GSO-SLAM: 3D Gaussian Splatting SLAM

• GSO-SLAM is an emerging SLAM framework that uses differentiable 3D Gaussian Splatting to achieve dense, photorealistic reconstructions with high tracking precision.
• It integrates bidirectional coupling between visual odometry and 3D map optimization using robust photometric losses and gradient-based Gaussian initialization.
• The framework supports real-time, sub-centimeter tracking and scalable mapping through modular submap strategies and EM-based joint optimization.

GSO-SLAM denotes an emerging family of SLAM frameworks that leverage 3D Gaussian Splatting (3DGS) as a differentiable, compact scene representation, with various coupling strategies to classical visual odometry, direct or feature-based SLAM, and scene understanding. Originating from advances in both neural rendering and real-time mapping, GSO-SLAM variants are characterized by dense, photorealistic reconstructions, sub-centimeter tracking performance, and scalability to large environments, enabled by the fusion of differentiable splatting, robust front-end tracking (direct, feature, or hybrid), and modular optimization/initialization approaches. Recent systems, including “GSO-SLAM: Bidirectionally Coupled Gaussian Splatting and Direct Visual Odometry” (Yeon et al., 12 Feb 2026), “MGSO: Monocular Real-time Photometric SLAM with Efficient 3D Gaussian Splatting” (Hu et al., 2024), and “Large-Scale Gaussian Splatting SLAM” (LSG-SLAM) (Xin et al., 15 May 2025), define the current state of the art.

1. System Architectures and Pipeline Variants

GSO-SLAM systems share a two-threaded (or modular) architecture: a front end for pose and depth estimation, and a back end for 3DGS map construction and refinement. The following key modules are identified across recent approaches:

• Front-end Tracking: Most implementations use a Direct Sparse Odometry (DSO) or DSO-like backbone, which tracks photometrically stable points and optimizes windowed bundle adjustment over poses and depths (Yeon et al., 12 Feb 2026, Hu et al., 2024). Some systems, such as LSG-SLAM (Xin et al., 15 May 2025), exploit stereo input and integrate geometric (feature-matching, ICP) and photometric priors for robust initialization.
• Back-end Mapping: The scene is modeled as a set of 3D or 2D Gaussians, continuously optimized via differentiable rendering losses that include photometric, depth, and structural (e.g., normal consistency, scale regularization) terms.
• Bidirectional Coupling: A defining feature of GSO-SLAM (Yeon et al., 12 Feb 2026) is the bidirectional optimization between the front-end (VO) and the 3DGS map: VO-derived depths guide initial splat construction and map regularization, while rendered splat depths regularize the VO’s map, closing the loop between tracking and mapping.
• Submaps and Scalability: For large outdoor or unbounded scenarios, as in LSG-SLAM (Xin et al., 15 May 2025), the workspace is partitioned into GS submaps with local descriptors, enabling memory-efficient representations and globally consistent pose graphs.

2. Gaussian Splatting Scene Representation

All GSO-SLAM methods adopt an anisotropic 3D (or sometimes 2D) Gaussian mixture, where each splat is parameterized by mean $\mu_i \in \mathbb{R}^3$, covariance $\Sigma_i \in \mathbb{R}^{3 \times 3}$ (often decomposed as $$R_i \mathrm{diag}(\sigma_{i,1}^2, \sigma_{i,2}^2, \sigma_{i,3}^2) R_i^\top$$), color $c_i$, and opacity $\alpha_i$. Differentiable rasterization accumulates alpha-weighted colors along camera rays:

$$C(p) = \sum_{i} \alpha_i' c_i, \quad \alpha_i' = \alpha_i \prod_{j < i}(1 - \alpha_j)$$

with Gaussians sorted by projected depth (Hu et al., 2024). Scene rendering is fully differentiable with respect to both pose and splat parameters.

Gaussian initialization is a critical bottleneck. Recent approaches (e.g., (Yeon et al., 12 Feb 2026)) propose closed-form derivation of splat covariances from image gradient distributions and multi-view associations, which dramatically accelerates convergence (requiring an order of magnitude fewer gradient steps than KNN- or isotropic-seeded schemes).

3. Joint Optimization Objectives and Optimization Algorithms

GSO-SLAM optimization is typically staged as alternating parallel modules:

• Photometric SLAM Objective: Front-end bundle adjustment minimizes photometric reprojection errors for tracked pixels across a sliding window of keyframes:

$$E_{\mathrm{photo}} = \sum_{i, j, u} \rho(I_i(u) - I_j(\pi(T_{ij} X_u)))$$

where $\rho$ is a robust loss and $\pi$ denotes the camera projection (Hu et al., 2024).

• Gaussian Splatting Map Objective: The 3DGS map is refined using a combination of pixelwise color matching ($\ell_1$), SSIM-based similarity, depth consistency, and regularization penalties for covariance and opacity:

$\Sigma_i \in \mathbb{R}^{3 \times 3}$0

with periodic split/merge/prune steps to maintain compactness (Hu et al., 2024, Yeon et al., 12 Feb 2026).

