Skip to content

Gaussian Splatting XR Viewers — Visual and Performance Analysis

This directory contains the XR visualization tools evaluation conducted on Gaussian Splatting reconstructions generated during the trainer analysis phase.

The analysis focuses on the behavior of two different XR rendering environments when visualizing identical cleaned reconstructions produced by multiple training pipelines.

The evaluation includes both:

  • qualitative visual inspection (artifact behavior, motion stability, depth consistency), and
  • quantitative runtime measurements (FPS and frame timing).

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

Contents

  • XR Viewers — Comparative analysis of two XR Gaussian Splatting viewers, including visual inspection, cross-viewer comparison, and runtime performance evaluation.

Evaluation Scope

The viewer analysis builds directly on the trainer evaluations:

  • the same 151-frame indoor dataset is used,
  • the reconstructions correspond to the models generated by the five training pipelines and analyzed in Indoor,
  • only the cleaned versions of these reconstructions — whose post-processing impact is examined in Indoor vs Outdoor — are used for the XR evaluation,
  • the cleaned .ply files are loaded locally into the XR viewers.

The indoor scenario was selected based on the structural and post-processing analysis reported in Indoor vs Outdoor.

XR Viewers Evaluated

Two Unity-based XR Gaussian Splatting viewers are compared:

  • Unity-VR-Gaussian-Splatting (ninjamode)
  • GaussianSplattingVRViewerUnity (clarte53)

Both are evaluated in immersive VR conditions using the same hardware setup.

Hardware Configuration

All tests were conducted on:

  • Windows 11
  • Single NVIDIA RTX 4060 GPU
  • Meta Quest 3 headset

Visual Inspection Protocol

Qualitative inspection is based on a 5-point ordinal scale evaluating:

  • surface solidity,
  • motion parallax stability,
  • edge stability,
  • transparency and depth ordering,
  • near-object robustness.

This stage identifies reconstruction-dependent artifacts and viewer-dependent motion behavior.

Cross-Viewer Comparison

A dedicated cross-viewer comparison is included to directly compare rendering behavior between the two XR environments under identical conditions.
This comparison highlights differences in motion stability, artifact manifestation, depth consistency, and overall immersive comfort.

Runtime Performance Measurements

After selecting the more stable viewer-Unity-VR-Gaussian-Splatting (ninjamode)-, quantitative measurements were conducted to compare runtime performance across training pipelines.

Metrics are:

  • Average FPS
  • Average frame time (ms)
  • Average maximum frame time (ms)

Measurements were recorded over a fixed acquisition window of 30s after a load time of 5s.

Relation to Trainer Benchmarks

This directory complements the material in Trainer Benchmarks.

While the trainer benchmarks analyze reconstruction structure and post-processing behavior, this section evaluates: how those reconstructions behave in immersive XR environments and how viewer implementation affects visual stability.

Together, both directories provide a full pipeline perspective:

Training → Cleaning → XR Visualization → Runtime Behavior

How to Navigate the Results