---

title: >- [Paper Note] Manifold-Aware Spectral Compaction: A Graph Signal Processing Perspective on Online Gaussian Reduction for 3DGS SLAM description: >- [ECCV 2026][3D Vision][3DGS SLAM] Online Gaussian map reduction reformulated via Graph Signal Processing (MASC) using Symmetrized KL divergence, incremental Nyström spectral embedding, and Bures-Wasserstein barycenters, cutting 60% of primitives with 2.8x memory reduction and improved tracking. tags: - ECCV 2026 - 3D Vision - 3DGS SLAM - Model Compaction - Optimal Transport - Spectral Embedding date: 2026-09-19 content_hash: 1b2819105e4df219

Manifold-Aware Spectral Compaction: A Graph Signal Processing Perspective on Online Gaussian Reduction for 3DGS SLAM

Conference: ECCV 2026
Paper: ECCV Official
Code: https://media.eventhosts.cc/Conferences/ECCV2026/pdfs/4373.pdf
Area: 3D Vision
Keywords: 3D Gaussian Splatting, SLAM, Model Compaction, Optimal Transport, Spectral Embedding

TL;DR

This paper introduces Manifold-Aware Spectral Compaction (MASC), a principled online reduction framework for 3DGS SLAM that reformulates Gaussian reduction as spectral signal reconstruction on a Riemannian manifold via Symmetrized KL topology modeling, incremental Nyström embedding, and closed-form Bures-Wasserstein barycenters, achieving a 2.8× GPU memory reduction, 2.4× rendering speedup, and an ATE of 0.98 cm on TUM RGB-D while discarding ~60% of redundant primitives.

Background & Motivation

3D Gaussian Splatting (3DGS) has emerged as a cornerstone representation for dense visual SLAM, delivering real-time photorealistic rendering through differentiable anisotropic Gaussian primitives. However, integrating 3DGS into dynamic online SLAM pipelines triggers continuous densification whenever novel views reveal photometric residuals. Consequently, the total primitive count scales almost linearly with trajectory length, creating an unsustainable scalability bottleneck that leads to explosive GPU memory consumption and severe rendering latency on compute-constrained edge platforms.

Existing compaction strategies predominantly originate from offline post-processing of static scenes, relying on heuristic pruning criteria such as accumulated ray contribution, opacity thresholds, or Hessian sensitivities. When applied to streaming SLAM pipelines, these isolated pruning heuristics sever the continuous manifold topology: deleting primitives based on simplistic Euclidean proximity or hard thresholds induces structural holes and tears on thin surfaces (reflected in sharp LPIPS error spikes). Furthermore, executing full eigenspace decomposition incurs a prohibitive \(O(N^3)\) computational cost, and abrupt primitive pruning destabilizes gradient continuity, triggering severe camera tracking drift and map fragmentation.

This paper's core insight is that geometric redundancy in mature 3DGS SLAM maps is fundamentally a spectral redundancy phenomenon on a Riemannian manifold: densely packed neighboring Gaussians on smooth surfaces exhibit near-identical attributes, causing the graph Dirichlet energy to vanish as signal energy concentrates overwhelmingly in the low-frequency Laplacian subspace. Core idea: reformulate online Gaussian reduction as continuous signal reconstruction in the spectral domain, constructing a statistical topology graph via Symmetrized KL divergence, executing scalable incremental Nyström spectral embedding, and collapsing redundant low-frequency primitives into moment-preserving Super-Gaussians via closed-form Bures-Wasserstein barycenters.

