Skip to content

Gaussian Splatting Trainers — Technical Benchmark and Visual Analysis

This directory contains the full benchmarking material used to compare multiple open-source Gaussian Splatting implementations under controlled and reproducible conditions.

Evaluations are organized by scene type (indoor vs outdoor) and include both: - quantitative evaluations (model size, Gaussian count, training time), and - qualitative visual analyses performed before and after post-processing with SuperSplat.

All experiments were executed on the same hardware platform and following consistent protocols to enable fair cross-method comparisons.

Contents

  • Indoor — Indoor benchmark: training statistics, raw vs cleaned visual inspection, and cleaning impact.
  • Outdoor — Outdoor benchmark: training statistics, raw vs cleaned visual inspection, and cleaning impact.
  • Indoor vs Outdoor — Cross-scenario analysis highlighting structural differences between indoor and outdoor reconstructions.

Benchmark Scope

The benchmarks evaluate multiple Gaussian Splatting pipelines under comparable conditions:

  • fixed datasets per scenario,
  • consistent hardware usage,
  • uniform inspection and cleaning workflow in SuperSplat,
  • export to .ply format for inspection and post-processing.

Both raw reconstructions and cleaned outputs are analyzed to assess:

  • reconstruction density and spatial extent,
  • artifact distribution (halos, streaks, floating clusters),
  • impact of post-processing on size and structure,
  • preservation of relevant structural elements.

Trainers Evaluated

The following open-source pipelines are benchmarked in both scenarios:

  • Inria gaussian-splatting
  • gsplat
  • OpenSplat
  • Nerfstudio
  • LichtFeld Studio (MCMC densification enabled)

Run instructions are linked in the tables inside Indoor and Outdoor.

Hardware Configuration

All tests were conducted on:

  • Windows 11
  • Single NVIDIA RTX 4060 GPU

Experimental Protocol

All methods were trained with:

  • 30,000 optimization iterations
  • the same 151-frame image set per scenario
  • default hyperparameters
  • exported reconstructions converted to .ply for inspection and downstream processing

Timing and Size Conventions

  • Training time refers only to the pipeline’s automatic training phase.
  • Output size (MB) refers to the exported .ply file size.

SuperSplat Inspection and Cleaning Workflow

All reconstructions follow a consistent three-stage qualitative workflow.

Raw inspection

The exported .ply is opened in SuperSplat to assess: - compactness of the reconstructed volume, - outlier/halo distribution, - detached clusters and far-field artifacts, - streak / spike-like structures.

Cleaning

The scene is cleaned in SuperSplat using a consistent set of principles: - spatial restriction (when feasible), - distance pruning (conservative outdoors), - opacity filtering, - scale-based filtering (x/y/z), - surface-area filtering, - manual refinement to preserve meaningful geometry,
then re-exported as a cleaned .ply.

Post-cleaning inspection

This stage focuses exclusively on the appearance of the cleaned models, highlighting:

  • changes in spatial compactness,
  • removal of peripheral noise and floating clusters,
  • persistence of residual streak artifacts,
  • preservation of structural and environmental elements.

Each benchmark file includes screenshots and orbit videos captured in SuperSplat for both raw and cleaned reconstructions.

Scene Categories

Two scene categories are considered, each represented by a single benchmark scene:

Indoor

A single confined environment captured in an indoor setting.
See: Indoor.

Outdoor

A single large-scale environment captured in an outdoor setting.
See: Outdoor.

Cross-Scenario Analysis

A dedicated comparison between indoor and outdoor results is provided in: Indoor vs Outdoor.

This section discusses:

  • differences in post-processing difficulty,
  • spatial extent of raw reconstructions,
  • cleaning efficiency and pruning limits,
  • quantitative and qualitative trade-offs.

Relation to XR Viewer Evaluation

This directory is complemented by the material in XR Viewers.

This section evaluates reconstructions structure, density and post-processing impact, while the XR viewer analysis evaluates how the cleaned reconstructions behave when rendered in immersive environments.

Together, both directories provide a full pipeline perspective:

Training → Cleaning → XR Visualization → Runtime Behavior

How to Navigate the Results