PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting

Abstract: Recent advances in 3D Gaussian Splatting (3DGS) have enabled impressive real-time photorealistic rendering. However, conventional training pipelines inherently assume full multi-view consistency among input images, which makes them sensitive to distractors that violate this assumption and cause visual artifacts. In this work, we revisit an underexplored aspect of 3DGS: its inherent ability to suppress inconsistent signals. Building on this insight, we propose PDF-GS (Progressive Distractor Filtering for Robust 3D Gaussian Splatting), a framework that amplifies this self-filtering property through a progressive multi-phase optimization. The progressive filtering phases gradually remove distractors by exploiting discrepancy cues, while the following reconstruction phase restores fine-grained, view-consistent details from the purified Gaussian representation. Through this iterative refinement, PDF-GS achieves robust, high-fidelity, and distractor-free reconstructions, consistently outperforming baselines across diverse datasets and challenging real-world conditions. Moreover, our approach is lightweight and easily adaptable to existing 3DGS frameworks, requiring no architectural changes or additional inference overhead, leading to a new state-of-the-art performance. The code is publicly available at https://github.com/kangrnin/PDF-GS.

• The paper presents a novel progressive filtering framework that incrementally removes view-inconsistent distractors to achieve robust 3D Gaussian Splatting reconstructions.
• It refines the Gaussian parameter space through multi-phase optimization using semantic discrepancy measures and structure-oriented loss to prevent color drift and geometric bias.
• Experimental results show improved PSNR, SSIM, and LPIPS on challenging datasets, effectively surpassing previous state-of-the-art methods.

PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting

Background and Motivation

3D Gaussian Splatting (3DGS) has achieved significant advancements in real-time photorealistic rendering, especially for novel view synthesis. Despite its efficiency and high-quality reconstruction capabilities, existing 3DGS pipelines are fundamentally limited by the assumption that all training images depict a static, view-consistent scene. This assumption rarely holds in uncontrolled real-world environments, where transient content such as pedestrians, vehicles, or dynamic shadows introduces multi-view inconsistencies. These distractors corrupt scene representation, causing blurred regions, artifacts, and unstable geometry. Prior approaches have attempted explicit distractor modeling via mask prediction and scene decomposition, but these methods often incur architectural modifications, inference overhead, or rely on pretrained predictors.

The paper introduces PDF-GS, a robust and lightweight framework for 3DGS that exploits an inherent but underutilized property: 3DGS’s tendency to suppress signals inconsistent across views. Capitalizing on this self-filtering phenomenon, PDF-GS reframes 3DGS as an active inconsistency filter, utilizing progressive multi-phase optimization to amplify this property and achieve state-of-the-art, distractor-free reconstructions.

Methodology

Progressive Filtering Phases

PDF-GS operates through multiple filtering phases, each designed to further cleanse the Gaussian parameter space of distractors. Each phase consists of:

1. Distractor Localization: Scene regions exhibiting discrepancies between the rendered view and the input image are flagged as distractors, with discrepancies quantified through patch-level semantic features (DINOv3), yielding robust, semantic-aware masking.
2. Masking and Optimization: Identified distractor regions are excluded from training supervision via binary masks. The Gaussian parameter set is re-initialized from Structure-from-Motion (SfM) points at each phase—preventing color drift and geometric bias accumulation. Optimization utilizes a structure-oriented loss (SSIM-focused), sidelining fine-grained appearance modeling in favor of robust geometric consistency.
3. Sparse Color Updates: To reinforce multi-view coherence and prevent color parameters from overfitting to transient, single-view appearances, color updates are only applied sparsely throughout training iterations.

As phases progress, the discrepancy maps and masks become increasingly refined, recursively isolating view-inconsistent distractor regions and yielding a purified Gaussian representation.

Reconstruction Phase

After the filtering phases, the reconstruction phase reintroduces the standard 3DGS objective for fine-grained detail recovery. Critically, the purified Gaussian set from the final filtering phase is used as initialization, ensuring stable, view-consistent geometry as the scaffold for appearance refinement. The masking from the last filtering phase persists, preventing previously excised distractor regions from influencing optimization.

