Skip to content

Outdoor Monocular SLAM with Global Scale-Consistent 3D Gaussian Pointmaps

Conference: ICCV 2025 arXiv: 2507.03737
Code: Project Page
Area: 3D Vision Keywords: 3DGS SLAM, Monocular Vision, Scale Consistency, Pointmap, Outdoor Scenes

TL;DR

This paper proposes S3PO-GS, a monocular RGB-only outdoor SLAM system that anchors pose estimation to 3DGS-rendered pointmaps for scale self-consistency, and employs a patch-based dynamic mapping mechanism, achieving high-accuracy localization without cumulative scale drift and high-fidelity novel view synthesis.

Background & Motivation

3D Gaussian Splatting (3DGS) has become a popular choice for SLAM due to its high-fidelity real-time rendering, but faces two core challenges in RGB-only outdoor scenes:

Lack of geometric priors: Tracking methods based on differentiable rendering pipelines (e.g., MonoGS) tend to fall into local optima in complex outdoor environments and suffer from convergence difficulties.

Scale drift: Introducing independent tracking modules (e.g., Photo-SLAM, OpenGS-SLAM) supplements geometric constraints but requires maintaining scale alignment between the external module and the 3DGS map. Under large rotations and translations, cumulative errors lead to severe scale drift.

Core Insight: Pre-trained pointmap models (e.g., MASt3R) can provide geometric priors, but their scale is inconsistent with the 3DGS scene. The authors propose not involving the pre-trained model in pose estimation itself, but using it solely to establish pixel-level 2Dโ€“3D correspondences โ€” with 3D coordinates derived entirely from 3DGS-rendered pointmaps โ€” fundamentally eliminating the scale alignment problem.

Method

Overall Architecture

S3PO-GS comprises three core modules: 3DGS scene representation, self-consistent 3DGS Pointmap Tracking, and Patch-based Dynamic Mapping. The system initializes the 3DGS map using MASt3R (optimized for 1000 steps), then performs tracking and mapping for each new frame.

Key Designs

  1. Pointmap-Anchored Pose Estimation (PAPE):

    • A 3DGS depth map \(D_{ak}\) is rendered from the viewpoint of an adjacent keyframe; the rendered pointmap \(X_{ak}^r\) is obtained by back-projecting through camera intrinsics.
    • A pre-trained model (MASt3R/DUSt3R) generates pre-trained pointmaps \(X_{ak}^p, X_n^p\) for the current frame and keyframe; inter-frame pixel correspondences are established via nearest-neighbor matching.
    • The correspondence chain \(X_{ak}^r \leftrightarrow I_{ak} \leftrightarrow I_n\) is propagated to yield 2Dโ€“3D correspondences.
    • RANSAC + PnP is applied to solve the relative pose \(\mathbf{T}_n^{rel}\).
    • Key advantage: 3D coordinates originate from 3DGS rendering and are inherently scale-consistent with the scene; the pre-trained model serves only as a "bridge" without participating in pose computation.

  2. Pose Refinement:

    • Starting from the PnP-initialized pose, the photometric loss \(L_{pho} = \|I(\mathcal{G}, T) - \bar{I}\|_1\) is minimized through the 3DGS differentiable rendering pipeline.
    • The pose \(T \in SE(3)\) is linearized in the Lie algebra \(\mathfrak{se}(3)\), with gradients computed explicitly within the CUDA pipeline.
    • Only 5 iterations are required to achieve accuracy comparable to 100 iterations in conventional methods.

  3. Patch-Level Scale Alignment:

    • The rendered pointmap \(X^r\) and pre-trained pointmap \(X^p\) are divided into \(P \times P\) patches.
    • Patches with similar statistical distributions (mean/standard deviation differences below thresholds) are selected for normalization.
    • A set of "correct points" \(CP\) is identified in the normalized space, and the scaling factor is computed as: \(\sigma' = \frac{\mu(X^r[CP])}{\mu(X^p[CP])}\)
    • Iterative alignment continues until the scaling factor stabilizes, yielding \(\hat{X}^p = \sigma \times X^p\).
    • If insufficient correct points are found, a remedial scaling factor is computed using the aligned pointmap of an adjacent keyframe.

