation-dominant, narrow-baseline indoor acquisition (Han et al., 8 Jul 2025).