PairDropGS: Paired Dropout-Induced Consistency Regularization for Sparse-View Gaussian Splatting

Technical Summary: PairDropGS

Problem Statement
Sparse-view 3D Gaussian Splatting (3DGS) faces significant challenges due to severe overfitting and insufficient geometric constraints caused by limited input observations. While recent dropout-based regularization methods attempt to alleviate overfitting by randomly suppressing Gaussian primitives during training, they exhibit critical limitations:

1. Instability: Different dropped Gaussian subsets often produce inconsistent structural responses, leading to unstable reconstruction and fluctuating rendering quality across training rounds.
2. Over-constraint on Ambiguity: Existing sophisticated dropout strategies often fail to distinguish between reliable structural cues and ambiguous high-frequency details, inadvertently constraining necessary flexibility or failing to suppress stochastic inconsistencies in local textures.
3. Rigid Scheduling: Most methods employ fixed dropout strategies throughout training, which may either over-constrain early optimization or under-constrain stochastic inconsistency in later stages, resulting in suboptimal convergence.

Methodology: PairDropGS
The authors propose PairDropGS, a Paired Dropout-Induced Consistency Regularization framework. Rather than designing increasingly complex dropout strategies, PairDropGS focuses on the consistency regularization of the optimization process itself. The framework consists of three core components:

1. Paired Dropout-Induced Optimization:
   Instead of rendering a single randomly dropped subset per iteration, PairDropGS constructs a pair of independently dropped Gaussian subsets ($\xi_a, \xi_b$) from the same shared Gaussian field under the same training view. Both subsets are supervised by the same ground-truth image. This formulation encourages the shared Gaussian field to learn stable structural responses that remain consistent despite different dropout masks.

2. Low-Frequency Consistency Regularization:
   To avoid over-constraining ambiguous high-frequency details while ensuring structural stability, the method applies a low-pass filter (Gaussian blur) to the rendered images of both branches. A consistency loss is then computed between these low-frequency representations. Crucially, an asymmetric formulation with a stop-gradient operator is used to stabilize optimization. This mechanism enforces consistency in scene layout and coarse geometry while preserving flexibility for high-frequency variations.

3. Progressive Consistency Scheduling:
   Recognizing that the impact of dropout-induced inconsistency varies during training, PairDropGS employs a progressive scheduling strategy. The consistency weight ($\lambda_t$) is linearly warmed up during the early stages of training to allow sufficient flexibility for establishing reasonable scene structures. As training progresses, the constraint is gradually strengthened to suppress accumulated inconsistency, overfitting, and structural drift.

Key Contributions

• Framework Redesign: The paper re-evaluates dropout-based sparse-view 3DGS from a consistency regularization perspective, proposing a simple yet effective framework that improves stability and robustness without relying on external priors or complex dropout strategies.
• Low-Frequency Mechanism: A novel regularization mechanism that explicitly aligns reliable structural responses in the low-frequency subspace, balancing scene stability with the preservation of high-frequency details.
• Adaptive Scheduling: A progressive consistency scheduling strategy that adaptively strengthens subset consistency constraints to optimize robustness throughout the training lifecycle.
• Plug-and-Play Nature: The method is designed to be easily integrated into existing dropout-based 3DGS pipelines without requiring architectural changes.

Experimental Results
Extensive experiments were conducted on three widely-used sparse-view benchmarks: LLFF, MipNeRF-360, and Blender.

• Quantitative Performance: PairDropGS consistently outperforms existing dropout-based methods (e.g., DropoutGS, DropGaussian, DropAnSH-GS) and other state-of-the-art sparse-view approaches. On the LLFF 3-view dataset, it achieved a PSNR of 20.89 dB, surpassing the previous best 3DGS-based method by 0.21 dB.
• Training Stability: Stability analysis across 10 random seeds demonstrates that PairDropGS significantly reduces performance fluctuations. For instance, on the LLFF dataset, the standard deviation of PSNR was reduced from 0.1724 (DropGaussian) to 0.0726, indicating a much more robust optimization process.
• Qualitative Improvements: Visual comparisons show that PairDropGS produces cleaner geometry, sharper boundaries, and fewer artifacts (such as blurred boundaries or local inconsistencies) compared to baselines.
• Compatibility: The method was successfully integrated into other 3DGS variants (FSGS, CoR-GS, DropGaussian), consistently improving their reconstruction quality with minimal computational overhead.
• Ablation Studies: Experiments confirmed that both the low-frequency consistency loss and the progressive scheduling strategy are essential for performance gains, and that a two-branch design offers the optimal trade-off between reconstruction quality and computational cost.

Significance
The paper claims that PairDropGS provides a new perspective for analyzing and enhancing dropout-based sparse-view Gaussian optimization. By shifting the focus from designing complex dropout policies to stabilizing the stochastic behavior of the optimization itself, the method achieves superior training stability and reconstruction quality. The authors position PairDropGS as a strong, extensible, and generalizable baseline for future research in sparse-view Gaussian splatting, emphasizing its simplicity and plug-and-play capability for improving existing dropout-based methods.