Method

Overall Architecture

MASC functions as a modular, plug-and-play map maintenance component embedded seamlessly within 3DGS SLAM backbones. As RGB-(D) keyframes are ingested and dense Gaussians accumulate, MASC first constructs a statistical topology graph weighted by Symmetrized Kullback-Leibler (SKL) divergence and appearance similarity. Next, it employs an incremental Nyström strategy with farthest point sampling (FPS) landmarks to derive low-dimensional spectral coordinates \(\tilde{\mathbf{U}}\) at \(O(nm)\) complexity. Redundant primitives are grouped via spectral clustering and subsequently aggregated into compact Super-Gaussians (\(\mathcal{S}_k\)) using closed-form Bures-Wasserstein barycenter iterations and weighted attribute low-pass filtering. An adaptive objective incorporating volumetric regularization determines the optimal retention ratio \(\rho_k\).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Dense Streaming Gaussian Map<br/>means / covariances / colors / opacities"] --> B["Statistical Topology Graph Construction<br/>Symmetrized KL divergence + appearance kernel"]
    B --> C["Incremental Nyström Spectral Embedding<br/>FPS landmark sampling & O(nm) projection"]
    C --> D["Spectral Redundancy Clustering<br/>grouping low-frequency structural components"]
    D --> E["Moment-Preserving Super-Gaussian Aggregation<br/>closed-form Bures-Wasserstein barycenters"]
    E --> F["Volume-Regularized Optimization<br/>balancing spectral reconstruction & dilation"]
    F --> G["Compact High-Fidelity SLAM Map<br/>~60% primitive reduction"]

Key Designs

1. Statistical Topology Graph Construction: capturing manifold connectivity via statistical overlap

Conventional reduction approaches group primitives using Euclidean spatial distance or KD-trees, which frequently causes erroneous links across thin structural barriers where surfaces are spatially close but topologically disjoint. MASC interprets each Gaussian \(g_i = (\mu_i, \Sigma_i, \alpha_i, \mathbf{c}_i)\) as a continuous normal probability density function \(\mathcal{N}_i(\mu_i, \Sigma_i)\) in \(\mathbb{R}^3\), quantifying statistical interaction via Symmetrized Kullback-Leibler (SKL) divergence:

\[D_{SKL}(\mathcal{N}_i \parallel \mathcal{N}_j) = \frac{1}{2} \left[ \text{tr}(\Sigma_j^{-1}\Sigma_i) + \text{tr}(\Sigma_i^{-1}\Sigma_j) + (\mu_i - \mu_j)^\top (\Sigma_i^{-1} + \Sigma_j^{-1})(\mu_i - \mu_j) - 2d \right]\]

with ambient dimension \(d=3\). The graph edge weight \(W_{ij}\) is formulated via a truncated kernel combining geometric statistical divergence and spherical harmonics feature discrepancy:

\[W_{ij} = \exp\left( -\frac{D_{SKL}(\mathcal{N}_i \parallel \mathcal{N}_j)}{\sigma_{geo}^2} - \frac{\|\mathbf{c}_i - \mathbf{c}_j\|_2^2}{\sigma_{app}^2} \right), \quad j \in k_{nn}(i)\]

Enforcing sparsity by restricting edges strictly to the spatial \(k\)-nearest neighbors \(k_{nn}(i)\) ensures that the resulting graph models genuine manifold topology without spurious cross-surface connections.

2. Scalable Incremental Spectral Embedding: bypassing the \(O(N^3)\) online bottleneck

While the graph Dirichlet energy \(\mathcal{E}_{smooth}(\mathbf{F}) = \text{tr}(\mathbf{F}^\top \mathbf{L}_N \mathbf{F}) = \frac{1}{2} \sum_{i,j} W_{ij} \|f_i - f_j\|_2^2 \to 0\) demonstrates that redundant attributes concentrate within low-frequency Laplacian eigenvectors, standard eigendecomposition requires \(O(N^3)\) computation. MASC bypasses this bottleneck via block-wise Nyström approximation. Using Farthest Point Sampling (FPS), it selects a tiny landmark set \(A\) of size \(m \ll N\) (typically 1%–5% of \(N\)) to construct landmark submatrix \(\mathbf{W}_{AA}\) and cross-affinity block \(\mathbf{C} \in \mathbb{R}^{N \times m}\). Following the eigendecomposition \(\mathbf{W}_{AA} = \mathbf{U}_A \mathbf{\Lambda}_A \mathbf{U}_A^\top\), the full approximated spectral coordinates \(\tilde{\mathbf{U}}\) are obtained analytically:

\[\tilde{\mathbf{U}} = \mathbf{C} \mathbf{U}_A \mathbf{\Lambda}_A^{-1}\]

When new keyframes introduce fresh primitives into the map, only the corresponding rows in \(\mathbf{C}\) are evaluated, cutting computation to \(O(nm)\) and enabling real-time streaming spectral clustering in milliseconds.