Experimental Results

Quantitative Performance

PDF-GS delivers consistent improvements across PSNR, SSIM, and LPIPS metrics on both NeRF On-the-go and RobustNeRF datasets, surpassing previous state-of-the-art methods including SpotLessSplats, DeSplat, RobustSplat, and WildGaussians. The improvements are attributed to the progressive removal of distractors, rather than explicit mask prediction or scene decomposition.

For example, in the NeRF On-the-go dataset (high occlusion settings), PDF-GS achieves mean scores of PSNR=23.32, SSIM=0.82, LPIPS=0.15, outperforming RobustSplat (PSNR=22.87, SSIM=0.84, LPIPS=0.15) and DeSplat (PSNR=22.38, SSIM=0.83, LPIPS=0.16).

In the RobustNeRF benchmark, PDF-GS records mean scores of PSNR=29.65, SSIM=0.90, LPIPS=0.13, exhibiting improvements over alternative approaches.

Qualitative Evaluation

Qualitative comparisons demonstrate that PDF-GS effectively removes distractors while preserving scene detail and static content fidelity. Reconstruction artifacts induced by transient objects visible in baselines are noticeably reduced in the PDF-GS outputs.

Ablation Studies and Analysis

• Number of Filtering Phases: Increasing filtering phases yields incremental gains in reconstruction quality, saturating at three phases; a single pass is inadequate for reliable distractor suppression.
• Structure-Oriented Loss and Sparse Color Updates: Both components are essential; their removal degrades performance by either overfitting to local color inconsistencies or causing color drift.
• Threshold Scheduling: A progressively decreasing threshold for mask refinement outperforms static criteria, achieving better balance between detail preservation and artifact removal.
• Gaussian Re-initialization: Periodic re-initialization at each phase prevents error accumulation and drift, resulting in cleaner reconstructions.
• Masking Metrics: Discrepancy measures using DINOv3 features maximize performance, although the progressive filtering strategy itself remains robust across other metrics like SSIM and PSNR.

Integration and Efficiency

PDF-GS is architecturally non-intrusive—requiring no changes to the 3DGS pipeline and incurring no additional inference-time overhead. Training times remain comparable to other state-of-the-art baselines, as feature extraction is decoupled from the inner optimization loop.

The progressive filtering strategy is compatible with more sophisticated learned masking approaches. When integrated with mask predictors from prior works, further gains in reconstruction quality are observed.

Implications and Future Directions

PDF-GS’s design shifts the paradigm for robust 3D reconstruction away from explicit semantic distractor modeling towards an intrinsic, progressive signal-consistency filtering. This approach is agnostic to network architecture, enabling seamless adoption atop existing 3DGS frameworks. The results substantiate the utility of iterative, semantic-aware refinement for distilling static scene geometry and appearance from noisy, distractor-laden input images.

From a practical standpoint, PDF-GS facilitates high-fidelity reconstructions in unconstrained, real-world scenarios—be it AR/VR content generation, digital twins, or autonomous perception—without reliance on specialized mask predictors or auxiliary networks. The progressive filtering mechanism is extensible to incorporate more advanced feature representations or masking strategies.

Theoretically, the methodology prompts further investigation into optimization landscapes in multi-view geometry and the interaction between self-filtering phenomena and semantic supervision. Potential future directions include:

• Integrating more advanced vision foundation models for feature extraction and discrepancy computation.
• Extending the progressive filtering framework to dynamic scene representations, temporal coherence, and non-SfM initializations.
• Reduction in training cost via adaptive scheduling or phase-wise sample selection.

Conclusion

PDF-GS introduces a progressive, discrepancy-guided filtering framework for robust 3D Gaussian Splatting, leveraging the inherent multi-view consistency filtering property of 3DGS. Through successive filtering phases and a reconstruction phase, PDF-GS consistently surpasses prior methods in both quantitative and qualitative metrics, yielding distractor-free, high-fidelity reconstructions in challenging real-world conditions. The framework is lightweight, architecturally agnostic, and amenable to integration with advanced masking techniques, setting a new benchmark for robust 3D scene synthesis (2604.12580).