FGO-SLAM: Consistent Gaussian SLAM

• FGO-SLAM is a dense Gaussian SLAM system that represents scenes as globally consistent opacity radiance fields, enhancing both pose estimation and surface reconstruction.
• It integrates feature-based tracking with loop closure and bundle adjustment, thereby reducing drift and improving geometric accuracy.
• The method employs geometric regularization and tetrahedral level-set recovery to directly convert 3D Gaussians into precise, complete meshes.

Searching arXiv for papers on FGO-SLAM and closely related methods to ground the article. FGO-SLAM most specifically denotes a 2025 dense Gaussian SLAM system that combines 3D Gaussian Splatting with a globally consistent opacity radiance field, feature-based tracking, global bundle adjustment, and direct surface extraction via tetrahedral level-set recovery (Zhu et al., 1 Sep 2025). In this formulation, the central objective is not only visual rendering quality, but also globally consistent pose estimation, geometrically meaningful scene reconstruction, and mesh extraction from a Gaussian representation. In adjacent literature, however, “FGO-SLAM” can also appear more loosely as a descriptor for factor-graph-optimization-style continuous-time SLAM; this naming overlap is important because the 2025 Gaussian method is architecturally distinct from continuous-time GP-SLAM formulations that merely share an optimization-centered viewpoint (Kapushev et al., 2020).

1. Nomenclature and problem framing

The term FGO-SLAM is associated most directly with “FGO-SLAM: Enhancing Gaussian SLAM with Globally Consistent Opacity Radiance Field,” a dense SLAM system motivated by three deficits in prior Gaussian SLAM systems: limited pose optimization, weak global consistency, and difficulty in extracting surfaces from disconnected Gaussians (Zhu et al., 1 Sep 2025). Its proposed remedy is to treat the Gaussian map as an opacity radiance field that is globally consistent in scene space, while coupling that representation to classical tracking and global adjustment.

A separate strand of literature uses “FGO-SLAM-style” in a broader methodological sense. In “Random Fourier Features based SLAM,” the term denotes continuous-time SLAM methods that perform nonlinear least-squares or MAP estimation in the factor-graph optimization tradition, typically with GP priors and iterative linearization (Kapushev et al., 2020). In that setting, the emphasis is continuous-time trajectory-and-landmark estimation rather than dense Gaussian radiance-field mapping. This suggests that the acronym is polysemous across the SLAM literature: in one usage it identifies a concrete Gaussian SLAM system, while in another it describes an optimization paradigm.

Within dense visual SLAM, the 2025 FGO-SLAM is best understood as part of the post-3DGS family that seeks to preserve the explicit rendering efficiency of Gaussians while introducing stronger geometric structure. Its closest contrasts in the provided literature are GO-SLAM, which pursues global consistency with a neural implicit SDF-based map rather than a Gaussian opacity field, and FGS-SLAM, which also uses Gaussians but focuses on Fourier-guided densification and sparse–dense map fusion rather than global opacity-field consistency and tetrahedral surface extraction (Zhang et al., 2023, Xu et al., 3 Mar 2025).

2. Scene representation as a globally consistent opacity radiance field

FGO-SLAM represents the scene as a set of 3D Gaussian primitives, but it does not treat them merely as renderable splats. Each primitive is written as

$$\mathcal{G}=\delta N(\mathbf{\mu},\mathbf{\Sigma}),$$

where $\delta \in [0,1]$ is the opacity, $\mathbf{\mu}\in\mathbb{R}^{3\times1}$ is the Gaussian mean in world coordinates, $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$ is the covariance parameterized by scale $\mathbf{S}$ and rotation $\mathbf{R}$, and $\mathbf{c}\in\mathbb{R}^{3\times1}$ is the associated color (Zhu et al., 1 Sep 2025).

The paper’s central representational move is to define opacity as a scene-space field. In the local Gaussian coordinate system, for a point $\mathbf{x}_g$,

$$\mathcal{G}(\mathbf{x}_g)=\exp(-\frac12\mathbf{x}_g^T\mathbf{x}_g),$$

with

$\mathbf{x}_g=\mathbf{o}_g+d\mathbf{r}_g.$

Substitution gives a 1D Gaussian along the camera ray: $\delta \in [0,1]$0 The text further states that this function attains its maximum at

$\delta \in [0,1]$1

noting an apparent typographical inconsistency in the denominator notation, but clearly intending a maximizing depth along the ray.

The opacity of a point $\delta \in [0,1]$2 is then defined as the minimum opacity across views: $\delta \in [0,1]$3 with single-view opacity accumulated by alpha-compositing: $\delta \in [0,1]$4 The corresponding per-Gaussian mixing weight is

$\delta \in [0,1]$5

This representation is significant because it turns a collection of splats into a scalar field that can be queried in 3D. A plausible implication is that FGO-SLAM moves Gaussian SLAM closer to the surface-centric logic of implicit-field methods, while retaining explicit Gaussian parameterization and fast rendering pathways. That distinction underlies its later mesh extraction stage.

