PaperlayerView exact revision on arXiv

Source-linked AI summary

DomeGS: Perceptually Consistent Background Completion for Unbounded 3D Gaussian Splatting in Virtual Reality

Anastasiya Pechko, Agata Szymańska, Piotr Borycki, Joanna Szelewska, Michał Mnich, Sławomir Tadeja, Przemysław Spurek

arXiv:2610.37172v1cs.GR

TL;DR

Unbounded outdoor 3DGS can leave distant skies and horizons incomplete, while a separately chosen panorama may not blend with the captured scene. DomeGS integrates a generated background as a finite Gaussian skydome, and a user study reports improved background consistency, spatial presence, and realism versus a skybox baseline.

  • Problem

    Sparse or missing far-field capture data limits 3DGS reconstruction of plausible distant geometry, and independently composited panoramas can mismatch the captured scene.

    Source: §1 · pp. 1–2

  • Method

    DomeGS uses 360°-aware generative outpainting to supervise refinement of a finite-radius Gaussian skydome unified with the near-field scene.

    Source: §1 · p. 2

  • Results

    DomeGS significantly improved background consistency under motion, spatial presence, and experienced realism compared with the classical skybox baseline.

    Source: §8 · p. 9

  • Takeaways & Limitations

    The unified 3DGS asset is designed to preserve depth cues and provide coherent background completion without degrading the primary reconstruction.

    Source: §1 · p. 2

    The formulation requires radiometric coherence with the reconstructed scene and depth-consistent parallax under translational head movement.

    Source: §3.1 · p. 3

Abstract

from arXiv · show

The growing need for high-quality 3D content remains among the most prevalent challenges in virtual reality (VR). Recently, novel view synthesis techniques have enabled rapid content creation, with 3D Gaussian Splatting (3DGS) emerging as one of the most promising methods for generating high-fidelity 3D scenes. However, using 3DGS to reconstruct unbounded outdoor scenes can leave horizons incomplete due to missing data or a lack of distant geometric constraints. Such incompleteness can break immersion, compromising VR's core characteristic. A typical workaround in traditional VR systems is a textured cube (skybox) or sphere (skydome) surrounding the environment to create the illusion of a distant background. However, these approaches struggle to blend seamlessly with the reconstructed 3DGS geometry. To address these issues, we introduce the DomeGS framework, which incorporates the skydome concept not as a flat background shell, but as a part of the reconstructed Gaussian Splatting 3D scene. By integrating the horizon and sky into the 3DGS representation, our method provides a unified, scene-consistent environment bridging foregrounds and backgrounds for greater realism. Experimental evaluation with N=36 users demonstrates that DomeGS significantly increases spatial presence (p=0.042) and realism (p=0.042) compared to the skybox baseline.

1 INTRODUCTION

Unbounded outdoor 3DGS can leave distant regions incomplete, while independently composited backgrounds can mismatch the reconstructed scene. DomeGS integrates a generated distant background into the Gaussian representation, and its user study reports higher presence and realism than a skybox baseline.

1 INTRODUCTION

  • Sparse or absent far-field capture data leaves 3DGS without sufficient parallax and landmarks to reconstruct plausible distant geometry, risking gaps, floaters, and jagged boundaries in VR.The text identifies skies, distant facades, and the horizon as affected regions.Source: §1 · pp. 1–2
  • A separately composited panoramic background can create color or illumination discrepancies at the boundary with near-field Gaussians.The paper contrasts this mismatch with its scene-integrated background approach.Source: §1 · p. 2
  • DomeGS significantly increased sense of presence and realism versus the skybox baseline, while the study also reported improved background consistency under motion and comparable behavior and comfort.The introduction reports p = 0.042 for both presence and realism.Source: §1 · p. 2
  • DomeGS completes missing distant regions with 360°-specialized generative outpainting, then uses the completed map to supervise Gaussian refinement on a finite-radius sphere.This unifies the far-field skydome and near-field scene in one 3DGS asset.Source: §1 · p. 2

2 RELATED WORKS

Prior work covers distant-background representations, generative environment completion, and 3DGS reconstruction, but the supplied review identifies a gap in completing existing 3DGS scenes with scene-consistent generated surroundings.

2.1 Distant Background Representations for VR

  • A skybox maps panoramic imagery onto an effectively infinite cubemap, whereas a finite skydome can provide depth-dependent motion parallax under viewer translation.The review notes that skydomes require custom mesh and shader integration.Source: §2.1 · p. 2

2.2 Generative Environment Completion

  • Generative methods synthesize unseen environment content, but independently selected panoramas can mismatch appearance and illumination already baked into captured 3DGS scenes.Prior panorama-to-Gaussian methods generate new scenes rather than completing an existing reconstruction under explicit scene-consistency constraints; DomeGS targets this gap with a finite Gaussian skydome.Source: §2.2 · pp. 2–3

2.3 3DGS for Unbounded Scene Reconstruction

  • For unbounded outdoor 3DGS, distant regions may remain unobserved; prior approaches mask or separately model them, but do not synthesize content entirely absent from input imagery.The review also notes that prior evaluations typically use static viewpoints rather than movement in immersive environments.Source: §2.3 · p. 3

3 METHODOLOGY

DomeGS turns panorama completion into a staged process that converts missing background regions into a geometrically grounded multi-view dataset and fuses it with the original scene as one 3DGS representation.

