RoGER-SLAM: A Robust Gaussian Splatting SLAM System for Noisy and Low-light Environment Resilience

TL;DR

RoGER-SLAM tackles robust SLAM in challenging sensing environments by enhancing 3D Gaussian Splatting with three key components: Structure-Preserving Robust Fusion (SP-RoFusion) to fuse rendered appearance, depth, and edges; an adaptive, residual-balanced tracking objective; and a selectively activated CLIP-based enhancement module for severe degradations. The method also exploits the implicit low-pass property of Gaussian rendering to suppress high-frequency noise while preserving structure, enabling stable tracking and high-fidelity mapping. Extensive experiments on Replica, TUM, and real-world sequences show substantial gains in trajectory accuracy (e.g., up to 6.82 cm drift under compounded degradations reduced to ~0.60 cm) and image-quality metrics (PSNR, SSIM) compared with prior 3DGS-SLAM systems, validating robustness to noise and low-light. The results demonstrate practical potential for robust dense mapping in safety-critical robotics and pave the way for open-source release and future extensions to outdoor and multi-sensor setups.

Abstract

The reliability of Simultaneous Localization and Mapping (SLAM) is severely constrained in environments where visual inputs suffer from noise and low illumination. Although recent 3D Gaussian Splatting (3DGS) based SLAM frameworks achieve high-fidelity mapping under clean conditions, they remain vulnerable to compounded degradations that degrade mapping and tracking performance. A key observation underlying our work is that the original 3DGS rendering pipeline inherently behaves as an implicit low-pass filter, attenuating high-frequency noise but also risking over-smoothing. Building on this insight, we propose RoGER-SLAM, a robust 3DGS SLAM system tailored for noise and low-light resilience. The framework integrates three innovations: a Structure-Preserving Robust Fusion (SP-RoFusion) mechanism that couples rendered appearance, depth, and edge cues; an adaptive tracking objective with residual balancing regularization; and a Contrastive Language-Image Pretraining (CLIP)-based enhancement module, selectively activated under compounded degradations to restore semantic and structural fidelity. Comprehensive experiments on Replica, TUM, and real-world sequences show that RoGER-SLAM consistently improves trajectory accuracy and reconstruction quality compared with other 3DGS-SLAM systems, especially under adverse imaging conditions.

Figures (11)

• Figure 1: Qualitative and quantitative comparison under noise and low-light conditions. The first column shows the degraded input images, the second column presents the rendered reconstructions, and the third column reports the quantitative results. Compared with SplaTAM under noise inputs (top) and compound inputs (middle), our method (bottom) achieves significantly sharper reconstructions with clearer structural details and superior quantitative performance in terms of ATE, PSNR, SSIM, and LPIPS.
• Figure 2: Framework Overview. Given RGB-D inputs, Gaussian spheres are initialized and densified (Sec.III-A), and updated through structure-preserving robust fusion (Sec.III-B). Camera poses are estimated via an adaptive tracking objective (Sec.III-C). A degradation judgment module selectively activates the CLIP-based enhancement module (Sec.III-D). The pipeline produces photorealistic reconstructions and consistent trajectories.
• Figure 3: Structure-preserving robust fusion (SP-RoFusion) module. The rendered image $I_{\text{render}}$, input depth $D^{[3]}$, and structural edge map $G^{[3]}$ (where $[3]$ denotes channel replication to three dimensions) are linearly fused into $I_{\text{fuse}}$ with weights $\alpha$, $\beta$, and $\gamma$. Depth, color, and illumination residuals are then computed, and combined into the final mapping loss $\mathcal{L}_{\text{map}}$ with an adaptive illumination weight.
• Figure 4: Workflow of the CLIP Image Enhancement Module. The module addresses both noise reduction and low-light enhancement, and is selectively activated through a dual-condition judgment strategy based on global brightness and residual noise variance. A CLIP-based image encoder extracts multi-scale features from degraded inputs, which are reconstructed by the image decoder to suppress noise. The Neural Representation Normalization (NRN) further mitigates degradation artifacts, while the Light Enhancement Module (LEM) restores brightness and structural details. The enhancement process is constrained by CLIP-based supervision, aligning visual and text features via cosine similarity. Note that for visualization, the light-condition judgment is presented in reciprocal form for clarity, while in our implementation it is applied directly on the raw brightness values.
• Figure 5: Examples from the constructed natural-noise dataset. columns (a) and (b) show two representative scenes, where the top row (a1, b1) corresponds to clean inputs and the bottom row (a2, b2) contains noisy counterparts generated with shot and read noise. The residual variance $\sigma_R^2$ quantifies the noise level, demonstrating the increase from mild to stronger perturbations after noise synthesis.