LiDAR-centric 3D Gaussian Splatting

• The paper integrates LiDAR into every stage of the 3D Gaussian lifecycle, ensuring accurate geometric mapping and metric fidelity.
• It employs LiDAR-based initialization, allocation, and confidence scoring to enhance real-time localization and detailed scene reconstruction.
• Applications include SLAM, autonomous driving simulation, and large-scale 3D reconstruction, demonstrating improved performance in sparse or dynamic environments.

LiDAR-centric 3D Gaussian Splatting denotes a family of 3D Gaussian Splatting (3DGS) methods in which LiDAR is treated as the primary source of metric geometry, while cameras typically provide appearance, dense photometric supervision, or completion in LiDAR-blind regions. In this literature, LiDAR is used not only to seed Gaussian centers, but also to determine Gaussian allocation, local surface orientation, confidence, depth constraints, scan-to-map registration, and, in some systems, direct LiDAR rendering itself (Lee et al., 23 Jan 2025). A useful reading of the field is that it spans a spectrum from LiDAR-augmented initialization to fully LiDAR-centered representations and optimization loops, with applications including robot localization, outdoor SLAM, aerial remote sensing, autonomous-driving simulation, calibration, and large-scale reconstruction (Jiang et al., 2024).

1. Conceptual scope and taxonomy

A useful distinction suggested by the literature is between LiDAR-augmented, LiDAR-guided, and LiDAR-centric 3DGS. In the weakest form, LiDAR is fused into the initial point cloud before otherwise standard 3DGS training. "LiDAR-3DGS" (Lim et al., 2024) is explicit about this design: it fuses LiDAR and SfM point clouds before training and does not modify the underlying 3DGS algorithm. "LiDAR-enhanced 3D Gaussian Splatting Mapping" (Shen et al., 7 Mar 2025) moves further by using LiDAR for pose estimation, extrinsic refinement, Gaussian initialization, and sparse depth supervision, but it still presents itself as a LiDAR-enhanced mapping framework rather than a LiDAR-native renderer. By contrast, "Geometry-Aware Gaussian Splatting (GeomGS)" (Lee et al., 23 Jan 2025) argues that a localization-ready Gaussian map requires LiDAR to influence the scene representation itself, the confidence assigned to primitives, the optimization weighting, and the eventual localization algorithm.

Several papers explicitly criticize the reduction of LiDAR to sparse seed points. "TCLC-GS: Tightly Coupled LiDAR-Camera Gaussian Splatting for Autonomous Driving" (Zhao et al., 2024) states that urban-scene GS methods that initialize 3D Gaussians directly with 3D LiDAR points underutilize LiDAR data capabilities. "LI-GS: Gaussian Splatting with LiDAR Incorporated for Accurate Large-Scale Reconstruction" (Jiang et al., 2024) goes further by converting LiDAR into plane-constrained multimodal Gaussian Mixture Models (GMMs) that are used during initialization, optimization, density control, and mesh extraction. A plausible implication is that, in this literature, LiDAR-centricity is less about the mere presence of LiDAR and more about where geometric authority resides during optimization.

The field also contains deliberately hybrid systems. "A Constrained Optimization Approach for Gaussian Splatting from Coarsely-posed Images and Noisy Lidar Point Clouds" (Peng et al., 12 Apr 2025) is best understood, in its own framing, as LiDAR-assisted or hybrid image-LiDAR 3DGS rather than fully LiDAR-centric, because LiDAR provides the initial 3D geometry and coarse SLAM pose prior, while refinement remains driven mainly by image rendering and image-based geometric consistency. This distinction matters because many claims about LiDAR-centric 3DGS hinge on whether LiDAR remains a persistent supervisory signal or only an initialization source.

2. Gaussian representations and LiDAR-conditioned initialization

Most LiDAR-centric systems retain the basic 3DGS primitive

$$G(x)=e^{-\frac{1}{2}(x-\mu)^{\top}\Sigma^{-1}(x-\mu)},$$

with anisotropic covariance

$\Sigma = RSS^{\top}R^{\top},$

and standard front-to-back alpha compositing, while changing how Gaussians are initialized, shaped, or supervised (Lee et al., 23 Jan 2025). This is true of GeomGS, LiV-GS, RSGaussian, LVI-GS, TCLC-GS, Structured-Li-GS, and Gaussian-LIC2, all of which preserve differentiable splatting while inserting LiDAR into the geometry pipeline (Xiao et al., 2024).

