Indoor vs Outdoor — Comparative Analysis

This document compares the behavior of the evaluated Gaussian Splatting pipelines across two different scene categories: an indoor and an outdoor setting. The comparison focuses on differences in:

raw reconstruction structure,
background modeling,
artifact distribution,
cleaning difficulty,
quantitative impact of post-processing,
emerging trends across environments.

All results discussed here are derived from the detailed analyses reported in indoor.md and outdoor.md.

Scope of the Comparison

The same five pipelines—Inria gaussian-splatting, gsplat, OpenSplat, Nerfstudio, and LichtFeld Studio—were evaluated on one indoor scene and one outdoor scene.

For both scenarios:

training was executed on the same hardware platform,
reconstructions were exported to .ply format,
a quantitative benchmarking analysis was conducted,
qualitative inspection and post-processing were performed using SuperSplat,
similar cleaning principles were applied, with parameters adapted to each scene.

The objective of this comparison is to highlight how each pipeline behaves when confronted with markedly different spatial layouts.

Quantitative Benchmarking Comparison

Expand to view cross-scenario results table and charts.

Tool	Gaussians Indoor	Gaussians Outdoor	Size Indoor (MB)	Size Outdoor (MB)	Training Indoor (min)	Training Outdoor (min)
Inria gaussian-splatting	955,819	777,067	226.1	183.8	120	60
gsplat	1,265,239	1,031,707	284.8	232.2	50	45
OpenSplat	510,870	589,291	120.8	139.4	60	50
Nerfstudio	170,150	197,545	40.2	46.7	30	25
LichtFeld Studio	1,000,000	1,000,000	236.5	236.5	60	50

Quantitative Visualizations

xychart-beta
    title "Indoor vs Outdoor — Raw Gaussian Count"
    x-axis ["Inria","gsplat","OpenSplat","Nerfstudio","LichtFeld"]
    y-axis "Gaussians" 0 --> 1400000
    line [955819,1265239,510870,170150,1000000]
    line [777067,1031707,589291,197545,1000000]

Legend: lighter line = indoor, darker line = outdoor

xychart-beta
    title "Indoor vs Outdoor — Training Time"
    x-axis ["Inria","gsplat","OpenSplat","Nerfstudio","LichtFeld"]
    y-axis "minutes" 0 --> 140
    line [120,50,60,30,60]
    line [60,45,50,25,50]

Legend: lighter line = indoor, darker line = outdoor

Observations

Inria gaussian-splatting produces fewer Gaussians and a smaller model for the outdoor scenario, together with a substantially shorter training time.
gsplat remains the densest pipeline in both scenarios, with consistently large models and only limited differences between indoor and outdoor.
OpenSplat increases both Gaussian count and file size for the outdoor scenario.
Nerfstudio remains the most compact pipeline in both settings, although the outdoor scene leads to a slightly larger model.
LichtFeld Studio reaches the same Gaussian budget in both scenarios, but trains faster on the outdoor one.

Raw Reconstructions: Structural Differences

Clear structural differences emerge between indoor and outdoor raw reconstructions.

Indoor

Indoor scenes are characterized by:

compact spatial envelopes,
well-defined scene boundaries,
limited far-field geometry,
peripheral Gaussians not associated with interior geometry,
streak artifacts mostly concentrated near the main structure.

Across pipelines, raw indoor reconstructions tend to occupy relatively constrained volumes, making spurious splats clusters easier to identify.

Outdoor

Outdoor scenes present fundamentally different challenges:

large spatial extents,
far-field environmental geometry corresponding to sky and vegetation,
peripheral halos of Gaussians around the main structure,
elongated streak artifacts extending from the core,
detached background elements that may still represent valid scene content.

Compared to indoor cases, outdoor raw reconstructions consistently exhibit wider spatial footprints and more complex splats distributions.

Cleaning Difficulty and Strategies

The qualitative differences between indoor and outdoor scenes are reflected directly in the complexity of post-processing.

Indoor Cleaning

Indoor cleaning operations were generally straightforward:

the scene core was easy to localize,
bounding volumes could be tightly defined,
aggressive spatial restriction was often feasible,
many peripheral Gaussians could be removed without damaging structural elements,
manual refinement was limited and primarily used for final inspection.

As a result, large portions of removable geometry could be pruned confidently.

Outdoor Cleaning

Outdoor cleaning proved substantially more challenging:

far-field Gaussians frequently corresponded to valid environmental elements,
background geometry often formed detached clusters rather than continuous shells,
sky-related streaks and large surface Gaussians required careful handling,
aggressive spatial pruning risked removing legitimate scene content,
manual inspection and selective filtering were critical at later stages.

Cleaning strategies therefore remained conservative, prioritizing preservation of environmental context over maximal size reduction.

Quantitative Impact of Cleaning

Expand to view cleaning impact comparison table and chart.