  4. Pointmap Replacement Mechanism:

    • When inserting new Gaussians at keyframes, the aligned pre-trained pointmap \(\hat{X}^p\) is used to detect "erroneous points" in the rendered pointmap \(X^r\).
    • Points with deviation exceeding the threshold are replaced: \(\hat{X}^r(x) = \hat{X}^p(x)\) if \(|X^r(x) - \hat{X}^p(x)| > \epsilon_m \times \hat{X}^p(x)\)
    • Random sparse downsampling controls the number of Gaussians.

Loss & Training

The total loss jointly optimizes poses and the Gaussian map over a keyframe window with three terms:

\[\min_{T_k, \mathcal{G}} \sum_{k \in \mathcal{W}} \alpha L_{pho}^k + (1-\alpha) L_{geo}^k + \lambda_{iso} L_{iso}\]
  • Photometric loss \(L_{pho}\): L1 rendering reconstruction loss
  • Geometric loss \(L_{geo} = \|X^r - \hat{X}^p\|_1\): pointmap alignment supervision
  • Isotropy regularization \(L_{iso}\): prevents excessive elongation of Gaussian ellipsoids

Key Experimental Results

Main Results

Comprehensive evaluation on three outdoor datasets โ€” Waymo, KITTI, and DL3DV:

Dataset Metric S3PO-GS (Ours) OpenGS-SLAM MonoGS GlORIE-SLAM
Waymo ATEโ†“ 0.622 0.839 8.529 0.589
Waymo PSNRโ†‘ 26.73 23.99 21.80 18.83
KITTI ATEโ†“ 1.048 3.224 9.493 1.134
KITTI PSNRโ†‘ 20.03 15.61 14.78 15.49
DL3DV ATEโ†“ 0.032 0.141 0.274 0.492
DL3DV PSNRโ†‘ 29.97 24.75 24.99 16.20

NVS PSNR improvements across all datasets are +2.73/+4.42/+4.98; tracking achieves best performance on KITTI and DL3DV, and is comparable to GlORIE-SLAM on Waymo.

Ablation Study

Configuration ATEโ†“ PSNRโ†‘ Note
w/o pose refinement 1.79 24.45 Tracking degrades without refinement
w/o scale alignment 3.50 23.49 Unaligned pointmaps introduce erroneous supervision
w/o point replacement 1.35 25.59 Incorrect Gaussian insertion degrades map quality
w/o \(L_{geo}\) 3.73 25.70 Lack of geometric supervision reduces translation awareness
Full model 0.62 26.73 โ€”

Convergence iteration ablation (Waymo 405841 scene, ATE):

Iterations MonoGS OpenGS-SLAM S3PO-GS
5 12.6 4.17 0.55
50 2.98 0.85 0.49
100 1.70 0.80 0.46

Key Findings

  • The PAPE module enables pose estimation to converge in only 5 iterations (MonoGS requires 50+).
  • Directly incorporating pre-trained pointmaps into MonoGS without the proposed processing leads to severe geometric blurring due to scale drift.
  • Patch-level local alignment is more robust than global alignment, avoiding contamination from outliers.

Highlights & Insights

  • Elegant design: The pre-trained model is used solely as a "matchmaker" (establishing correspondences) without directly participating in pose computation, eliminating scale alignment issues at the architectural level.
  • Plug-and-play: The approach generalizes to other 3DGS SLAM systems that need to integrate external geometric priors.
  • Rapid convergence: Stable accuracy is achieved in just 5 iterations, making it highly compatible with real-time systems.

Limitations & Future Work

  • Loop closure detection is not incorporated; drift may still accumulate over long sequences.
  • Dependence on the inference speed of pre-trained pointmap models may constrain real-time applicability.
  • Evaluation is limited to static scenes; dynamic objects are not addressed.
  • MASt3R/DUSt3R: Provide powerful pre-trained pointmap priors, but direct application to SLAM introduces scale inconsistency.
  • MonoGS: A pioneering 3DGS SLAM system, but suffers from convergence difficulties in outdoor scenes.
  • GlORIE-SLAM: Achieves high tracking accuracy via inter-frame relationships, but does not support NVS.

Rating

  • Novelty: โญโญโญโญ (Scale self-consistent design is conceptually novel)
  • Technical Depth: โญโญโญโญ (Complete tracking-mapping pipeline)
  • Experimental Thoroughness: โญโญโญโญ (Three datasets + thorough ablations)
  • Practical Value: โญโญโญโญ (Directly targets outdoor scenarios such as autonomous driving)