Introduction
3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction with high-quality, real-time rendering. However, it suffers from a "noisy neighbor" problem: when initialized with low-quality data, such as sparse point clouds from Structure-from-Motion, it often produces floating artifacts. These blurry, detached structures hover in the scene and significantly degrade visual fidelity.
The Problem: Rapid Collapse and Frequency Bias
The researchers identified a fundamental "frequency bias" in 3DGS training. Gaussians tend to collapse in size extremely rapidly during the first 5,000 iterations. If they shrink smaller than the Nyquist sampling interval before they have learned the basic low-frequency structural information of an object, they become "stuck" as under-optimized high-frequency noise, manifesting as floaters.
Methods: EFA-GS
The authors propose Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS), which centers on three key strategies to manage Gaussian optimization:
- Low-Frequency-Come-First (LFCF) Algorithm: Instead of allowing unchecked shrinkage, LFCF monitors cumulative gradients. If a Gaussian struggles to fit the scene, the algorithm selectively expands it, "resetting" it to a lower frequency state to relearn underlying structures.
- Depth-based Strategy: To prevent over-blurring distant backgrounds, EFA-GS assigns lower expansion factors to deeper Gaussians which naturally have lower sampling rates.
- Scale-based Strategy: The method employs volume-preserving expansion. When triggered, it expands the shortest axis while shrinking the longest, helping the primitive cover local geometry without ballooning out of control.
Key Results and Insights
EFA-GS was evaluated on standard benchmarks and a new RWLQ (Real-World Low-Quality) dataset. The results demonstrate significant improvements in both quantitative metrics and visual quality.
- Quantitative Gains: On the RWLQ dataset, EFA-GS improved PSNR by 1.68 dB over Mip-splatting and 0.99 dB over vanilla 3DGS.
- Efficiency: Surprisingly, EFA-GS often trains faster than vanilla 3DGS (e.g., 29.2 mins vs 32.3 mins) because the LFCF algorithm reduces the total number of unnecessary Gaussian splits and clones.
- Downstream Impact: When integrated into 3D editing pipelines like GaussianEditor, EFA-GS successfully prevented the "clown-like" artifacts that typically appear during text-to-3D modifications.
Conclusion
By identifying the link between Gaussian covariance and the Nyquist-Shannon sampling theorem, the authors have provided a theoretical and practical solution to the floater problem. EFA-GS ensures that low-frequency structural information is prioritized, leading to cleaner, more robust 3D reconstructions from imperfect real-world data.
Summary updated