3. Moment-Preserving Super-Gaussian Aggregation: preserving anisotropic volume via Wasserstein barycenters

Direct arithmetic averaging or linear interpolation of covariance matrices within a redundancy cluster \(\mathcal{C}_k\) collapses the anisotropic volume and orientation of the original surface Gaussians. To guarantee strict conservation of local radiance field moments, MASC computes the exact 2-Wasserstein barycenter on the Gaussian probability distributions. The optimal covariance matrix \(\mathbf{\Sigma}_{\mathcal{S}_k}\) minimizes the opacity-weighted squared Wasserstein distance:

\[\mathbf{\Sigma}_{\mathcal{S}_k} = \arg\min_{\mathbf{\Sigma}} \sum_{i \in \mathcal{C}_k} w_i \, d_W^2(\mathcal{N}(0, \mathbf{\Sigma}), \mathcal{N}(0, \mathbf{\Sigma}_i))\]

where normalized opacity weights are \(w_i = \alpha_i / \sum_{j \in \mathcal{C}_k} \alpha_j\). The global solution is computed via fixed-point iterations:

\[\mathbf{\Sigma}_{\mathcal{S}_k}^{(t+1)} = \left(\mathbf{\Sigma}_{\mathcal{S}_k}^{(t)}\right)^{-1/2} \left( \sum_{i \in \mathcal{C}_k} w_i \left( \left(\mathbf{\Sigma}_{\mathcal{S}_k}^{(t)}\right)^{1/2} \mathbf{\Sigma}_i \left(\mathbf{\Sigma}_{\mathcal{S}_k}^{(t)}\right)^{1/2} \right)^{1/2} \right)^2 \left(\mathbf{\Sigma}_{\mathcal{S}_k}^{(t)}\right)^{-1/2}\]

Spatial centroids and appearance vectors are aggregated using a weighted graph low-pass filter: \(\mu_{\mathcal{S}_k} = \sum_{i \in \mathcal{C}_k} w_i \mu_i\) and \(\mathbf{c}_{\mathcal{S}_k} = \sum_{i \in \mathcal{C}_k} w_i \mathbf{c}_i\). This dual formulation preserves the first- and second-order spatial moments of the local surface patch, preventing over-smoothing and floaters.

Loss & Training

The map retention ratio \(\rho_k = K / |\mathcal{P}|\) is adaptively guided through a joint objective balancing spectral embedding reconstruction fidelity against primitive geometric dilation:

\[\min_{\{\mathcal{C}_k\}_{k=1}^K} \sum_{k=1}^K \sum_{i \in \mathcal{C}_k} \|\mathbf{u}_i - \mathbf{z}_k\|_2^2 + \lambda \sum_{k=1}^K \text{Vol}(\mathbf{\Sigma}_{\mathcal{S}_k})\]

where \(\mathbf{u}_i\) denotes the \(i\)-th row of \(\tilde{\mathbf{U}}\), \(\mathbf{z}_k\) is the cluster centroid in spectral space, and \(\lambda\) balances fidelity against map sparsity. The regularizer \(\text{Vol}(\mathbf{\Sigma}_{\mathcal{S}_k}) \propto \sqrt{\det(\mathbf{\Sigma}_{\mathcal{S}_k})}\) explicitly penalizes the creation of excessively dilated Gaussians that could induce blurry artifacts during novel view synthesis.

