Gaussian Splatting: Anisotropic Visibility Field

• The paper introduces GAVIS, which augments standard 3D Gaussian Splatting with an anisotropic, direction-dependent visibility field for improved uncertainty estimation.
• It models per-Gaussian visibility using spherical harmonics, enabling constant-time queries and integrating this metric into a Bayesian, uncertainty-aware rendering pipeline.
• The framework achieves real-time performance at around 200 FPS and effectively guides active mapping by steering views towards underexplored regions.

Searching arXiv for the specified paper and closely related Gaussian Splatting visibility work. Searching arXiv for "Gaussian Splatting Anisotropic Visibility Field". Gaussian Splatting Anisotropic Visibility Field (GAVIS) is a framework that augments 3D Gaussian Splatting with an explicit, direction-dependent visibility field for uncertainty quantification and active mapping. It is motivated by the observation that regions unseen from the training views yield unreliable predictions from 3DGS, and it addresses this by defining the visibility field as the anisotropic visibility of each particle with respect to the training views, representing that field with spherical harmonics, and integrating it into a Bayesian Network-based uncertainty-aware 3DGS rasterizer. The resulting system performs real-time uncertainty quantification for synthesized views at 200 FPS and uses the resulting entropy estimates in a maximum information gain formulation for active mapping (Xue et al., 28 May 2026). In architectural terms, GAVIS is a post-hoc extension of the real-time, visibility-aware, anisotropic Gaussian rendering pipeline introduced by 3D Gaussian Splatting (Kerbl et al., 2023).

1. Problem formulation and conceptual scope

3D Gaussian Splatting represents a scene as a set of trainable 3D Gaussian particles with position, covariance, opacity, and view-dependent color parameterized via spherical harmonics. Rendering projects each Gaussian into the image, computes per-pixel opacity contributions and view-dependent color, and composites front-to-back along each ray. This representation provides high-quality, real-time novel-view synthesis, but active mapping requires not only rendering quality but also a view-conditioned estimate of predictive uncertainty (Kerbl et al., 2023).

The active mapping objective used by GAVIS is the standard maximum-entropy approximation to information gain,

$$\tau^* = \argmax_{\tau} \; \mathcal{H}\!\left(\mathbf{Z}_\tau\right),$$

where $\tau$ is a candidate pose and $\mathbf{Z}_\tau$ is the random image that would be observed there (Xue et al., 28 May 2026). The central requirement is therefore a per-pixel, per-pose uncertainty estimate that is fast enough to evaluate over many candidate views.

GAVIS is positioned against two families of prior approaches. FisherRF and VIMC focus on epistemic parameter uncertainty in 3DGS, but the motivating claim is that for active mapping there is a more direct issue: if a region has never been seen in any training view, its predictions should be considered unreliable regardless of the learned model’s apparent confidence. NVF had already emphasized the role of visibility in NeRF, but it learned an isotropic visibility field $V(\mathbf{x})$ with an additional neural network and ignored view direction. GAVIS therefore targets four specific goals: modeling per-Gaussian, direction-dependent visibility; computing this field analytically and efficiently with no extra network; integrating visibility into a Bayesian interpretation of 3DGS rasterization; and using the resulting entropy for active mapping (Xue et al., 28 May 2026).

A common misconception is to treat visibility as an isotropic occupancy-like scalar. GAVIS explicitly rejects that view. Seeing one side of a wall does not imply visibility of the opposite side, so the relevant latent quantity is directional visibility rather than position-only coverage (Xue et al., 28 May 2026).

2. Anisotropic visibility field

For each 3DGS particle $i$, GAVIS defines an anisotropic visibility field

$V^{(i)}(\mathbf{d}) \in [0,1],$

where $\mathbf{d}\in\mathbb{S}^2$ is a viewing direction. Intuitively, this is the probability that Gaussian $i$ is visible, meaning unoccluded and inside the field of view, from direction $\mathbf{d}$ given all training views (Xue et al., 28 May 2026).

The construction begins with visibility relative to a single training view with pose $\mathbf{p}\in\mathcal{P}$:

$\tau$0

where $\tau$1 indicates whether Gaussian $\tau$2 lies in the camera field of view, $\tau$3 is the accumulated transmittance from the camera center to the Gaussian along the camera ray, and

$\tau$4

is the direction from the training camera to the particle. The directional term is defined as

$\tau$5

so that $\tau$6 and visibility decays smoothly with angular deviation. The paper reports a default $\tau$7 in experiments (Xue et al., 28 May 2026).

