Decoupled Gaussian Aggregator (DGA)

• DGA is a probabilistic Gaussian-to-voxel aggregator that decouples geometry (via opacity-based occupancy) from conditional semantic predictions in 3D semantic scene completion.
• It computes voxel occupancies using localized Gaussian kernels and integrates semantic logits separately, thereby mitigating outlier-induced artifacts seen in earlier methods.
• Empirical results on Occ-ScanNet show DGA’s efficiency with superior IoU/mIoU metrics and reduced latency compared to prior aggregators like GaussianFormer and PGS-based methods.

The Decoupled Gaussian Aggregator (DGA) is a probabilistic Gaussian-to-voxel aggregation module introduced as the final stage of SplatSSC, a monocular 3D Semantic Scene Completion (SSC) framework that represents a scene with sparse 3D Gaussian primitives and splats them into a dense voxel grid. In that setting, DGA is designed to be robust to outlier Gaussians by decoupling geometry from semantics during aggregation. A separate forecasting literature uses the closely related term “Gaussian aggregator” for the revealed conditional-probability aggregator in a Gaussian partial information model; in the supplied literature, that object is explicitly connected to DGA terminology, although the older paper itself does not use the name “Decoupled Gaussian Aggregator” (Qian et al., 4 Aug 2025, Ernst et al., 2016).

1. Definition and functional role

Within SplatSSC, DGA is the final aggregation module that converts an object-centric Gaussian representation into a dense voxel-based SSC output. The upstream pipeline begins with a single RGB image $\mathcal{I}_\mathrm{rgb}$, uses an image encoder built from EfficientNet and FPN to produce multi-scale image features $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$, and uses a frozen Depth-Anything-V2 model to produce depth features $\mathcal{F}_d$. A depth branch with Group-wise Multi-scale Fusion (GMF) and Group Cross-Attention (GCA) yields fused depth features $\mathcal{F}_d'$, from which a small depth head predicts a refined depth map $\mathcal{I}_d$. These features are lifted into 3D to generate an initial set of Gaussian primitives $\mathcal{G}_o$, which are then refined by multiple encoder blocks, as in EmbodiedOcc, into $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$. DGA consumes these refined Gaussians and outputs voxel-wise semantic probabilities $\hat{\mathcal{V}}_\mathrm{agg}$ over $C$ classes, including empty (Qian et al., 4 Aug 2025).

The motivation for DGA is explicitly tied to two failure modes in prior Gaussian-based SSC and occupancy methods such as GaussianFormer, GaussianFormer-2, and EmbodiedOcc. First, random primitive initialization produces redundant primitives in empty space and many outlier primitives that are misplaced or poorly supervised. Second, prior splatting and aggregation strategies are fragile: GaussianFormer uses an additive aggregator, while GaussianFormer-2 introduces Probabilistic Gaussian Superposition (PGS), a GMM-based formulation that entangles opacity and semantics. DGA is introduced to provide a principled probabilistic aggregation, to use opacity as geometry confidence rather than as a semantic prior, and to suppress outlier-induced “floater” artifacts.

A common misconception is that DGA is a learnable neural subnetwork analogous to an extra attention block or decoder head. The supplied formulation states the opposite: no extra MLPs, attention layers, or CNNs are introduced inside DGA; it is a purely probabilistic aggregator over Gaussian kernels and semantic probabilities.

2. Gaussian primitive parameterization and voxel aggregation

Each refined primitive

$G_i \in \mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$

has a mean $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$0, a scale $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$1, a rotation quaternion $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$2, an opacity $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$3, and semantic logits $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$4 for non-empty classes. Its covariance is

$$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$5

where $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$6 converts a quaternion to a rotation matrix. Let $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$7 denote the softmax-normalized semantic probabilities. DGA aggregates only over a local neighborhood $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$8 for each voxel center $$\mathcal{F}_\mathrm{rgb} = \{ f_\mathrm{rgb}^l \}_{l=1}^L$$9, rather than over all Gaussians globally (Qian et al., 4 Aug 2025).

The unnormalized geometric influence of a primitive at position $\mathcal{F}_d$0 is

$\mathcal{F}_d$1

Algorithmically, DGA proceeds per voxel by selecting $\mathcal{F}_d$2, computing Gaussian kernels and densities for neighboring primitives, aggregating geometry, aggregating conditional semantics, and then fusing occupied and empty probabilities into the final voxel distribution. The outputs form a full 3D grid of semantic probabilities used both for loss computation and evaluation.

This design makes DGA a deterministic probabilistic interface between a sparse continuous representation and a dense discrete one. A plausible implication is that the module’s importance is not merely mathematical elegance but representation transfer: it formalizes how a sparse, depth-aligned set of Gaussians becomes a dense SSC prediction without introducing an additional learned fusion block.