Key Experimental Results

Main Results

MASC was rigorously evaluated on synthetic indoor scenes from Replica and real-world handheld sequences from TUM RGB-D under both Monocular and RGB-D modalities. Crucially, all MASC results were benchmarked under an aggressive ~60% primitive reduction constraint (\(\rho_k \approx 40\%\)).

Table 1: Quantitative results on Replica across Monocular and RGB-D modalities (RTX 4090 benchmark)

Modality	Method	Localization RMSE (cm) ↓	PSNR (dB) ↑	SSIM ↑	LPIPS ↓	R-FPS ↑	GPU Memory (GB) ↓
Monocular	Photo-SLAM	1.09	33.30	0.926	0.078	911.3	6.0
Monocular	MASC + Photo-SLAM (Ours)	0.35	34.15	0.935	0.065	1150.0	3.5
Monocular	DROID-Splat	0.27	39.47	1.000	0.030	145.0	6.0
Monocular	MASC + DROID-Splat (Ours)	0.27	39.20	1.000	0.031	350.0	2.3
RGB-D	Photo-SLAM	0.60	34.96	0.940	0.059	1084.0	5.0
RGB-D	MASC + Photo-SLAM (Ours)	0.30	36.39	0.950	0.048	1103.0	3.2
RGB-D	DROID-Splat	0.29	39.66	1.000	0.030	145.0	14.0
RGB-D	MASC + DROID-Splat (Ours)	0.29	39.16	1.000	0.037	350.0	5.0

Table 2: Performance across varying retention ratios on TUM RGB-D (from paper Table 3)

Backbone	Retention \(\rho_k\)	Map Size (k) ↓	Peak Memory (GB) ↓	Tracking FPS ↑	ATE (cm) ↓	PSNR (dB) ↑	SSIM ↑
Photo-SLAM	100% (Uncompressed)	~50.0	5.00	40	1.32	21.90	0.763
Photo-SLAM	80%	40.0	4.50	45	1.20	22.40	0.778
Photo-SLAM	60%	30.0	3.70	49	1.08	22.90	0.793
Photo-SLAM	40% (MASC)	20.0	3.50	55	0.98	23.43	0.807
DROID-Splat	100% (Uncompressed)	180.4	8.50	160	1.47	27.58	0.990
DROID-Splat	60%	108.2	4.50	295	1.43	27.54	0.990
DROID-Splat	40% (MASC)	70.3	3.00	450	1.42	27.52	0.990

Ablation Study

Table 3: Compaction strategies ablation on Replica (\(\rho_k = 40\%\), DROID-Splat backbone, paper Table 4)

Strategy	PSNR (dB) ↑	SSIM ↑	LPIPS ↓	Memory (GB) ↓	ATE (cm) ↓	Analysis
(a) Uncompressed (100%)	39.66	1.000	0.030	14.00	0.29	Full baseline
(b) Random Subsampling	28.15	0.812	0.185	5.00	1.42	Severely destroys manifold integrity
(c) Heuristic Pruning	35.20	0.945	0.085	5.00	0.45	Spikes perceptual error via holes
(d) Euclidean K-Means	36.85	0.968	0.072	5.00	0.40	Fails across statistical boundaries
(e) MASC (Ours)	39.16	1.000	0.037	5.00	0.33	Near-lossless fidelity & high localization

Table 4: Algorithmic component ablation on TUM RGB-D (\(\rho_k = 40\%\), paper Table 5)

Configuration	PSNR (dB) ↑	SSIM ↑	LPIPS ↓	R-FPS ↑	ATE (cm) ↓	Note
(1) Spatial-only (\(\sigma_{app} \to \infty\))	19.45	0.712	0.185	430	1.48	Lacks appearance consistency
(2) Sparse Landmarks (\(m = 100\))	18.20	0.685	0.210	460	1.55	Inadequate Nyström approximation
(3) W/O \(D_{SKL}\) (\(L_2\) only)	19.85	0.725	0.165	440	1.46	Ignores covariance statistical overlap
(4) Simple Mean Aggregation	20.32	0.730	0.158	450	1.45	Volume collapse and geometric bias
(5) Full MASC (\(m = 1000\))	20.96	0.741	0.150	450	1.42	Optimal spectral fidelity & tracking

