Frequency-Guided 3D Gaussian Splatting for Challenging Low-Light View Synthesis
Zhaoyuan Mai, Bi Zeng, Boquan Zhang, Tianle Zeng, Jingxuan Lu, Jiarong Feng
AI summary
Problem
Existing 3D Gaussian Splatting methods degrade severely in low-light environments due to high sensor noise and exposure inconsistencies, producing unstable geometry and flickering artifacts that hinder robotic navigation and mapping.
Approach
The method processes illumination features in the frequency domain to decouple noise from structural details, applies an adaptive mask to filter unstable Gaussian primitives based on rendering statistics, and enforces multi-view frequency consistency to ensure global illumination coherence.
Key results
- State-of-the-art reconstruction quality on challenging low-light datasets
- 46.4% reduction in model storage via adaptive Gaussian pruning
- Effective suppression of sensor noise and exposure artifacts while preserving fine textures
- Maintains real-time rendering speeds suitable for onboard robotic deployment
Why it matters
Enables reliable, real-time 3D scene understanding for autonomous robots and mapping systems operating in dark or poorly lit environments.
Abstract
Robust 3D scene understanding is crucial for autonomous robots, but degrades sharply in low-light envi- ronments where sensor noise and illumination inconsistencies corrupt visual inputs. Even 3D Gaussian Splatting (3DGS), while efficient for real-time reconstruction, produces unsta- ble and artifact-prone results under such conditions, limit- ing its reliability for navigation and mapping. To address these challenges, we propose a 3DGS-based framework for reconstructing clear scenes under low-light conditions. Firstly, we employ a frequency-aware modulator that operates on spectral components to decouple and suppress sensor noise from structural signals, providing a clean input for reconstruction. To refine the 3D model and ensure its compactness for onboard deployment, we introduce an adaptive denoising mask guided by dynamically updated statistics of rendering contribution and stability, which filters transient artifacts caused by sensor noise. Finally, a multi-view frequency consistency constraint is enforced to ensure the global coherence of the reconstructed model’s appearance, which is critical for consistent mapping. Experiments on challenging low-light datasets demonstrate that our method achieves state-of-the-art reconstruction quality while significantly reducing model storage by approximately 46.4% and maintaining real-time rendering speeds.