3. Decoupling geometry and semantics

The defining property of DGA is the separation of geometry and semantics. Geometry is represented by the occupancy probability $\mathcal{F}_d$3, while semantics is represented by a conditional distribution $\mathcal{F}_d$4 over non-empty classes. The revised occupancy model is

$\mathcal{F}_d$5

Here, $\mathcal{F}_d$6 is the geometric influence of primitive $\mathcal{F}_d$7 at $\mathcal{F}_d$8, and $\mathcal{F}_d$9 is its global existence confidence. Their product is the effective occupancy contribution. The product term is the probability that none of the local Gaussians occupies $\mathcal{F}_d'$0, so its complement is the probability that at least one does (Qian et al., 4 Aug 2025).

Semantics is then modeled conditionally on occupancy. For class $\mathcal{F}_d'$1,

$\mathcal{F}_d'$2

The key difference from PGS is that opacity $\mathcal{F}_d'$3 does not appear in this expression. Semantics depends only on geometric proximity via $\mathcal{F}_d'$4 and on the learned semantic probabilities $\mathcal{F}_d'$5. DGA therefore assigns opacity exclusively to the geometry channel and semantic probabilities exclusively to the semantic channel.

The final voxel probabilities are

$\mathcal{F}_d'$6

and

$\mathcal{F}_d'$7

This formulation is often summarized as “decoupling,” but the decoupling is not independence. Semantics is conditioned on occupancy, and the final marginal semantic probability is multiplicatively gated by occupancy. The supplied derivation therefore rules out a second misconception: DGA does not ignore geometry in semantic prediction; it isolates where geometry should act.

4. Outlier robustness and comparison with prior Gaussian aggregators

The immediate target of DGA is the outlier failure mode of PGS in GaussianFormer-2. The supplied analysis considers an outlier primitive $\mathcal{F}_d'$8 located near some voxel $\mathcal{F}_d'$9 but far from the true surface. In PGS, the posterior weight

$\mathcal{I}_d$0

can become dominant even when the opacity $\mathcal{I}_d$1 is small, because the denominator is controlled by the outlier’s likelihood once other primitives have negligible likelihood. The semantic expectation then collapses to the outlier’s semantic logits, and geometry can also become spuriously high at the same voxel. The result is a semantically labeled blob floating in empty space (Qian et al., 4 Aug 2025).

DGA prevents this specific normalization failure. If an outlier has low opacity $\mathcal{I}_d$2 but high local likelihood, its semantic weight inside $\mathcal{I}_d$3 can still be large, but the final class probabilities

$\mathcal{I}_d$4

remain small because $\mathcal{I}_d$5 is suppressed by the low occupancy contribution $\mathcal{I}_d$6. The effect is not that the outlier disappears from the conditional semantic mixture; rather, its semantic influence is prevented from activating a strongly occupied voxel.

The empirical ablation on Occ-ScanNet-mini shows the practical consequences. Without depth-guided GMF, GF.agg yields IoU 11.64 and mIoU 12.62, GF2.agg yields IoU 27.54 and mIoU 17.27, and DGA yields IoU 48.85 and mIoU 36.91. With GMF enabled, GF.agg yields IoU 16.63 and mIoU 10.45, GF2.agg yields IoU 57.70 and mIoU 45.13, and DGA yields IoU 60.61 and mIoU 48.01, corresponding to a gain of $\mathcal{I}_d$7 IoU and $\mathcal{I}_d$8 mIoU over GF2.agg. Figure 1 is described as showing that GF.agg and GF2.agg produce floaters, whereas DGA suppresses them. In the full SplatSSC system on Occ-ScanNet, the model achieves IoU 62.83 and mIoU 51.83, compared with RoboOcc at IoU 56.48 and mIoU 47.67, an improvement of $\mathcal{I}_d$9 IoU and $\mathcal{G}_o$0 mIoU; the abstract also states reductions in latency and memory consumption by more than $\mathcal{G}_o$1 (Qian et al., 4 Aug 2025).

5. Interaction with initialization, losses, and design choices

DGA is coupled operationally to the depth-guided initialization strategy. GMF fuses image and depth features to produce $\mathcal{G}_o$2 and a depth map $\mathcal{G}_o$3; a downsampled image-plane grid, for example $\mathcal{G}_o$4, is lifted using $\mathcal{G}_o$5 to generate a compact set of Gaussian centers, about 1200 Gaussians with initial scales in $\mathcal{G}_o$6 m. These primitives are refined before they reach DGA. The supplied interpretation is direct: DGA operates on a sparse, depth-aligned set of Gaussians instead of a huge random cloud, reducing both the number of overlapping Gaussians per voxel and the chance of strong outliers, while still requiring robustness to the outliers that remain (Qian et al., 4 Aug 2025).

The losses reinforce the same decomposition. Stage 2 uses

$\mathcal{G}_o$7

with

$\mathcal{G}_o$8

The Probability Scale Loss is

$\mathcal{G}_o$9

Earlier layers therefore receive weaker geometry supervision, while the last layer is fully weighted. The paper reports that combining geometry and semantic scale losses is suboptimal for SplatSSC and that using only $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$0 plus focal and Lovász yields the best IoU and competitive mIoU. This suggests that DGA’s decoupling is reflected not only in inference but also in supervision.

