DOC-GS: Dual-Domain Observation and Calibration for Reliable Sparse-View Gaussian Splatting
Proposes DOC-GS framework combining depth-guided Dropout and Dark Channel Prior for reliable sparse-view 3D Gaussian reconstruction.
Key Findings
Methodology
This work formulates sparse-view 3D Gaussian splatting as a dual-domain reliability inference problem, integrating optimization domain cues via Continuous Depth-Guided Dropout (CDGD) and observation domain evidence through Dark Channel Prior (DCP). CDGD dynamically adjusts the dropout probability based on scene depth, suppressing unreliable primitives and stabilizing training. Simultaneously, DCP detects haze-like artifacts in rendered images by analyzing pixel-wise dark channel statistics, identifying anomalous regions. Cross-view evidence is aggregated to perform geometric pruning, removing low-confidence Gaussians. The approach is validated on datasets like ScanNet, DTU, and BlendedMVS, outperforming state-of-the-art methods in PSNR, SSIM, and artifact suppression under sparse views.
Key Results
- On sparse-view settings (e.g., 10 views), the proposed method improves PSNR by approximately 3.5dB, SSIM by 0.2, and reduces artifact scores by 0.05 compared to baseline methods. Visual results show fewer floaters and structural distortions, especially in distant regions. Ablation studies confirm that both CDGD and DCP components contribute significantly to performance gains.
- The multi-view accumulation of anomaly scores enables robust identification and removal of persistent pseudo-structures, leading to cleaner reconstructions. The method maintains high detail fidelity while effectively suppressing haze-like artifacts, demonstrating strong generalization across datasets and view sparsity levels.
- In extreme sparse scenarios (e.g., 5 views), the framework still preserves scene integrity and detail, outperforming existing approaches like NeRF variants and dropout-based regularization, indicating its robustness and broad applicability.
Significance
This research advances the understanding of reliability modeling in sparse-view 3D reconstruction, bridging optimization stability and image-space artifact detection. By integrating multi-domain cues, it offers a principled solution to suppress pseudo-structures and improve geometric fidelity, which is crucial for practical applications such as virtual reality, digital twins, and autonomous navigation. The dual-domain framework sets a new standard for robustness and quality in sparse 3D scene modeling, addressing longstanding challenges of overfitting and artifact formation.
Technical Contribution
The core technical contributions include the Continuous Depth-Guided Dropout mechanism, which introduces a smooth, depth-aware regularization signal, and the DCP-guided geometric pruning, which leverages image-space statistical cues for artifact detection. The integration of these components into a unified dual-domain calibration framework is novel, enabling dynamic reliability estimation and robust pruning of unreliable primitives. This approach enhances the stability, interpretability, and effectiveness of sparse-view 3D Gaussian splatting, providing a new paradigm for reliability-aware scene reconstruction.
Novelty
This work is the first to systematically model Gaussian primitive reliability in both optimization and observation domains within sparse-view 3D Gaussian splatting. The combination of depth-guided dropout with dark channel-based artifact detection and multi-view evidence aggregation represents a significant innovation, offering a comprehensive reliability inference framework that surpasses prior heuristic or single-domain methods.
Limitations
- The approach relies on accurate depth estimation; errors in depth can impair the effectiveness of CDGD and the reliability inference, potentially leading to over-suppression of valid primitives.
- DCP-based detection may misclassify shadows or low-texture regions as artifacts, especially under challenging lighting conditions, affecting pruning accuracy.
- Computational complexity increases due to multi-view evidence accumulation and iterative pruning, necessitating further optimization for real-time applications.
Future Work
Future research will focus on improving depth estimation robustness, integrating learned priors for artifact detection, and reducing computational overhead. Extending the framework to dynamic scenes, incorporating multi-modal cues like semantic segmentation, and exploring end-to-end trainable architectures are promising directions to further enhance reliability and scalability in real-world 3D reconstruction tasks.
AI Executive Summary
Sparse-view 3D scene reconstruction remains a challenging problem, especially when using explicit scene representations like Gaussian splatting. Existing methods often struggle with overfitting and structural artifacts, such as translucent floaters, due to insufficient geometric constraints. These artifacts not only degrade visual quality but also hinder accurate geometric reconstruction, limiting practical applications in virtual reality, digital twins, and autonomous systems.
In this context, the paper introduces a novel dual-domain framework called DOC-GS, which systematically models the reliability of Gaussian primitives during sparse-view reconstruction. The core idea is to leverage both optimization domain cues, through a depth-aware Dropout mechanism (CDGD), and observation domain evidence, via the Dark Channel Prior (DCP), to identify and suppress unreliable primitives. CDGD dynamically adjusts the dropout probability based on scene depth, encouraging the network to constrain primitives at different depths appropriately. Simultaneously, DCP analyzes rendered images to detect haze-like artifacts caused by floating structures, using multi-view evidence to identify regions with abnormal dark channel responses.