Initialization strategy is therefore one of the main differentiators. GeomGS initializes from LiDAR scans accumulated in the world frame rather than from SfM points (Lee et al., 23 Jan 2025). LVI-GS initializes Gaussians from colourized LiDAR points and uses all colourized LiDAR points for the first frame (Zhao et al., 2024). TCLC-GS does not initialize directly from raw LiDAR points; instead, LiDAR sweeps are merged into a global point cloud, used to train a hierarchical octree implicit representation, converted into a colorized mesh, and that mesh then initializes the Gaussian representation (Zhao et al., 2024). Structured-Li-GS anchors Gaussians to a sub-sampled dense colorized LiDAR point cloud, initializes their ellipsoids from local surface geometry, rotates them to match point-cloud normals, and flattens them along the normal direction (Weng et al., 25 Jun 2026).

Not all systems retain full anisotropic 3DGS. LiGSM simplifies the Gaussian primitive to an isotropic, view-independent form with center, RGB color, opacity, and scalar radius, explicitly sacrificing full anisotropic covariance in favor of a lighter multimodal mapper (Shen et al., 7 Mar 2025). At the other extreme, LI-GS and Splat-LOAM replace volumetric 3D Gaussians with surface-like primitives. LI-GS adopts 2D Gaussian surfels, parameterized by center, tangent directions, radii, normal, opacity, and spherical-harmonic appearance, precisely to improve surface alignment in large-scale scenes (Jiang et al., 2024). Splat-LOAM likewise uses 2D Gaussian surface primitives defined by opacity, centroid, two tangential vectors, and a 2D scale vector, and renders only range, normal, and opacity rather than RGB (Giacomini et al., 21 Mar 2025).

Several recent methods push initialization beyond direct LiDAR seeding. "DensifyBeforehand" (Patt et al., 24 Nov 2025) uses sparse mobile LiDAR together with Metric3D-v2 monocular depth to build a dense point cloud beforehand, then initializes 3DGS from a LiDAR-assisted, ROI-aware sampled version of that dense cloud. "Gaussian-LIC2" (Lang et al., 5 Jul 2025) initializes from colorized LiDAR points in LiDAR-covered regions and supplements LiDAR-blind regions through a zero-shot depth completion model, SPNet, that combines RGB appearance cues with sparse LiDAR depth. According to its abstract, "ES-Gaussian" (Chen et al., 2024) also targets low-cost indoor reconstruction by combining a low-altitude camera with single-line LiDAR and introducing a single-line-LiDAR-guided 3DGS initialization.

3. LiDAR supervision, allocation, and geometry control

The decisive step from LiDAR-augmented to LiDAR-centric 3DGS is the use of LiDAR during optimization, not just at initialization. GeomGS exemplifies this by assigning each Gaussian a learnable Geometric Confidence Score (GCS) and coupling it to the nearest-neighbor distance between Gaussian centers and the accumulated LiDAR cloud. Its central probabilistic distance constraint,

$$\mathcal{L}_{prob} = \frac{1}{N} \sum_{i=1}^{N} \left( \ln (1-\gamma_i) + \frac{d_i}{(1-\gamma_i)} \right),$$

makes high-confidence Gaussians remain close to LiDAR-supported structure while allowing lower-confidence Gaussians greater photometric freedom (Lee et al., 23 Jan 2025). The method is explicit that this is not a full LiDAR ray model; LiDAR enters through nearest-point geometric support rather than beam-wise visibility or occupancy.

Other systems replace pointwise supervision with richer geometric priors. LI-GS constructs plane-constrained multimodal 4D GMMs from LiDAR point clouds, then uses them for initialization, geometric supervision, geometry-aware density control, and mesh extraction (Jiang et al., 2024). TCLC-GS derives a colorized mesh and dense depth maps from LiDAR-camera data, then uses dense mesh-rendered depth to supervise Gaussian optimization (Zhao et al., 2024). RSGaussian constrains densification and pruning by nearest-neighbor relations to LiDAR points, projects split directions onto local LiDAR tangent planes, and optimizes depth and plane-consistency losses derived from LiDAR-supported geometry (Yao et al., 2024). Structured-Li-GS combines photometric, flattening, offset, depth, and normal losses, all guided by dense colorized LiDAR point clouds, while avoiding Gaussian densification (Weng et al., 25 Jun 2026).

Allocation under a fixed Gaussian budget has emerged as another explicitly LiDAR-centric topic. GTLR-GS formulates Gaussian allocation as a geometry-texture-aware sampling problem over a registered LiDAR point cloud. For each point, it computes a curvature score $K_i$ from the eigenvalues of the local covariance matrix and a texture score $T_i$ from local RGB variation, then samples according to

$$P_i = \frac{\alpha K_i + \beta T_i}{\sum_{j=1}^{N} (\alpha K_j + \beta T_j)},$$

with $\alpha=\beta=0.5$ in experiments (Fang et al., 24 Mar 2026). The same work then uses curvature-adaptive splitting during training and confidence-aware metric depth regularization from projected LiDAR depth maps.