• EM Formulation and Bidirectional Coupling: Some systems (notably (Yeon et al., 12 Feb 2026)) formalize the entire system as an Expectation-Maximization procedure, alternating Gaussian parameter refinement (E-step, conditioned on VO) and pose/depth update (M-step, regularized by the 3DGS map).

4. Scalability and Submap Strategies for Large-Scale SLAM

To address memory and computational constraints in large environments, LSG-SLAM (Xin et al., 15 May 2025) divides the trajectory and the global map into sequential submaps, each holding a local set of Gaussians and associated keyframe database. Only the active submap and its neighbors reside in GPU memory at any time. Loop closure is performed at the submap level, using place recognition (TransVPR descriptors), feature-matching, and joint optimization of loop-pose constraints. Structure refinement is conducted submap-wise, supporting anisotropic scaling of Gaussians to capture local geometry detail.

This submap-based design allows runtime and memory to remain approximately constant regardless of global trajectory length, enabling scaling to full KITTI/EUROC sequences on a single GPU (Xin et al., 15 May 2025).

5. Tracking, Mapping, and Loop Closure Performance

Empirical results across leading GSO-SLAM implementations confirm high accuracy and real-time performance:

• On Replica (monocular), GSO-SLAM (Yeon et al., 12 Feb 2026) achieves tracking ATE RMSE of 0.46 cm, PSNR of 34.48 dB, and depth L₁ error of 8.12 cm, outperforming contemporaries such as Photo-SLAM (ATE 1.03 cm, PSNR 30.91 dB).
• On TUM-RGBD (monocular), GSO-SLAM achieves avg. ATE RMSE of 3.07 cm and PSNR 20.52 dB.
• For large-scale outdoor mapping (EuRoC, KITTI), LSG-SLAM (Xin et al., 15 May 2025) achieves ATE RMSE of 0.17 m on EuRoC (all sequences, no loop closure), and 0.06 m with loop closure—exceeding MonoGS and Photo-SLAM baselines. Mapping PSNR reaches 31.4 dB with structural refinement.

Frame rates of 20–30 Hz are consistently reported on both desktop (RTX4090) and laptop (RTX3080) hardware.

6. Limitations and Directions for Future Work

Despite the strengths of GSO-SLAM, several limitations remain:

• Dynamic Scene Handling: Current systems assume scene rigidity; moving objects induce “ghost” splats or degraded local maps. No explicit outlier masking or dynamic object segmentation is performed (Xin et al., 15 May 2025).
• Initialization Sensitivity: Accurate Gaussian initialization is crucial; suboptimal parameterization increases map redundancy and/or convergence time (Yeon et al., 12 Feb 2026).
• Parameter Robustness: Performance and map compactness depend on tuning silhouette thresholds, depth regularization weights, and split/prune heuristics.
• Loop Closure Beyond Submaps: While current submap pose-graphs are effective, global loop closure and dense BA at the Gaussian level remain computationally expensive at city scale.

Research directions include semantic integration for robust tracking (as in Go-SLAM (Pham et al., 2024)), real-time dynamic-object masking, learned/fast splat initialization, and monocular or monocular+IMU extensions for scale-robust outdoor mapping.

7. Comparison with Related Gaussian Splatting SLAM Systems

To contextualize the GSO-SLAM paradigm, the following table summarizes distinctive features and reported benchmarks for major Gaussian Splatting SLAM variants:

System (Year)	Front-End	Splat Init	Loop Closure	ATE RMSE (Replica)	PSNR (Replica)	Submaps
GSO-SLAM [(Yeon et al., 12 Feb 2026), 2026]	DSO	Gradient-based EM	No	0.46 cm	34.48 dB	No
MGSO [(Hu et al., 2024), 2024]	DSO	DSO point cloud	No	1.11 cm	31.41 dB	No
LSG-SLAM [(Xin et al., 15 May 2025), 2025]	Stereo+SuperPt	Local geometric+KNN	Submap-level	0.17 m (EuRoC)	31.4 dB	Yes

These results indicate that GSO-SLAM and its derivatives deliver state-of-the-art photometric and geometric SLAM performance, with unique combinations of EM-based optimization, fast splat initialization, submap scalability, and robust feature or direct tracking.

---

References:

• GSO-SLAM: Bidirectionally Coupled Gaussian Splatting and Direct Visual Odometry (Yeon et al., 12 Feb 2026)
• MGSO: Monocular Real-time Photometric SLAM with Efficient 3D Gaussian Splatting (Hu et al., 2024)
• Large-Scale Gaussian Splatting SLAM (Xin et al., 15 May 2025)
• Go-SLAM: Grounded Object Segmentation and Localization with Gaussian Splatting SLAM (Pham et al., 2024)