ELoG-GS: Dual-Branch Gaussian Splatting with Luminance-Guided Enhancement for Extreme Low-light 3D Reconstruction

Abstract: This paper presents our approach to the NTIRE 2026 3D Restoration and Reconstruction Challenge (Track 1), which focuses on reconstructing high-quality 3D representations from degraded multi-view inputs. The challenge involves recovering geometrically consistent and photorealistic 3D scenes in extreme low-light environments. To address this task, we propose Extreme Low-light Optimized Gaussian Splatting (ELoG-GS), a robust low-light 3D reconstruction pipeline that integrates learning-based point cloud initialization and luminance-guided color enhancement for stable and photorealistic Gaussian Splatting. Our method incorporates both geometry-aware initialization and photometric adaptation strategies to improve reconstruction fidelity under challenging conditions. Extensive experiments on the NTIRE Track 1 benchmark demonstrate that our approach significantly improves reconstruction quality over the baselines, achieving superior visual fidelity and geometric consistency. The proposed method provides a practical solution for robust 3D reconstruction in real-world degraded scenarios. In the final testing phase, our method achieved a PSNR of 18.6626 and an SSIM of 0.6855 on the official platform leaderboard. Code is available at https://github.com/lyh120/FSGS_EAPGS.

• The paper introduces a dual-branch Gaussian splatting framework that decouples restoration and reconstruction to counteract geometric collapse in extreme low-light conditions.
• It integrates a learning-based depth initialization with adaptive luminance-guided post-processing to enhance photometric fidelity and suppress sensor noise.
• Experimental results on the NTIRE 2026 challenge show notable improvements in PSNR and SSIM, confirming its advantage over conventional 3D reconstruction methods.

Extreme Low-light 3D Reconstruction via Dual-Branch Gaussian Splatting and Luminance-Guided Enhancement

Introduction

The paper "ELoG-GS: Dual-Branch Gaussian Splatting with Luminance-Guided Enhancement for Extreme Low-light 3D Reconstruction" (2604.12592) addresses the computational challenges of recovering accurate 3D scene geometry and photorealistic visual fidelity from multi-view images obtained under extreme low-light and sparse-view (≈30 images) conditions. Standard Structure-from-Motion (SfM) pipelines and vanilla 3D Gaussian Splatting (3DGS) are susceptible to geometric collapse and color drift in such scenarios, due to sensor noise and limited feature visibility. The proposed ELoG-GS pipeline integrates robust, learning-based geometric initialization, sophisticated photometric restoration, and dual-branch splatting architectures, followed by adaptive luminance-guided post-processing to achieve high-fidelity 3D reconstruction.

Technical Contributions

Restoration-then-Reconstruction Decoupling

ELoG-GS utilizes an explicit restoration-then-reconstruction approach. In preprocessing, the Retinexformer network, pre-trained on the LOL_v2_real dataset, is applied in a zero-shot fashion to denoise and recover latent information from severely underexposed multi-view images. This preprocessing step significantly mitigates sensor noise and exposure-related degradation, providing high-quality inputs for the subsequent stages.

Robust Point Cloud Initialization

Classical SfM, particularly optimization-based methods like COLMAP, consistently fail in extreme low-light due to unreliable feature matching, resulting in excessively sparse or erroneous point clouds. To circumvent this, ELoG-GS employs a VGGT-based monocular depth estimation pipeline. Depth maps are back-projected and fused via spatial voxelization, leading to a dense and multi-view consistent point cloud that is formatted for direct use in downstream 3D Gaussian Splatting. This learning-based strategy effectively suppresses noise and guarantees geometric reliability in initialization.

Hybrid Dual-branch Reconstruction

The core of ELoG-GS comprises a dual-branch pipeline:

• Branch 1 (Regularized FSGS): Modified few-shot Gaussian Splatting is initialized with randomly distributed Gaussians, eschewing dependence on unreliable SfM. Regularization via tuned photometric losses and position learning rates ensures stable convergence and global geometric integrity, especially in background regions.
• Branch 2 (Geometry-guided EAP-GS): Transforming camera extrinsics/intrinsics into a COLMAP-compatible format, this branch augments the VGGT-derived point cloud with monocular ZoeDepth priors for high-resolution geometric guidance. The EAP-GS protocol excels at preserving high-frequency structures and sharp foreground textures.

After parallel training, both branches are evaluated, and an expert-in-the-loop approach selects the better result on a per-scene basis, balancing global regularity with local detail preservation.

Luminance-Guided Post-Enhancement

Following reconstruction, luminance-guided post-processing is performed to further refine renderings:

• Brightness adjustment and gamma correction compensate for underexposed image regions.
• Contrast enhancement is achieved via mean-based scaling.
• Luminance-based saturation amplification enhances chromaticity without inducing hue shifts or highlight artifacts.
• Global histogram matching ensures that the rendered color distributions conform to physically meaningful scene statistics.

This post-processing is essential to eliminate color drift and to achieve visually plausible, photorealistic renderings.

Experimental Results

ELoG-GS was evaluated on the NTIRE 2026 3D Restoration and Reconstruction Challenge dataset, which provides real-world, highly sparse, and degraded imagery. Quantitative results substantiate the efficacy of the proposed method:

• ELoG-GS achieved a PSNR of 18.66 and an SSIM of 0.686, surpassing both LITA-GS (15.63 PSNR, 0.542 SSIM) and Luminance-GS (10.89 PSNR, 0.531 SSIM).
• The method was ranked 9th out of 148 participants, demonstrating strong competitive performance.
• Qualitative comparisons confirm superior geometric consistency and visual fidelity, with effective mitigation of common artifacts such as color drift and geometric collapse, which afflict traditional approaches under these challenging conditions.

Implications and Future Directions

The methodological decoupling of restoration and reconstruction, coupled with learning-based depth fusion and dual-branch Gaussian splatting, delineates a robust paradigm for low-light 3D reconstruction. ELoG-GS’s strong numerical results indicate that explicit, learned priors effectively overcome the limitations of traditional feature-based geometry extraction in degraded visual domains.

Practical applications include robust scene digitization in adverse environments, augmented/virtual reality under unconstrained lighting, and low-light robotics perception. The pipeline’s modular design also facilitates future integration with end-to-end differentiable optimization, fully automatic branch selection, and advanced primitive representations (e.g., Gabor splats or neural point splats) to enhance fidelity and generalizability.

Further research is warranted toward automated selection criteria for the branch architecture, joint optimization of photometric restoration and 3D reconstruction, and scalable application to diverse domains such as dynamic scenes and semantic-aware reconstruction.

Conclusion

ELoG-GS establishes a comprehensive framework for high-fidelity 3D reconstruction from extreme low-light and few-shot multi-view imagery by integrating learning-driven depth initialization, dual-branch adaptive Gaussian splatting, and luminance-guided photometric enhancement. Experimental results on a challenging industry-standard benchmark validate its superiority over existing state-of-the-art solutions in both geometric and visual dimensions, with practical implications for real-world 3D scene capture under adverse illumination. Future development should focus on full pipeline automation and extension to broader reconstruction tasks.