A related but distinct development is the use of LiDAR to reduce or replace adaptive density control. DensifyBeforehand argues that cloning-heavy adaptive density control produces floating artifacts, redundant Gaussians, and high memory cost, then moves densification out of the training loop by using LiDAR anchors to calibrate monocular depth globally and locally before 3DGS optimization (Patt et al., 24 Nov 2025). A plausible implication is that LiDAR-centricity can manifest either as stronger online supervision or as stronger offline geometry preparation.

4. SLAM, localization, and online Gaussian mapping

LiDAR-centric 3DGS has become closely tied to SLAM and localization because LiDAR-grounded Gaussian maps are attractive as both renderable scene models and registration targets. LiV-GS is explicit about this design: it directly aligns sparse, discrete LiDAR scans with a continuous differentiable Gaussian map in large-scale outdoor scenes, using covariance as the shared attribute between LiDAR points and Gaussian primitives (Xiao et al., 2024). Its front-end uses weighted point-to-plane scan-to-Gaussian alignment in a local Gaussian submap, and its Conditional Gaussian Constraint propagates reliable LiDAR-supported geometry into regions outside the LiDAR field of view.

LVI-GS extends this direction to a real-time LiDAR-Visual-Inertial SLAM architecture with two parallel threads, a hyper-primitives module containing point clouds, voxels, and Gaussians, LiDAR-based Gaussian initialization, LiDAR depth supervision, and pyramid-based training. The paper reports that photorealistic quality for a keyframe is reached after about 105 iterations, with roughly 3 seconds total runtime in its example (Zhao et al., 2024). Gaussian-LIC2 further tightens the coupling by using a continuous-time LiDAR-Inertial-Camera front-end, sparse LiDAR depth supervision for Gaussian optimization, dense depth completion for LiDAR-blind regions, and optional photometric constraints from the Gaussian map back into the factor graph for robust tracking under LiDAR degradation (Lang et al., 5 Jul 2025).

Localization-specific Gaussian maps appear in GeomGS. There, the learned GCS values weight ICP between a query LiDAR scan and the Gaussian map, after which pose is refined photometrically in the spirit of iNeRF. In the reported configuration, each localization iteration uses one weighted ICP step followed by 20 image-refinement steps, and the full process runs for 20 iterations (Lee et al., 23 Jan 2025). GSFusion takes a complementary route: it separates a sparse surfel map for localization from a dense Gaussian map for rendering, then uses surfel-to-surfel constraints in a global pose graph to achieve the global alignment precision required by photorealistic Gaussian maps (Park et al., 31 Jul 2025).

At the most LiDAR-only end of the spectrum, Splat-LOAM discards cameras entirely and builds a LiDAR odometry and mapping pipeline that exclusively relies on Gaussian surface primitives. It uses spherical projection, differentiable LiDAR rendering of range and normals, and a tracking loss that combines point-to-plane registration with image-space range consistency in the LiDAR panorama (Giacomini et al., 21 Mar 2025). This suggests that LiDAR-centric 3DGS is no longer restricted to multimodal radiance-field systems; it also includes purely LiDAR-driven Gaussian SLAM.

5. Application domains and domain-specific formulations

The domain breadth of LiDAR-centric 3DGS is unusually wide. Autonomous-driving reconstruction and rendering motivated early tightly coupled LiDAR-camera systems such as TCLC-GS and large-scale geometric systems such as LI-GS (Zhao et al., 2024). Outdoor SLAM and robotics drove LiV-GS, LVI-GS, Gaussian-LIC2, GSFusion, and Splat-LOAM (Xiao et al., 2024). Aerial remote sensing motivated RSGaussian, which addresses wide baselines, long sensor-to-scene distances, sparse angular coverage, and geo-alignment via distortion-aware LiDAR-image registration and LiDAR-constrained densification (Yao et al., 2024). Targetless extrinsic calibration motivated GeoP-Calib, which treats the Gaussian scene as a LiDAR-faithful geometric proxy rather than as a purely image-driven renderer (Kwak et al., 18 Jun 2026). Direct multi-sensor simulation motivated SplatAD, which is the first 3DGS-based method for realistic, real-time rendering of both camera and LiDAR data in dynamic autonomous-driving scenes (Hess et al., 2024).