Several design choices bound the module’s practical regime. The best reported configuration uses 1200 primitives and a scale range of $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$1 meters, achieving IoU 61.47 and mIoU 48.87 on Occ-ScanNet-mini. Using 4800 or 19200 primitives yields marginal gains or degradation and higher memory and time, while large scale ranges can lead to overlaps and even OOM. The neighborhood $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$2 is not fully specified, and Appendix Table 10 reports that performance is batch-size sensitive, with mIoU rising from 36.09 at batch size 2 to 51.83 at batch size 8. The local-view constraint is also explicit: SplatSSC is currently trained per frame, and extension to global multi-frame scenes would require primitive pruning and fusion.

6. Relation to the Gaussian partial information framework

A distinct line of work studies a “Gaussian aggregator” for calibrated probability forecasts under a Gaussian partial information model. In that framework, there is an underlying probability space $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$3, a binary event $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$4, and forecasters with information $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$5-fields $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$6. Each forecast is calibrated, $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$7, and the theoretically optimal revealed aggregator is

$\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$8

The Gaussian model introduces a centered Gaussian process $\mathcal{G}_\mathrm{loc} = \{G_i\}_{i=1}^N$9 with covariance $\hat{\mathcal{V}}_\mathrm{agg}$0, defines the event as $\hat{\mathcal{V}}_\mathrm{agg}$1, and represents each forecaster’s signal through an observed subset $\hat{\mathcal{V}}_\mathrm{agg}$2 (Ernst et al., 2016).

For two forecasters in the symmetric one-shot setting, the information structure is

$\hat{\mathcal{V}}_\mathrm{agg}$3

with prior $\hat{\mathcal{V}}_\mathrm{agg}$4. If $\hat{\mathcal{V}}_\mathrm{agg}$5 and $\hat{\mathcal{V}}_\mathrm{agg}$6, then the fixed-parameter Gaussian aggregator is

$\hat{\mathcal{V}}_\mathrm{agg}$7

The Bayesian one-shot aggregator integrates over the posterior of $\hat{\mathcal{V}}_\mathrm{agg}$8 and is given in closed piecewise form by

$\hat{\mathcal{V}}_\mathrm{agg}$9

The connection between this forecasting aggregator and the SSC module called DGA is structural rather than application-level. In both cases, aggregation is derived from an explicitly Gaussian model rather than from ad hoc feature fusion. In the forecasting setting, “decoupling” refers to separating the structural Gaussian model from parameter estimation by placing a prior on the overlap parameter $C$0 and integrating it out. In SplatSSC, “decoupling” refers to separating opacity-driven occupancy from likelihood-driven semantics. These are different technical uses of the same broad aggregation idea, and conflating them is a recurrent source of confusion.

7. Position within related methods and conceptual significance

Relative to standard 3D Gaussian splatting for radiance fields, DGA operates in 3D voxel space rather than along image rays, and it introduces explicit probabilistic modeling of occupancy versus semantics. Relative to object-centric Gaussian SSC methods, GaussianFormer uses an additive splatter without principled semantic conditioning, while GaussianFormer-2 uses PGS, which treats opacity as a GMM prior and thereby entangles geometry and semantics. EmbodiedOcc, EmbodiedOcc++, and RoboOcc focus on online and incremental occupancy in indoor scenes but are described as relying on aggregators that do not fully decouple geometry and semantics. Relative to grid-aligned voxel pooling or transformer-based attention methods such as VoxFormer and OccFormer, DGA is object-centric and continuous: occupancy is a gated probabilistic OR over Gaussians, and semantics is a conditional GMM over Gaussian densities (Qian et al., 4 Aug 2025).

The significance of DGA in the SplatSSC formulation lies in how it resolves a specific representational tension. Sparse Gaussians are efficient and flexible, but sparsity increases sensitivity to initialization errors and local outliers. A purely additive or posterior-normalized aggregator can turn those outliers into occupied semantic artifacts. DGA addresses that failure mode by assigning different inferential roles to opacity and semantic class probabilities, then recombining them only at the final marginalization step. In the supplied evidence, this separation is not merely conceptual; it is tied to the reported state-of-the-art performance on Occ-ScanNet, to reduced latency and memory consumption, and to visibly cleaner reconstructions.

In that sense, DGA names a family resemblance more than a single universal formula. In monocular 3D SSC, it denotes a decoupled Gaussian-to-voxel occupancy-semantic aggregator. In the Gaussian partial information framework, it denotes a revealed probability aggregator derived from Gaussian signal structure and, in the one-shot Bayesian case, from prior integration over information overlap. What unifies these usages is the premise that aggregation should be specified by an explicit probabilistic model of how local evidence, confidence, and shared structure combine.