Visibility over all training views is then aggregated as

$\tau$8

This exact expression is direction-dependent and multi-view, but naïve evaluation would require iterating over all past views for each query direction (Xue et al., 28 May 2026).

To obtain constant-time querying, GAVIS introduces

$\tau$9

and represents it with spherical harmonics:

$\mathbf{Z}_\tau$0

The directional kernel itself has an SH expansion,

$\mathbf{Z}_\tau$1

with coefficients

$\mathbf{Z}_\tau$2

leading to the accumulated per-particle coefficients

$\mathbf{Z}_\tau$3

In practice, the paper states that $\mathbf{Z}_\tau$4 works well, giving $\mathbf{Z}_\tau$5 coefficients per Gaussian (Xue et al., 28 May 2026).

The exact product form for $\mathbf{Z}_\tau$6 is then approximated through the AM–GM inequality:

$\mathbf{Z}_\tau$7

Querying visibility therefore consists of evaluating the SH basis at $\mathbf{Z}_\tau$8, reconstructing $\mathbf{Z}_\tau$9, and applying the analytic bound. This makes query cost $V(\mathbf{x})$0 per Gaussian and independent of the number of training views (Xue et al., 28 May 2026).

3. Bayesian rasterization and uncertainty-aware rendering

Standard 3DGS rendering in GAVIS follows the usual front-to-back compositing rule,

$V(\mathbf{x})$1

where $V(\mathbf{x})$2 are transmittance-weighted alpha-compositing coefficients and $V(\mathbf{x})$3 is the view-dependent color of Gaussian $V(\mathbf{x})$4 (Xue et al., 28 May 2026). GAVIS retains this renderer but reinterprets it probabilistically.

The paper adopts the uncertainty-aware volume rendering viewpoint used in NVF and models per-Gaussian emitted color as a Gaussian random variable,

$V(\mathbf{x})$5

with the recursive Bayesian relation

$V(\mathbf{x})$6

which unrolls into a Gaussian mixture,

$V(\mathbf{x})$7

This formulation alone is insufficient in unseen regions, because learned color uncertainty may remain small even where training visibility is poor (Xue et al., 28 May 2026).

GAVIS therefore introduces a visibility variable $V(\mathbf{x})$8 and conditions the color model on visibility:

$V(\mathbf{x})$9

Here $i$0 is a large covariance representing an uninformative, high-uncertainty prior, and the visibility probability is

$i$1

Opacity is also corrected through the visibility-compensated form

$i$2

where $i$3 controls trust in opacities for invisible regions and $i$4 is a prior opacity. The paper reports $i$5 by default (Xue et al., 28 May 2026).

Using

$i$6

the final pixel-color distribution becomes

$i$7

The first term aggregates visible Gaussians, while the second collects invisible mass into a high-variance prior component. This directly couples uncertainty to directional visibility rather than to parameter uncertainty alone (Xue et al., 28 May 2026).

Per-pixel uncertainty is quantified by the entropy of this Gaussian mixture. Since exact entropy has no closed form, GAVIS uses Huber’s upper bound:

$i$8

with $i$9 for RGB. The paper states that this computation is integrated into a modified CUDA 3DGS rasterizer and yields entropy maps at about 200 FPS on an A40 GPU (Xue et al., 28 May 2026).

4. Active mapping pipeline and visibility-field construction

The planning rule is the maximum-information-gain approximation

$V^{(i)}(\mathbf{d}) \in [0,1],$0

where image-level entropy is approximated by a depth-corrected sum of per-pixel entropies,

$V^{(i)}(\mathbf{d}) \in [0,1],$1

This correction is inherited from NVF and compensates for spatial correlation between neighboring pixels (Xue et al., 28 May 2026).