Tool	Δ Gaussians Indoor (%)	Δ Size Indoor (%)	Δ Gaussians Outdoor (%)	Δ Size Outdoor (%)
Inria gaussian-splatting	−45.8	−45.8	−52.5	−52.4
gsplat	−45.4	−45.4	−46.1	−46.1
OpenSplat	−21.2	−21.2	−21.3	−21.3
Nerfstudio	−25.4	−25.4	−13.5	−13.5
LichtFeld Studio	−20.0	−20.0	−58.4	−58.3

Quantitative Visualizations

xychart-beta
    title "Indoor vs Outdoor — Cleaning Impact (Δ Gaussians %)"
    x-axis ["Inria","gsplat","OpenSplat","Nerfstudio","LichtFeld"]
    y-axis "% removed" 0 --> 65
    line [45.8,45.4,21.2,25.4,20.0]
    line [52.5,46.1,21.3,13.5,58.4]

Legend: lighter line = indoor, darker line = outdoor

Observations

The numerical effect of post-processing differs markedly between indoor and outdoor settings.

In indoor reconstructions:

Inria gaussian-splatting and gsplat experience the largest reductions in Gaussian count and file size (≈ −45%),
OpenSplat and LichtFeld Studio show more moderate decreases (≈ −21% and ≈ −20%),
Nerfstudio occupies an intermediate regime (≈ −25%).

In outdoor reconstructions:

reductions are more variable across pipelines, ranging from limited (Nerfstudio, ≈ −13%) to very large (LichtFeld Studio, ≈ −58%),
Inria gaussian-splatting exhibits a substantial reduction after cleaning (≈ −52%),
gsplat also undergoes a strong reduction (≈ −46%),
OpenSplat again shows a moderate reduction (≈ −21%).

These results confirm that post-processing efficiency depends strongly on the spatial organization of the raw reconstruction.

Across environments, LichtFeld Studio and Inria gaussian-splatting are pruned more aggressively in outdoor reconstructions than in the indoor reconstructions, while gsplat and OpenSplat show smaller cross-scenario differences.

Nerfstudio behaves differently across settings: it undergoes moderate pruning in the indoor scene but only limited reductions in the outdoor scene, suggesting that a larger fraction of its outdoor Gaussians contribute to retained background geometry and cannot be safely removed without degrading scene completeness.

Qualitative Impact of Cleaning

Visual inspection across the two scenarios reveals systematic differences in how the same pipelines distribute Gaussians and how effectively artifacts can be controlled.

Raw Reconstructions

Across tools, the indoor scenes generally produced more spatially constrained raw outputs than outdoor scenes:

Inria gaussian-splatting remains relatively compact in both environments, but the outdoor reconstruction contains more detached peripheral elements and streak artifacts than indoors.
gsplat remains relatively compact in both environments, but the outdoor scene includes detached environmental splats and more far-field streaks extending from the scene core.
OpenSplat produces streak artifacts and peripheral clusters in both scenarios; however, in the outdoor scene these form broader halos compared to the indoor case.
Nerfstudio is characterized by sparse central reconstructions in both environments, with a wide halo of streaks, spike-like structures, and clusters of splats.
LichtFeld Studio reconstructs dense scene cores in both settings, with a large spatial footprint and peripheral halo (stronger in the outdoor reconstructuion).

Overall, outdoor raw reconstructions systematically contain larger far-field regions, halos around the scene core, and the sky reconstruction is a dominant source of large Gaussians and streak artifacts.

After Cleaning

Post-processing reduces clutter in both scenarios, but different residual patterns remain:

Inria gaussian-splatting converges to a compact reconstruction in both environments, though the outdoor scene retains more peripheral streaks than the indoor scene.
gsplat becomes tightly focused after cleaning in the indoor scene, whereas the outdoor reconstruction still preserves some streak artifacts near the core and in sky-related regions.
OpenSplat achieves strong spatial compaction in both cases, but the outdoor scene continues to show more significant residual elongated structures than the indoor one.
Nerfstudio transitions to very tightly cropped reconstructions in both environments, with the outdoor scene still retaining peripheral remnants and streaks.
LichtFeld Studio shows the largest qualitative contrast: while the indoor scene becomes compact with only faint residual halos, the outdoor scene preserves a dense core surrounded by thin but persistent peripheral halos.

In summary, indoor reconstructions tend to converge toward tightly bounded reconstructions after cleaning across pipelines, whereas outdoor reconstructions consistently preserve larger envelopes and detached structures, reflecting the increased complexity of environmental geometry and sky regions.

Conclusions

Considering the structural differences observed between the indoor and outdoor reconstructions, the indoor scene is selected as the reference case for the XR viewers analysis, as it provides a more controlled, spatially bounded, and structurally consistent benchmark across pipelines.

The analysis conducted on the reconstructions using the XR viewers is available here: analysis/viewers/xr-viewers.md

Summary

Outdoor reconstructions show larger far-field regions and more complex splats distributions than indoor reconstructions.
Indoor cleaning is generally more straightforward due to clearer scene boundaries and a more localized scene core.
Outdoor cleaning requires more cautious and selective pruning, since far-field Gaussians may correspond to valid environmental elements (e.g., sky and vegetation).
Quantitatively, outdoor post-processing amplifies inter-method variability (e.g., very large reductions for LichtFeld Studio and Inria gaussian-splatting, but limited reductions for Nerfstudio), while some pipelines (e.g., OpenSplat) show similar reduction rates across environments.
Combining indoor and outdoor benchmarks highlights tool behaviors that would not be visible when evaluating a single scene category.