Paper	Primary domain	LiDAR role
"ES-Gaussian" (Chen et al., 2024)	Indoor reconstruction	low-altitude camera + single-line LiDAR
"RSGaussian" (Yao et al., 2024)	Aerial remote sensing NVS	LiDAR-constrained densification and depth/plane supervision
"SplatAD" (Hess et al., 2024)	Autonomous-driving simulation	direct LiDAR rendering in spherical coordinates
"GeoP-Calib" (Kwak et al., 18 Jun 2026)	Targetless calibration	LiDAR-faithful Gaussian proxy with dense depth anchoring
"Splat-LOAM" (Giacomini et al., 21 Mar 2025)	LiDAR odometry and mapping	LiDAR-only Gaussian surface representation

Among these, SplatAD is the clearest example of LiDAR-centricity at the sensor-model level. It introduces a dedicated LiDAR rendering branch that projects Gaussians into spherical coordinates,

$$\mu_i^S = \begin{bmatrix} \phi\ \omega\ r \end{bmatrix} = \begin{bmatrix} \arctan2(y,x)\ \arcsin(z/r)\ \sqrt{x^2+y^2+z^2} \end{bmatrix},$$

renders range, intensity, and ray-drop probability, models rolling-shutter-like scan timing for spinning LiDAR, and uses beam-aware tiling instead of image-like tiling (Hess et al., 2024). This is a qualitatively different use of LiDAR from systems that only use sparse depth regularization.

Large-scale reconstruction papers emphasize another recurring theme: LiDAR often matters most when scenes are unbounded, sparse-view, weak-texture, or metrically demanding. LI-GS reports large gains in geometric accuracy for large-scale reconstruction through LiDAR-derived GMM supervision (Jiang et al., 2024). Structured-Li-GS explicitly targets fewer Gaussians and reconstruction without densification by training on dense colorized LiDAR point clouds (Weng et al., 25 Jun 2026). GTLR-GS frames the entire problem as geometry-conditioned allocation and refinement under a fixed representational budget (Fang et al., 24 Mar 2026). This suggests that LiDAR-centric 3DGS has become not only a multimodal rendering topic, but also a methodology for controlling model compactness and metric fidelity.

6. Limitations, misconceptions, and research directions

A common misconception is that LiDAR initialization alone makes a method LiDAR-centric. Several papers argue the opposite. LiDAR-3DGS is explicit that it is a simple multimodal extension in which LiDAR point clouds are fused with the image/SfM point cloud before standard 3DGS training (Lim et al., 2024). TCLC-GS argues that direct LiDAR-point seeding underutilizes LiDAR’s geometric structure and continuity (Zhao et al., 2024). GeomGS criticizes approaches that use LiDAR only as initial points or simple Euclidean constraints and instead makes confidence, map optimization, and localization depend on LiDAR (Lee et al., 23 Jan 2025). A plausible implication is that LiDAR-centricity should be judged by whether LiDAR shapes the Gaussian lifecycle and downstream use of the map, not merely its starting point.

Another misconception is that stronger rendering quality necessarily implies stronger geometry. GeoP-Calib shows the opposite for calibration: it argues that image-centric 3DGS can optimize photometric quality while degrading the metric faithfulness of the Gaussian proxy, and it reports better calibration accuracy even when rendering PSNR is lower than image-centric baselines (Kwak et al., 18 Jun 2026). GeomGS makes a related point from robotics: photorealistic Gaussian maps are not necessarily metrically faithful enough for localization (Lee et al., 23 Jan 2025).

The field also contains an important terminological caution around “full integration.” GeomGS states that LiDAR is fully integrated into 3D Gaussian primitives via a probabilistic approach, yet the method still uses nearest-point Gaussian-to-LiDAR proximity rather than full LiDAR beam rendering or occupancy integration (Lee et al., 23 Jan 2025). This should not be read as a contradiction; it indicates that “full integration” is often relative to earlier 3DGS methods, not equivalent to a full probabilistic LiDAR sensor model.

Across papers, several limitations recur. Many systems depend on accurate sensor calibration and accurate or known poses to accumulate LiDAR in world coordinates and project it into images (Yao et al., 2024). Methods that rely on dense point-cloud anchors or projected depth supervision inherit sensitivity to SLAM quality, LiDAR density, and LiDAR-camera alignment (Weng et al., 25 Jun 2026). RSGaussian notes that performance weakens when LiDAR coverage is sparse and that satellite-borne scenarios remain difficult (Yao et al., 2024). Splat-LOAM explicitly identifies motion distortion and lack of loop closure as open issues (Giacomini et al., 21 Mar 2025). SplatAD models dynamic actors as rigid only and leaves non-rigid actor modeling and broader LiDAR sensor support as future work (Hess et al., 2024). GTLR-GS notes the need to extend the framework to more diverse mobile LiDAR conditions (Fang et al., 24 Mar 2026).

This suggests a current research trajectory rather than a settled endpoint. The literature appears to be converging on three design principles: geometry-first Gaussian allocation, LiDAR-aware supervision or rendering throughout optimization, and downstream use of Gaussian maps as metric objects for localization, calibration, or simulation. What remains open is how far this convergence can be pushed toward full beam-consistent LiDAR physics, large-scale online global consistency, and robust operation under sparse mobile LiDAR, asynchronous sensing, dynamic scenes, and severe calibration error.