The operational pipeline is iterative. First, 3DGS is trained or updated using the currently acquired images. Second, for every training view, a modified 3DGS rasterizer computes the single-view visibility terms $V^{(i)}(\mathbf{d}) \in [0,1],$2, and those quantities are accumulated into the SH coefficients $V^{(i)}(\mathbf{d}) \in [0,1],$3. Third, the system performs density control with virtual particles. Fourth, it samples candidate poses. Fifth, for each candidate it queries $V^{(i)}(\mathbf{d}) \in [0,1],$4 along rendered rays, computes per-pixel entropy with the uncertainty-aware rasterizer, and aggregates entropy across the image. Finally, it selects the pose with maximal entropy, acquires a new image there, and adds that observation to the training set (Xue et al., 28 May 2026).

The virtual-particle mechanism addresses a specific ambiguity in 3DGS density control: both genuinely empty space and underexplored regions may contain few or no Gaussians. GAVIS samples virtual particles uniformly in the scene volume, assigns them zero opacity, large scale, and identity rotation, and computes their multi-view isotropic visibility

$V^{(i)}(\mathbf{d}) \in [0,1],$5

Virtual particles with $V^{(i)}(\mathbf{d}) \in [0,1],$6 are pruned as well-seen free space, while low-visibility particles are kept and assigned $V^{(i)}(\mathbf{d}) \in [0,1],$7 for all directions. The paper reports that this adds only about $V^{(i)}(\mathbf{d}) \in [0,1],$8–$V^{(i)}(\mathbf{d}) \in [0,1],$9 more particles while allowing entropy to distinguish underexplored regions from already-confirmed empty space (Xue et al., 28 May 2026).

A second misconception is that GAVIS modifies the core 3DGS optimization objective. It does not. The method is explicitly post-hoc: it leaves the 3DGS training pipeline unchanged and builds the anisotropic visibility field analytically from the trained model and the training views (Xue et al., 28 May 2026).

5. Empirical performance, datasets, and quantitative results

GAVIS is evaluated on four datasets: NeRF Synthetic, Space, Gibson, and HM3D. The reported planning horizons are $\mathbf{d}\in\mathbb{S}^2$0 steps for NeRF Synthetic and Space, $\mathbf{d}\in\mathbb{S}^2$1 for Gibson, and $\mathbf{d}\in\mathbb{S}^2$2 for HM3D. Baselines use the same candidate pose sampler and differ only in the uncertainty module: FisherRF, VIMC, NVF, GAVIS, FisherRF+GAVIS, and VIMC+GAVIS (Xue et al., 28 May 2026).

The reported metrics include PSNR, SSIM, LPIPS, completion ratio (CR), scene visual coverage (VIS), uncertainty preparation time $\mathbf{d}\in\mathbb{S}^2$3, uncertainty quantification speed in FPS, and uncertainty quality metrics AUSE-D and AUSE-V. The paper states that GAVIS achieves the best AUSE-V among all methods, indicating that high-entropy pixels align well with low ground-truth visibility (Xue et al., 28 May 2026).

On NeRF Synthetic, GAVIS reports PSNR $\mathbf{d}\in\mathbb{S}^2$4, SSIM $\mathbf{d}\in\mathbb{S}^2$5, LPIPS $\mathbf{d}\in\mathbb{S}^2$6, CR $\mathbf{d}\in\mathbb{S}^2$7, VIS $\mathbf{d}\in\mathbb{S}^2$8, UQ speed $\mathbf{d}\in\mathbb{S}^2$9 FPS, and $i$0. On Space, it reports PSNR $i$1, SSIM $i$2, LPIPS $i$3, CR $i$4, VIS $i$5, UQ speed $i$6 FPS, and $i$7. On Gibson, it reports PSNR $i$8, SSIM $i$9, LPIPS $\mathbf{d}$0, CR $\mathbf{d}$1, VIS $\mathbf{d}$2, UQ speed $\mathbf{d}$3 FPS, and $\mathbf{d}$4. On HM3D, it reports PSNR $\mathbf{d}$5, CR $\mathbf{d}$6, VIS $\mathbf{d}$7, UQ speed $\mathbf{d}$8 FPS, and $\mathbf{d}$9 (Xue et al., 28 May 2026).

