Dense Geometric Gaussian Representation

• Dense geometric Gaussian representation is a method that employs Gaussian primitives to encode spatial parameters such as position, scale, and rotation for mapping and reconstruction.
• It integrates task-specific features—ranging from color and semantic channels to feature fields—to optimize visual localization, SLAM, and HD mapping.
• The approach uses varied methodologies from geometry-only to geometry-plus-appearance with adaptive density control to improve efficiency and accuracy.

Dense geometric Gaussian representation denotes a family of explicit spatial representations in which Gaussian primitives are used as geometric carriers rather than only as radiance carriers. Across visual localization, dense SLAM, online HD mapping, feed-forward reconstruction, and point-cloud densification, the shared pattern is a map or latent field whose state is organized around Gaussian position, scale, rotation, opacity, and, depending on the task, feature, color, or semantic channels. In this formulation, geometry is primary: Gaussian support, orientation, and projection behavior define correspondence, rendering, or planning structure, while appearance may be auxiliary, fixed, compressed, or removed entirely when it is not task-essential (Li et al., 18 May 2026, Zhou et al., 8 Jul 2026, Lyu et al., 30 Jun 2026, Zhang et al., 8 Apr 2026).

1. Conceptual scope

Dense geometric Gaussian representation is not a single canonical object. In some works it is a scene map in Euclidean $\mathbb{R}^3$; in others it is a BEV-plane representation, a surface-aligned 2D Gaussian field, or a local distribution attached to each input point. GaussianMap, for example, defines a set of Gaussian primitives on the BEV plane for online vectorized HD map construction, while Sparse2DGS uses planar 2D Gaussian primitives for sparse-view surface reconstruction, PU-Gaussian models the local neighborhood around each point using anisotropic 3D Gaussian distributions, and Fast-2DGS treats an image as a dense set of explicit 2D Gaussian primitives (Lyu et al., 30 Jun 2026, Wu et al., 29 Apr 2025, Khater et al., 24 Sep 2025, Wang et al., 14 Dec 2025). This suggests that “dense” is task-relative: it may refer to dense BEV support after splatting, dense surface coverage by tangent-plane splats, dense per-pixel feature rendering, or dense point generation by sampling learned local Gaussian distributions.

A second source of variation is whether the representation is geometry-only or geometry-plus-appearance. LiteLoc explicitly argues that the color field inherited from Feature 3DGS is “functionally useless” for localization and replaces the coupled color-feature primitive

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$

with the color-free form

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$

producing what it describes as a compact localization-specific dense feature field (Li et al., 18 May 2026). GeoGS goes further and removes appearance modeling entirely, retaining only spatial parameters in a geometry-only primitive

$$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$

with $\mathbf{s}_i\in\mathbb{R}^2$ because the primitive is a surface-aligned disk-like element rather than a full volumetric ellipsoid (Zhou et al., 8 Jul 2026). By contrast, SGS-SLAM and autonomy-oriented Gaussian maps attach semantic channels or language features to the same Gaussian geometry, making the primitive simultaneously geometric, photometric, and semantic (Li et al., 2024, Ong et al., 17 May 2025).

A common misconception is that “geometry-only” implies the absence of photometric supervision. The papers do not support that reading. GeoGS is geometry-only at the representation level, yet it is optimized with single-view and multi-view geometric and photometric supervision; Sparse2DGS freezes color and feature attributes but still uses image reconstruction; and SplatMap combines geometric constraints with photometric consistency while keeping a standard renderable 3DGS map (Zhou et al., 8 Jul 2026, Hu et al., 13 Jan 2025).

