Lumina-4DGS: Illumination-Robust Four-Dimensional Gaussian Splatting for Dynamic Scene Reconstruction.

Area of Science:

• Computer Vision
• Computer Graphics
• 3D Reconstruction

Background:

• High-fidelity 4D reconstruction of dynamic scenes is crucial for immersive simulations but challenged by photometric inconsistencies in multi-view sensor arrays.
• Standard 3D Gaussian Splatting (3DGS) struggles with auto-exposure (AE), auto-white-balance (AWB), and ISP processing, leading to geometric errors like flickering and floating artifacts.

Purpose of the Study:

• To develop a robust framework, Lumina-4DGS, that harmonizes spatiotemporal geometry modeling with hierarchical exposure compensation for accurate 4D scene reconstruction.
• To address limitations of existing methods by explicitly decoupling and correcting photometric variations while preserving geometric integrity.

Main Methods:

• Implemented a hierarchical exposure compensation strategy with a Global Exposure Affine Module for AE/AWB fluctuations and a Multi-Scale Bilateral Grid for non-linearities.
• Introduced SSIM-Gated Optimization to dynamically gate gradient flow, prioritizing geometric accuracy over photometric overfitting by ensuring enhancement only on structurally reliable geometry.

Main Results:

• Achieved state-of-the-art Full Image PSNR of 31.12 dB on the Waymo Open Dataset, with minimal geometric errors (Depth RMSE: 1.89 m, Chamfer Distance: 0.215 m).
• Demonstrated a 2.13 dB PSNR improvement over baselines on a challenging self-collected dataset with severe illumination shifts, outperforming recent driving-scene methods.

Conclusions:

• Lumina-4DGS provides photorealistic, exposure-invariant novel view synthesis for dynamic scenes.
• The framework maintains superior geometric consistency across heterogeneous camera inputs, effectively mitigating artifacts caused by photometric variations.