3. Tracking, loop closure, and global adjustment

FGO-SLAM deliberately avoids relying exclusively on differentiable photometric pose tracking. Its front end is feature-based and structurally similar to ORB-SLAM3: geometric feature points are extracted and matched between consecutive frames, and camera pose is estimated by minimizing reprojection error,

$\delta \in [0,1]$6

where $\delta \in [0,1]$7, $\delta \in [0,1]$8, $\delta \in [0,1]$9, $\mathbf{\mu}\in\mathbb{R}^{3\times1}$0 is the matched 2D keypoint, $\mathbf{\mu}\in\mathbb{R}^{3\times1}$1 is projection, $\mathbf{\mu}\in\mathbb{R}^{3\times1}$2 is the keypoint covariance, and $\mathbf{\mu}\in\mathbb{R}^{3\times1}$3 is the Huber loss. The pose estimate is obtained as

$\mathbf{\mu}\in\mathbb{R}^{3\times1}$4

This design choice is presented as a remedy for the lack of effective pose optimization in earlier Gaussian SLAM systems (Zhu et al., 1 Sep 2025).

Global consistency is enforced through loop detection and bundle adjustment. The system uses a Bag-of-Words model for place recognition and loop-keyframe verification. After loop closure is detected, local keyframe poses are corrected using similarity transforms, after which global bundle adjustment is performed jointly over the full keyframe graph and sparse point cloud: $\mathbf{\mu}\in\mathbb{R}^{3\times1}$5 The optimization is solved by Levenberg–Marquardt.

The role of this stage is broader than trajectory cleanup. The paper states that global refinement of camera poses and sparse map points removes odometry drift and “ghosting” in the Gaussian map (Zhu et al., 1 Sep 2025). This creates a bridge between classical sparse geometric SLAM and dense Gaussian mapping: a conventional robust front end produces globally corrected poses, and the Gaussian field is subsequently optimized against those poses.

This architecture invites comparison with GO-SLAM, which also emphasizes loop closing and online global optimization, but does so with a DROID-SLAM-like dense tracking backbone and a neural implicit surface representation rather than a classical feature front end and Gaussian opacity field (Zhang et al., 2023). The overlap is conceptual rather than representational: both systems treat global consistency as a first-order requirement, but implement it through different map models and tracking subsystems.

4. Mapping objectives and geometric regularization

The mapping stage optimizes the Gaussian scene representation from tracked poses and is explicitly designed to avoid requiring depth priors, allowing both monocular and RGB-D operation (Zhu et al., 1 Sep 2025). The basic color reconstruction loss combines L1 photometric error with D-SSIM: $\mathbf{\mu}\in\mathbb{R}^{3\times1}$6 Rendered color along a ray is computed by

$\mathbf{\mu}\in\mathbb{R}^{3\times1}$7

The full optimization objective is

$\mathbf{\mu}\in\mathbb{R}^{3\times1}$8

with $\mathbf{\mu}\in\mathbb{R}^{3\times1}$9 ranging from 100 to 1000 and $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$0; the implementation details specify $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$1 and $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$2. The two additional terms are depth distortion and normal consistency.

The depth distortion term is

$\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$3

where $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$4 and $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$5 are the depths at which the corresponding $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$6 functions attain their maxima. The stated purpose is to penalize depth spread of contributing Gaussians along the same ray, thereby reducing floating or blurred geometry. An implementation detail is that gradients of $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$7 and $\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$8 are decoupled, so the optimization concentrates on minimizing depth disagreement rather than suppressing the loss by weight manipulation alone.

The normal consistency term is

$\mathbf{\Sigma}\in\mathbb{R}^{3\times3}$9

where $\mathbf{S}$0 is the normal of the 3D Gaussian and $\mathbf{S}$1 is the normal estimated from the depth-map gradient. This enforces agreement between Gaussian geometry and depth-derived local surface orientation.

These regularizers are central to the method’s claim of improving geometry rather than only rendering. The ablation study reports that including regularization produces smoother extracted surfaces, especially on floors and other large planar regions (Zhu et al., 1 Sep 2025). This suggests that FGO-SLAM is best interpreted not as a pure radiance-field optimizer, but as a geometry-regularized Gaussian field whose opacity and surface orientation are constrained to support later surface recovery.

5. Surface extraction from Gaussians

A defining contribution of FGO-SLAM is direct mesh extraction from the Gaussian field. Instead of using TSDF fusion or Poisson reconstruction, the method constructs tetrahedral grids from the Gaussian primitives themselves (Zhu et al., 1 Sep 2025).

For each 3D Gaussian primitive, a bounding box is constructed at the $\mathbf{S}$2 level. Tetrahedral grids are then generated using the box centers and vertices. To prevent tetrahedralization from connecting unrelated Gaussians, the method removes a tetrahedral cell if the edge connecting two Gaussian primitives exceeds the sum of their maximum scales. This cell filtering acts as a geometric sanity check.

