TIDI-GS: Floater Suppression in 3D Gaussian Splatting for Enhanced Indoor Scene Fidelity

TL;DR

TIDI-GS presents a plug-in training framework that substantially reduces floater artifacts in indoor 3D Gaussian Splatting without altering the core rendering pipeline. It combines evidence-based floater pruning, detail-preserving guards, and isolation-aware removal with an uncertainty-weighted monocular depth regularizer to steer geometry toward plausible surfaces. The method yields cleaner, more stable reconstructions with improved LPIPS perceptual quality and depth consistency, while maintaining training efficiency comparable to baseline 3DGS. By aligning multi-view evidence and geometric uncertainty, TIDI-GS advances the reliability and usability of indoor digital twins and artifact-aware optimization in splat-based rendering.

Abstract

3D Gaussian Splatting (3DGS) is a technique to create high-quality, real-time 3D scenes from images. This method often produces visual artifacts known as floaters--nearly transparent, disconnected elements that drift in space away from the actual surface. This geometric inaccuracy undermines the reliability of these models for practical applications, which is critical. To address this issue, we introduce TIDI-GS, a new training framework designed to eliminate these floaters. A key benefit of our approach is that it functions as a lightweight plugin for the standard 3DGS pipeline, requiring no major architectural changes and adding minimal overhead to the training process. The core of our method is a floater pruning algorithm--TIDI--that identifies and removes floaters based on several criteria: their consistency across multiple viewpoints, their spatial relationship to other elements, and an importance score learned during training. The framework includes a mechanism to preserve fine details, ensuring that important high-frequency elements are not mistakenly removed. This targeted cleanup is supported by a monocular depth-based loss function that helps improve the overall geometric structure of the scene. Our experiments demonstrate that TIDI-GS improves both the perceptual quality and geometric integrity of reconstructions, transforming them into robust digital assets, suitable for high-fidelity applications.

Figures (10)

• Figure 2: Conceptual illustration of floater cues. Our method identifies floaters based on a combination of weak evidence signals, including low opacity, low multi-view visibility, low optimization activity (Gradient EMA), and high spatial isolation (High Distance Factor). These signals are used to trigger a periodic pruning pass to clean the scene representation.
• Figure 3: The pipeline for identifying the base candidate set $\mathcal{C}_{\text{base}}$ for pruning. A Gaussian is selected for this set if it falls below predefined thresholds for all four weak evidence signals shown: (i) its visibility count ($v_i$) indicates insufficient multi-view support; (ii) its opacity ($\alpha_i$) shows a negligible radiometric contribution; (iii) its learned importance ($\sigma(\omega_i)$) is low; and (iv) its position-gradient EMA ($||\nabla x_i||_{\text{EMA}}$) is small. This multi-criteria approach creates an initial set of potential floaters that will be further refined by the detail-preserving guards.
• Figure 4: Empirical pruning rate as a function of spatial isolation and learned importance. The high pruning rate (red) is concentrated in the region of high spatial isolation (large $k$-NN distance) and low learned importance, confirming our method's targeted removal of floaters.
• Figure 5: TIDI-GS training dynamics. (Left) Photometric (PSNR, LPIPS) and geometric (Depth Error) quality metrics improve concurrently throughout training. (Right) The total number of Gaussians grows and then stabilizes as the pruning activity (red bars) increases in the later stages, indicating a shift from a growth phase to a refinement phase.
• Figure 6: Visualization of the detail-aware pruning pipeline across different scenes. From left to right: (1) All Gaussians, (2) Candidate floaters ($\mathcal{C}_{\text{base}}$) in red, (3) Detail guarded candidates ($M_{\text{detail}}$) in blue, and (4) Finally removed Gaussians in green. The process effectively isolates sparse floaters while preserving the dense structure of the underlying geometry.