2. Primitive models and rendering operators

The most common primitive model follows standard anisotropic Gaussian splatting. GeoTexDensifier, SplatMap, and GS-SLAM all use covariance parameterizations of the form

$\Sigma = R S S^T R^T,$

with $S$ encoding scale and $R$ encoding rotation, so each primitive is an oriented ellipsoid with explicit local extent (Jiang et al., 2024, Hu et al., 13 Jan 2025, Yan et al., 2023). GS-SLAM represents the scene as

$$\mathbf{G} = \{ G_i:(\mathbf{X}_i, \mathbf{\Sigma}_i, \Lambda_i, \boldsymbol{Y}_i) \mid i=1,...,N\},$$

and renders color and depth by front-to-back alpha compositing,

$$\hat{\mathbf{C}}=\sum_{i \in N} \mathbf{c}_i \alpha_i \prod_{j=1}^{i-1}\left(1-\alpha_j\right), \qquad \hat{D} = \sum_{i \in N} d_i \alpha_i \prod_{j=1}^{i-1}\left(1-\alpha_j\right),$$

which turns the Gaussian cloud into a dense differentiable RGB-D map (Yan et al., 2023).

Several papers specialize this model to make geometry more surface-like. Sparse2DGS uses planar 2D Gaussian primitives with

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$0

so the third scale is zero and the primitive is a tangent-plane disk rather than a volume (Wu et al., 29 Apr 2025). GeoGS likewise uses a disk-like primitive whose local frame is

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$1

and renders only geometric quantities such as depth and normals (Zhou et al., 8 Jul 2026). GeoGaussian imposes thinness on smooth connected areas by initializing surface Gaussians with

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$2

so the primitive behaves like a thin sheet aligned with the local surface normal (Li et al., 2024).

Task-specific renderers replace color with other dense signals. LiteLoc renders a dense feature field using normalized alpha compositing,

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$3

and trains the map against SuperPoint features rather than RGB (Li et al., 18 May 2026). GaussianMap uses BEV-plane Gaussians

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$4

then splats them to a dense BEV tensor by posterior responsibilities,

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$5

Here the dense field is induced by superposition rather than by storing a feature at every cell (Lyu et al., 30 Jun 2026).

3. Geometry construction, densification, and map growth

A central issue is how Gaussian geometry is created and refined. One strategy is to remove task-irrelevant channels so that primitive density follows geometry rather than photometric detail. LiteLoc attributes the redundancy of STDLoc to a coupled color-feature field that forces “excessively dense Gaussian primitives,” and reports that removing color allows Gaussian density to align to the lower-frequency structure of the feature field (Li et al., 18 May 2026). Another strategy is to keep the full primitive model but densify it under geometric constraints. GeoTexDensifier supplements adaptive density control with a texture-aware densification rule and a geometry-aware splitting rule based on monocular depth and normal priors. Its key acceptance criterion is the Validation of Depth Ratio Change,

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$6

with $\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$7, so child splats inconsistent with local surface depth are rejected (Jiang et al., 2024).

Anchor-based methods move the support of the representation from image pixels to 3D geometry. AnchorSplat extracts anchors by back-projecting predicted depths and poses and then downsampling by farthest point sampling,

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$8

after which each anchor predicts four Gaussians with centers

$\Theta_i = \{x_i, q_i, s_i, \alpha_i, c_i, f_i\}$9

The result is an anchor-aligned Gaussian representation that is independent of image resolution and number of views (Zhang et al., 8 Apr 2026). GeoGS-SLAM also uses learned geometry priors to drive map growth, but in an online monocular setting: VGGT predicts poses, intrinsics, depth, and confidence; DoG responses determine informative image locations; and a new primitive is spawned only when

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$0

with initial scale

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$1

This makes primitive placement explicitly dependent on predicted 3D structure and camera geometry (Gao et al., 13 Jul 2026).

Several systems use hierarchical or hybrid map structure instead of a single uniform Gaussian family. PointSLAM++ maintains primary anchors, which “cannot undergo splitting or deletion,” and secondary anchors, which are inserted when voxel-wise average gradient satisfies

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$2

and culled by opacity decay when they become unnecessary (Wang et al., 10 Jan 2026). HGS-Mapping partitions the scene into $\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$3, $\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$4, and $\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$5, corresponding to Sphere Gaussians for sky, 2D Gaussian Plane primitives for the road surface, and standard 3D Gaussians for roadside scenery and general 3D objects (Wu et al., 2024). GeoGaussian uses point-cloud normals to identify smoothly connected areas, initializes thin Gaussians there, and constrains clone and split operations to the local tangent plane so that offspring remain co-planar with the parent (Li et al., 2024).