Key Findings

• Compaction as geometric regularization: Discarding 60% of primitives actively improves camera tracking accuracy on dynamic TUM sequences (Photo-SLAM ATE drops from 1.32 cm to 0.98 cm). MASC regularizes noisy, unconstrained Gaussians and eliminates transient floaters caused by motion blur.
• Drastic memory and rendering gains: On DROID-Splat (RGB-D), MASC slashes peak GPU memory from 14.0 GB to 5.0 GB (a 64.3% reduction) and boosts rendering throughput from 145 FPS to 350 FPS (a 2.4× speedup), unlocking long-term operation on edge devices.
• Landmark saturation behavior: Reconstruction fidelity saturates cleanly at \(m = 1000\) landmarks, proving that a small landmark subset is sufficient to accurately approximate the global Laplace-Beltrami spectrum.

Highlights & Insights

• From discrete heuristics to continuous Graph Signal Processing: Reconceptualizing 3DGS map growth as spectral signal redundancy on a manifold provides a mathematically grounded foundation for why and where primitives can be safely reduced without surface tears.
• Optimal Transport moment preservation: Utilizing closed-form Bures-Wasserstein barycenters ensures that Super-Gaussians preserve both the orientation and volume of the underlying radiance field, preventing blur and hollow surfaces.
• Seamless plug-and-play adaptability: MASC operates independently of specific tracking backbones, enhancing both direct feature-matching pipelines (Photo-SLAM) and dense optical flow architectures (DROID-Splat).

Limitations & Future Work

• Curvature-agnostic landmark sampling: FPS sampling selects landmarks based solely on spatial distance; future improvements could incorporate curvature-aware sampling to allocate more landmarks to intricate geometric boundaries.
• Unbounded outdoor environments: The evaluation was conducted primarily on indoor datasets; adapting MASC to unbounded driving scenes requires multi-scale hierarchical graph decomposition.
• Absence of explicit dynamic motion tracking: While MASC filters out transient floaters, it does not explicitly track moving dynamic obstacles, which could be addressed by incorporating temporal velocity fields.

Related Work & Insights

• vs Heuristic pruning (e.g., LightGaussian, Pup 3D-GS): Prior pruning methods evaluate primitives independently using localized opacity or Hessian metrics, causing structure tearing and LPIPS degradation; MASC achieves topological consistency through spectral clustering and barycenter fusion.
• vs Offline reduction (e.g., Gaussian Herding, Scaffold-GS): Offline frameworks require fully converged static scenes and heavy iterations; MASC provides an \(O(nm)\) incremental Nyström formulation tailored for real-time streaming SLAM.

Rating

• Novelty: ⭐⭐⭐⭐⭐ [Pioneering integration of Graph Signal Processing and Bures-Wasserstein barycenters into online 3DGS SLAM]
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Extensive monocular/RGB-D benchmarks on Replica and TUM, complete ablation of components and retention ratios]
• Writing Quality: ⭐⭐⭐⭐⭐ [Rigorous mathematical derivations, clear motivation, and strong empirical substantiation]
• Value: ⭐⭐⭐⭐⭐ [Directly tackles the critical memory explosion bottleneck of 3DGS SLAM, enabling edge-device deployment]

Related Papers

• [ECCV 2026] GraphCPD: Coherent Point Drift for Point Cloud Registration via Graph Signal Processing
• [ICLR 2026] Signal Structure-Aware Gaussian Splatting for Large-Scale Scene Reconstruction
• [CVPR 2025] Spectral Informed Mamba for Robust Point Cloud Processing
• [ECCV 2026] VIGS-SLAM: Visual Inertial Gaussian Splatting SLAM
• [ECCV 2026] SLAM-Former: Putting SLAM into One Transformer