Opacity is then evaluated at tetrahedral vertices using the method of Gaussian Opacity Fields, but with selected keyframes rather than all training views to reduce redundant computation. The subsequent extraction procedure consists of three steps: evaluating opacity at tetrahedral vertices, identifying the level set by binary search, and applying Marching Tetrahedra to generate the mesh.

The significance of this pipeline is methodological. Standard Gaussian SLAM often provides visually plausible rendering without a principled surface model. FGO-SLAM instead supplies a scalar opacity field over space, and that scalar field supports isosurface extraction in a conventional geometric sense. The paper emphasizes that it can produce complete meshes and that, on Replica, mesh quality is visually better than NeRF-based methods in hole filling and global consistency (Zhu et al., 1 Sep 2025).

In relation to GO-SLAM, the contrast is again instructive: GO-SLAM extracts a mesh through marching cubes on queried SDF values from an implicit neural field, whereas FGO-SLAM uses tetrahedral level-set recovery over an opacity field induced by 3D Gaussians (Zhang et al., 2023). Both systems therefore translate dense learned or optimized fields into explicit surfaces, but they do so from different field semantics.

6. Empirical performance, positioning, and related systems

The reported implementation runs on an Intel i7-12700K CPU and an NVIDIA RTX 3090 Ti with 24 GB of memory, supports both monocular and RGB-D modes, and uses key frames rather than all views for surface extraction (Zhu et al., 1 Sep 2025). On Replica Office0, the reported average processing speed is tracking/frame $\mathbf{S}$3 s, mapping/frame $\mathbf{S}$4 s, and FPS $\mathbf{S}$5.

For tracking on TUM RGB-D, using ATE-RMSE [cm], the reported averages are 2.50 for NICE-SLAM, 2.11 for ESLAM, 1.31 for Photo-SLAM, 3.30 for SplaTAM, 1.58 for MonoGS, 1.56 for the monocular version of FGO-SLAM, and 0.98 for the RGB-D version. On ScanNet, also using ATE-RMSE [cm], the corresponding averages are 10.00, 7.51, 11.90, 13.51, 8.35, and 7.37 for NICE-SLAM, ESLAM, SplaTAM, MonoGS, FGO-SLAM (Mono), and FGO-SLAM (RGB-D), respectively. The paper characterizes these results as state-of-the-art tracking accuracy and mapping performance (Zhu et al., 1 Sep 2025).

For mapping and rendering on Replica, averaged over 8 scenes, the reported metrics are PSNR, SSIM, LPIPS, and Depth L1 [cm]. The RGB-D version of FGO-SLAM reports PSNR 38.35, SSIM 0.973, LPIPS 0.084, and Depth L1 0.71, while the monocular version reports PSNR 34.13, SSIM 0.956, LPIPS 0.094, and Depth L1 3.24. The paper states that the RGB-D version is best in rendering quality overall, whereas the monocular version is weaker in Depth L1, which is consistent with the absence of depth priors.

The ablation on global adjustment reports, on ScanNet scene 0000_00, ATE-RMSE 7.59 and PSNR 23.54 without global adjustment, versus ATE-RMSE 6.74 and PSNR 25.12 with global adjustment. This supports the paper’s thesis that loop closing and global bundle adjustment improve both tracking and mapping (Zhu et al., 1 Sep 2025).

Relative to other Gaussian systems in the supplied literature, FGO-SLAM occupies a distinct position. FGS-SLAM uses Fourier-domain analysis for adaptive Gaussian densification, separates sparse tracking and dense reconstruction maps, and reports real-time performance at 36 FPS on Replica and TUM RGB-D datasets, with 33.04 FPS on Replica/Office0 and 44.09 FPS on TUM RGB-D (Xu et al., 3 Mar 2025). Its principal concern is content-aware densification and efficient 3D tracking via GICP, not global opacity-field consistency or tetrahedral mesh extraction. GO-SLAM, by contrast, emphasizes online full bundle adjustment and implicit SDF mapping, reaching around 8 FPS on Replica while targeting global consistency in neural dense SLAM (Zhang et al., 2023).

A common misconception is that Gaussian SLAM quality can be judged adequately by rendering metrics alone. The FGO-SLAM formulation argues against that view by making surface extraction, geometric regularization, and global pose correction central design objectives (Zhu et al., 1 Sep 2025). Another possible source of confusion is the acronym itself: in continuous-time SLAM, “FGO-SLAM-style” may denote an optimization family rather than this specific Gaussian opacity-field system (Kapushev et al., 2020). Within current dense visual SLAM, however, FGO-SLAM most precisely refers to the 2025 method that couples feature-based tracking, global adjustment, and a globally consistent opacity radiance field over 3D Gaussians.