4. Supervision regimes and geometry-first optimization

The supervision used for dense geometric Gaussian representations is as varied as the representations themselves. LiteLoc is an extreme example of distilled feature-field optimization. It renders dense feature maps, supervises them with a SuperPoint feature map $\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$6, and uses only

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$7

with no RGB reconstruction loss (Li et al., 18 May 2026). Sparse2DGS also suppresses appearance flexibility by freezing Gaussian color and MVS feature values, then adds a cosine-similarity feature loss

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$8

together with Direct Gaussian Primitive Regularization. Its reparameterized disk sampling

$\Theta'_i = \{x_i, q_i, s_i, \alpha_i, f_i\},$9

allows multiview feature consistency to backpropagate directly to primitive position, orientation, and scale (Wu et al., 29 Apr 2025).

Geometry-only systems typically combine direct geometric supervision with indirect photometric constraints. GeoGS defines

$$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$0

where $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$1 contains normal-consistency, depth-smoothness, distortion, and optional depth/normal prior terms, and $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$2 contains multiview $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$3, SSIM, and NCC consistency after depth-based reprojection (Zhou et al., 8 Jul 2026). GeoGS-SLAM couples standard 3DGS rendering with a photometric loss

$$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$4

and a geometric loss

$$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$5

then jointly optimizes poses and Gaussians within the active window (Gao et al., 13 Jul 2026). SplatMap combines photometric consistency with edge-aware normal regularization, using

$$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$6

where $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$7 is computed on rendered normals derived from depth (Hu et al., 13 Jan 2025).

Multi-channel optimization extends the same principle to semantics. SGS-SLAM renders depth, color, and semantic images from the same Gaussian map and optimizes tracking and mapping with losses over all three channels, while the semantic channel is stored directly as a per-Gaussian semantic color and composited with the same visibility model as RGB (Li et al., 2024). A plausible implication is that dense geometric Gaussian representation is increasingly being used as a common substrate onto which different supervisory signals—depth, color, learned features, normals, semantics, and language embeddings—can be attached without changing the underlying geometric primitive.

5. Systems, domains, and empirical behavior

The representation has been instantiated across a wide range of tasks.

Domain	Representative role of Gaussians	Example
Visual localization	Color-free dense feature field for 2D–3D matching	LiteLoc
Dense SLAM	Explicit RGB-D, semantic, or geometry-only scene map	GS-SLAM, SGS-SLAM, GeoGS-SLAM
Online HD mapping	Adaptive BEV-plane latent for vectorized map prediction	GaussianMap
Surface reconstruction	Surface-aligned 2D Gaussian splats under sparse views	Sparse2DGS
Feed-forward scene reconstruction	Anchor-aligned renderable 3D Gaussians from geometric priors	AnchorSplat
Point-cloud upsampling	Local anisotropic Gaussian distributions for direct sampling	PU-Gaussian

The numerical behavior reported in these papers shows that geometry-first Gaussian design can improve both efficiency and reconstruction quality. In localization, LiteLoc reduces the number of Gaussians from $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$8k to $$\mathcal{G}_i=\{\mathbf{\mu}_i,\mathbf{s}_i,\mathbf{q}_i,\mathbf{\alpha}_i\},$$9k and memory from $\mathbf{s}_i\in\mathbb{R}^2$0MB to $\mathbf{s}_i\in\mathbb{R}^2$1MB, reported as a $\mathbf{s}_i\in\mathbb{R}^2$2 storage reduction; its condensing strategy reduces PnP from $\mathbf{s}_i\in\mathbb{R}^2$3 ms to $\mathbf{s}_i\in\mathbb{R}^2$4 ms, summarized as nearly $\mathbf{s}_i\in\mathbb{R}^2$5, and reduces four-iteration runtime from $\mathbf{s}_i\in\mathbb{R}^2$6 ms to $\mathbf{s}_i\in\mathbb{R}^2$7 ms, a $\mathbf{s}_i\in\mathbb{R}^2$8 reduction, while improving average 7-Scenes accuracy from $\mathbf{s}_i\in\mathbb{R}^2$9 cm / $\Sigma = R S S^T R^T,$0 to $\Sigma = R S S^T R^T,$1 cm / $\Sigma = R S S^T R^T,$2 (Li et al., 18 May 2026).