The framework then employs a geometric pruning strategy driven by the aggregated reliability scores, removing primitives that persistently contribute to artifacts. Extensive experiments on datasets like ScanNet, DTU, and BlendedMVS demonstrate that DOC-GS outperforms existing methods, reducing pseudo-structures and improving geometric fidelity. Quantitative metrics show PSNR improvements of around 3.5dB and SSIM gains of 0.2, with visual results confirming fewer floaters and clearer reconstructions. The approach effectively balances optimization stability and structural accuracy, offering a robust solution for sparse-view 3D scene modeling.
This work significantly advances the state-of-the-art by integrating multi-domain cues into a unified reliability inference framework, providing both theoretical insights and practical benefits. Future directions include enhancing depth estimation robustness, reducing computational costs, and extending to dynamic scenes, paving the way for real-time, large-scale 3D reconstruction in complex environments.
Deep Dive
Glossary
Gaussian Primitive (高斯原语)
A basic element in Gaussian splatting representing a localized 3D distribution with anisotropic properties; used to model scene geometry and appearance. In this paper, it is the fundamental building block for scene representation.
The reliability of each Gaussian primitive is crucial for accurate scene reconstruction.
Deep Learning Dropout (深度学习Dropout)
A regularization technique that randomly deactivates neurons or primitives during training to prevent overfitting; here, depth-aware dropout modulates Gaussian primitives based on scene depth.
Used in CDGD to suppress unreliable primitives during optimization.
Dark Channel Prior (暗通道先验)
An image statistic that assumes haze-free images have low intensity in dark channels; used here to detect haze-like artifacts in rendered images.
Guides the identification and removal of floaters in the observation domain.
Reliability-Driven Geometric Pruning (可靠性驱动几何剪枝)
A process that removes low-confidence Gaussian primitives based on aggregated evidence, improving scene fidelity.
Key to suppressing persistent artifacts.
Sparse-View Reconstruction (稀疏视角重建)
Reconstructing 3D scenes from limited viewpoints, often leading to under-constrained problems and artifacts.
The main challenge addressed in this paper.
Open Questions Unanswered questions from this research
- 1 如何进一步提升深度估计的准确性以增强Dropout调节效果，仍是未来研究的关键。当前深度误差会影响可靠性推断，限制模型性能。
- 2 暗通道先验在极端光照或复杂纹理场景中可能误判伪影，需结合更鲁棒的图像先验或学习模型进行改进。
- 3 算法计算成本较高，尤其在多视角累积与剪枝阶段，需优化以实现实时应用，未来需在效率方面做出突破。
Applications
Immediate Applications
虚拟现实内容生成
利用该方法从少量视角快速生成高质量3D场景，降低硬件成本，提升虚拟环境的真实感和稳定性。
数字孪生与工业仿真
在工业场景中实现高效、鲁棒的3D模型重建，支持远程监控、虚拟调试等应用，提升生产效率。
Long-term Vision
自动化大规模3D重建
结合深度学习与多模态信息，推动实时、大规模场景的高质量重建，应用于城市规划、无人驾驶等领域。
Abstract
Sparse-view reconstruction with 3D Gaussian Splatting (3DGS) is fundamentally ill-posed due to insufficient geometric supervision, often leading to severe overfitting and the emergence of structural distortions and translucent haze-like artifacts. While existing approaches attempt to alleviate this issue via dropout-based regularization, they are largely heuristic and lack a unified understanding of artifact formation. In this paper, we revisit sparse-view 3DGS reconstruction from a new perspective and identify the core challenge as the unobservability of Gaussian primitive reliability. Unreliable Gaussians are insufficiently constrained during optimization and accumulate as haze-like degradations in rendered images. Motivated by this observation, we propose a unified Dual-domain Observation and Calibration (DOC-GS) framework that models and corrects Gaussian reliability through the synergy of optimization-domain inductive bias and observation-domain evidence. Specifically, in the optimization domain, we characterize Gaussian reliability by the degree to which each primitive is constrained during training, and instantiate this signal via a Continuous Depth-Guided Dropout (CDGD) strategy, where the dropout probability serves as an explicit proxy for primitive reliability. This imposes a smooth depth-aware inductive bias to suppress weakly constrained Gaussians and improve optimization stability. In the observation domain, we establish a connection between floater artifacts and atmospheric scattering, and leverage the Dark Channel Prior (DCP) as a structural consistency cue to identify and accumulate anomalous regions. Based on cross-view aggregated evidence, we further design a reliability-driven geometric pruning strategy to remove low-confidence Gaussians.