The principal efficiency comparison is with NVF. NVF reports UQ speeds of $\mathbf{p}\in\mathcal{P}$0 FPS on NeRF Synthetic, $\mathbf{p}\in\mathcal{P}$1 FPS on Gibson, and $\mathbf{p}\in\mathcal{P}$2 FPS on HM3D, with uncertainty preparation times of roughly $\mathbf{p}\in\mathcal{P}$3, $\mathbf{p}\in\mathcal{P}$4, and $\mathbf{p}\in\mathcal{P}$5, respectively. GAVIS therefore retains comparable or better exploration performance while moving from seconds-to-minutes of uncertainty preparation to sub-second preparation and from single-digit or low-double-digit uncertainty rendering to roughly $\mathbf{p}\in\mathcal{P}$6–$\mathbf{p}\in\mathcal{P}$7 FPS (Xue et al., 28 May 2026).

The paper also emphasizes post-hoc compatibility with parameter-uncertainty estimators. FisherRF+GAVIS and VIMC+GAVIS reuse FisherRF or VIMC only to obtain per-Gaussian color covariances $\mathbf{p}\in\mathcal{P}$8, then inject visibility through the GAVIS rasterizer. A concrete example reported for HM3D is FisherRF+GAVIS improving FisherRF from PSNR $\mathbf{p}\in\mathcal{P}$9 to $\tau$00 and from CR $\tau$01 to $\tau$02. This suggests that directional visibility modeling and parameter uncertainty are complementary rather than exclusive (Xue et al., 28 May 2026).

Qualitatively, the paper reports that GAVIS steers viewpoints into unseen regions, including behind walls, doors, and around objects from multiple sides, whereas baseline methods often continue sampling already well-seen regions because they underestimate uncertainty there. The uncertainty maps shown in the paper reportedly track ground-truth visibility more closely than FisherRF, VIMC, or NVF (Xue et al., 28 May 2026).

6. Relation to adjacent Gaussian visibility research and methodological limitations

GAVIS sits within a broader line of work on Gaussian splatting, visibility, and directional effects. The original 3D Gaussian Splatting formulation already provided fully anisotropic 3D Gaussian primitives, projection to screen-space Gaussians, and visibility-aware front-to-back alpha compositing equivalent in image formation to NeRF-style volume rendering (Kerbl et al., 2023). GAVIS builds on that renderer but shifts the emphasis from photorealistic view synthesis to uncertainty and exploration.

Several neighboring methods clarify what GAVIS is and is not. Spec-Gaussian replaces spherical harmonics with an anisotropic spherical Gaussian appearance field to model specular and anisotropic view-dependent radiance, but its ASG field modulates color rather than explicit visibility or occlusion (Yang et al., 2024). VoD-3DGS introduces view-dependent opacity through a symmetric matrix and can be interpreted as a per-Gaussian anisotropic visibility or opacity field for reflections and highlights, but its purpose is improved rendering of view-dependent effects rather than uncertainty-aware mapping (Nowak et al., 29 Jan 2025). GVGS constructs a Gaussian visibility-aware multi-view geometric consistency constraint for surface reconstruction by aggregating Gaussian co-visibility across views, thereby using visibility for geometry supervision rather than planning (Su et al., 28 Jan 2026). REFINE derives a rendering-aware Hessian field and an anisotropic perceptual weight field for pruning, so its visibility treatment is tied to primitive importance rather than Bayesian predictive uncertainty (Chen et al., 8 Jun 2026).

These comparisons help delimit the specific meaning of “visibility field” in GAVIS. In GAVIS, visibility is neither merely an appearance parameter nor a geometry-consistency weight. It is an explicit directional memory of which Gaussians have been seen from which directions, queried during rendering to modulate both mixture weights and uncertainty (Xue et al., 28 May 2026).

The paper also states several limitations. GAVIS assumes static scenes and accurate camera poses and calibration. It does not model trajectory-level information gain, since pose selection is based on per-step entropy maximization. Its multi-view visibility is approximated with the lower bound

$\tau$03

which may slightly overestimate uncertainty because visibility is underestimated. Directional visibility is truncated to low-order SH, typically $\tau$04, so very sharp directional effects may not be captured perfectly. Finally, the distinction between underexplored space and free space depends on the density of virtual particles, introducing a trade-off between coverage fidelity and preprocessing cost (Xue et al., 28 May 2026).

Taken together, these properties define GAVIS as a visibility-centric extension of 3DGS: it turns the scene representation into a directionally indexed record of prior observation, embeds that record in an uncertainty-aware rasterizer, and uses the resulting entropy for real-time active mapping (Xue et al., 28 May 2026).