SplatAudit: Reference-Free Geometric Auditing of Floater Artifacts in 3D Gaussian Splatting

Ian Wai Si¹  ·  Elaine Huang²

¹Tsinghua University    ²Mission San Jose High School

KDD 2026 Undergraduate Consortium

Paper  |  KDD 2026  |  Installation  |  Usage  |  Citation

Official implementation of SplatAudit, a reference-free framework for auditing floater artifacts in trained 3D Gaussian Splatting (3DGS) scenes.

Overview

SplatAudit assesses the geometric coherence of pretrained 3DGS scenes without ground-truth geometry or retraining. It combines surface variation and mutual coverage over anisotropic 3σ support neighborhoods to produce:

• Per-Gaussian confidence for inspecting the learned representation.
• Artifact heatmaps and Floater Artifact Score (FAS) for view-dependent diagnostics. Lower FAS indicates less geometric artifact evidence.
• One-pass pruning, using geometric confidence alone or a harmonic combination with PointSplat importance, followed by refinement.

On seven Mip-NeRF 360 scenes with 24 training views, SplatAudit achieves the lowest FAS in every scene at 20% retention. The hybrid improves mean image quality over geometry alone at both 20% and 10% retention. See the paper for full results and ablations.

Installation

Requires Conda, an NVIDIA GPU, a compatible driver, and the CUDA 11.8 toolkit (including nvcc). The environment provides Python 3.10 and PyTorch 2.4.1.

git clone --recursive https://github.com/Gearoid522/SplatAudit.git
cd SplatAudit
conda env create -f environment.yml
conda activate splataudit
python -m pip install --no-build-isolation -e . third_party/gaussian-splatting/submodules/{diff-gaussian-rasterization,simple-knn,fused-ssim}

Conda packages use the Tsinghua TUNA mirror; PyTorch uses the official CUDA 11.8 wheel index. CUDA is required for model loading, including CPU geometry mode. If building for another GPU, set TORCH_CUDA_ARCH_LIST accordingly.

Usage

Audit a pretrained model

For an existing Graphdeco-format 3DGS model, export geometric confidence:

python cli/score.py \
  --scene /path/to/model \
  --method splataudit \
  --output /path/to/audit/splataudit_scores.npy \
  --geometry-device auto

--scene accepts a model directory (using its latest saved iteration) or a Graphdeco-format Gaussian PLY. The output contains one confidence value in [0, 1] per Gaussian, in PLY order; higher means stronger geometric support. Use --method hybrid to export the combined pruning score instead.

Train, prune, and evaluate

Prepare a Mip-NeRF 360 or COLMAP scene containing images/ and sparse/0/. Keep datasets and outputs outside the repository. The example below retains 20% of the baseline Gaussians.

# Train a sparse-24 baseline.
python cli/train.py \
  --scene-path /path/to/mipnerf360/bicycle \
  --seed 0 \
  --output /path/to/outputs/bicycle/baseline \
  --iterations 30000

# Prune once and refine without densification.
python cli/refine.py \
  --scene-path /path/to/outputs/bicycle/baseline \
  --train-path /path/to/mipnerf360/bicycle \
  --method splataudit \
  --keep 0.20 \
  --seed 0 \
  --output /path/to/outputs/bicycle/splataudit_20 \
  --iterations 5000

# Evaluate held-out views and save artifact heatmaps.
python cli/evaluate.py \
  --scene-path /path/to/outputs/bicycle/splataudit_20 \
  --source-path /path/to/mipnerf360/bicycle \
  --output /path/to/outputs/bicycle/splataudit_20/eval \
  --save-images

Training holds out every eighth camera and selects 24 evenly spaced training views; use --regime full for all remaining views. For hybrid pruning, change --method splataudit to --method hybrid. The paper's 10% setting uses --keep 0.10 --iterations 10000.

Refinement and evaluation use the view_split.json created during training. Evaluation saves PSNR, SSIM, LPIPS, and FAS in results.json, confidence in scores.npy, and optional images in renders/ and fas/.

The evaluation CLI normalizes FAS using the evaluated model's own alpha and foreground mask. The paper's cross-method FAS comparison uses the unpruned baseline's alpha and foreground reference.

Citation

If you use SplatAudit in your research, please cite the paper:

@misc{si2026splataudit,
  title  = {{SplatAudit}: Reference-Free Geometric Auditing of Floater Artifacts in {3D} Gaussian Splatting},
  author = {Si, Ian Wai and Huang, Elaine},
  year   = {2026},
  note   = {Accepted to the KDD 2026 Undergraduate Consortium},
  url    = {https://kdd2026.kdd.org/wp-content/uploads/2026/09/14-SplatAudit-Reference-Free-Geometric-Auditing-of-Floater-Artifacts-in-3D-Gaussian-Splatting.pdf}
}

Acknowledgements

Built on 3D Gaussian Splatting. The hybrid score uses PointSplat intrinsic importance. We evaluate on Mip-NeRF 360.