In dense mapping and SLAM, GeoGS reports that a standard appearance-aware primitive uses 59 parameters while GeoGS uses 10, “remains only 17% of the parameters required by a standard 3DGS primitive,” and on DTU Scan55 uses 23k Gaussians and 1.2 MB versus 161k and 37.5 MB for 2DGS. Within full SLAM it reports Replica accuracy 2.09 cm, completeness 3.95 cm, and completeness ratio 83.67%, and ScanNet++ average Chamfer Distance 8.67 cm (Zhou et al., 8 Jul 2026). GS-SLAM, which keeps full appearance but uses explicit Gaussian geometry, reports average Replica ATE RMSE 0.50 cm and rendering speed 386.91 FPS, illustrating the speed advantage of differentiable splatting over neural implicit dense SLAM (Yan et al., 2023).

In task-specific structured prediction, GaussianMap reaches 70.5 mAP in the camera-only setting and 78.3 mAP in camera-LiDAR fusion on nuScenes validation, and 70.1 and 79.0 mAP on Argoverse 2 validation, showing that a Gaussian representation can also function as an adaptive latent for vectorized map construction rather than as a scene renderer (Lyu et al., 30 Jun 2026). AnchorSplat shows a similar shift in feed-forward reconstruction: with 32 source views / 4 novel views it uses 247,153 Gaussians versus 5,550,940 for AnySplat and reports PSNR 21.48 and AbsRel 0.066 (Zhang et al., 8 Apr 2026). In point-cloud upsampling, PU-Gaussian achieves CD 0.079 and HD 1.443 on PUGAN at 16×, using one anisotropic Gaussian per input point and direct 3D sampling followed by refinement (Khater et al., 24 Sep 2025).

6. Limitations, misconceptions, and unresolved questions

The literature makes clear that dense geometric Gaussian representation is not a universal substitute for all other scene models. A first limitation is task specificity. LiteLoc’s color-free map is explicitly unsuitable for settings that also require photorealistic novel-view synthesis; GaussianMap is “fundamentally 2D BEV-based, not a full 3D Gaussian scene model”; and Fast-2DGS is confined to 2D image representation and does not address multi-view consistency, depth ordering, or 3D opacity modeling (Li et al., 18 May 2026, Lyu et al., 30 Jun 2026, Wang et al., 14 Dec 2025). A second limitation is dependence on external or learned priors. GeoTexDensifier depends on ZoeDepth-derived depth and normal maps, GeoGS-SLAM depends heavily on VGGT and MegaLoc, Sparse2DGS depends on CLMVSNet, and SGS-SLAM explicitly lists dependence on depth and 2D semantic inputs (Jiang et al., 2024, Gao et al., 13 Jul 2026, Wu et al., 29 Apr 2025, Li et al., 2024).

Another recurring issue is that dense geometric Gaussian maps often do not explicitly represent free space. The autonomy paper states that the Gaussian map “does not explicitly encode occluded or free space,” forcing exploration and collision checking to operate through custom uncertainty and pointwise collision tests rather than an occupancy grid or ESDF (Ong et al., 17 May 2025). Scalability is also unresolved. Several papers propose compactness mechanisms—masking and geometry codebooks in compact 3DGS SLAM, fewer primitives through geometry-only design in GeoGS, anchor-aligned decoding in AnchorSplat—but memory growth and long-horizon maintenance remain active concerns (Deng et al., 2024, Zhou et al., 8 Jul 2026, Zhang et al., 8 Apr 2026).

A final misconception is that more Gaussians necessarily imply better geometry. Sparse2DGS shows the opposite under sparse views: adaptive density control worsens CD from 1.201 to 1.234, while selective Gaussian update improves it to 1.152 (Wu et al., 29 Apr 2025). GeoTexDensifier and GeoGaussian make a related point from the reconstruction side: where and how Gaussians are added can matter as much as how many are added (Jiang et al., 2024, Li et al., 2024). This suggests that the central open problem is not simply densification, but principled density control: allocating Gaussian support where geometry is reliable, conserving it where evidence is weak, and attaching only the channels that the downstream task actually needs.