3.1 Preliminaries and Problem Formulation

  • DomeGS expands the initial scene with finite-radius background Gaussians, targeting radiometric coherence and depth-consistent parallax under head translation.The background is jointly rasterized with the original scene under arbitrary HMD poses.Source: §3.1 · p. 3

3.2 Stage I: Panorama Extraction and Outpainting

  • The panorama stage renders an equirectangular view from the trajectory centroid, masks unobserved sky and voids, then outpaints and harmonizes the 360° wrap seam.A global outpainting pass is followed by a shifted, seam-focused inpainting pass.Source: §3.2 · p. 3

3.3 Stage II: Scene-Adaptive Dataset Preparation

  • Scene-adaptive preparation estimates shell radius from visible scene depths and base pitch by balancing visible foreground density against unobserved sky voids.The pitch objective uses confident foreground Gaussians and visible primitives at each probe pitch.Source: §3.3 · pp. 3–4
  • The method prunes backside floaters and builds stratified ground, transition, and sky views with role-dependent supervision that places background behind foreground silhouettes without mesh segmentation.The role-dependent formulation uses physically grounded alpha compositing to guide background Gaussians into empty regions behind foreground geometry.Source: §3.3 · p. 4

3.4 Stage III: Unified Sky Gaussian Optimization

  • Optimization preserves near-field Gaussians, seeds an upper-hemispherical background shell, and applies radial-shell and planarity regularizers to background Gaussians.The shell thickness is set by γ = 0.99; seed Gaussians are initialized at scales adapted to local sampling density and optimized with the original scene.Source: §3.4 · p. 4
  • Seed Gaussians densify into the final background; seeds are protected from pruning while newly created Gaussians can be removed, and scale-based culling is disabled across the background.The result combines the fine-tuned original scene with the optimized background as one unified representation.Source: §3.4 · p. 4

4 ABLATION: GENERATIVE ENVIRONMENT COMPLETION

The authors compared generative outpainting methods for completing a 360° panorama, seeking continuity across the spherical image. They selected FLUX.2-klein-base-9B with a 360° ERP LoRA for the VR conditions.

4 ABLATION: GENERATIVE ENVIRONMENT COMPLETION

  • The same completed panorama was used in both VR conditions, with completion required to preserve continuity across the spherical image.Source: §4 · p. 4
  • Other tested generators showed distinct artifacts: FLUX.2-dev had color mismatches, FLUX.1-Fill-dev repeated scene parts, and 512-inpainting-ema and Nano Banana produced boundary seams.Figure 4 compares the generators on the same masked panorama.Source: §4 · p. 4
  • FLUX.2-klein-base-9B with a 360° ERP LoRA gave the best overall result, with better color consistency and reduced seam artifacts.The full candidate list and two-stage wrap procedure for removing the remaining ERP join are reported in the supplementary material.Source: §4 · p. 4

5 USER STUDY

The study used a counterbalanced, within-subjects design to compare DomeGS and Skybox across four scenes. Thirty-six participants completed the study in a wireless VR setup.

5 USER STUDY

  • Each participant experienced all four scenes with both DomeGS and Skybox; technology order was balanced and scene order randomized across six Latin-square sequences.The study received ethics committee approval.Source: §5 · p. 4

5.1 Participants

  • Thirty-six participants aged 20–37 completed the study, exceeding the analytic minimum of 34; their VR/AR experience was generally low.The sample included 31 male and 5 female volunteers; demographic variables did not differ significantly between presentation-order groups (all p > 0.22).Source: §5.1 · p. 5

5.2 Apparatus

  • Scenes were rendered in Unity at 90 Hz on a wireless Meta Quest Pro connected to a workstation with an NVIDIA GeForce RTX 4070 SUPER GPU.Participants used Meta Quest Touch Pro controllers.Source: §5.2 · p. 5

5.3 Experimental Task

A structured bird-search task guided broad visual exploration while focusing attention on the upper hemisphere and horizon. The within-subjects design crossed rendering technology with scene type across four outdoor reconstructions.

5.3 Experimental Task

  • The bird-search task directed attention to the upper hemisphere and horizon, where Skybox and DomeGS parallax differences are most prevalent, with about 2–3 minutes of exposure per scene.The focused task was chosen over free exploration to standardize visual attention and keep exposure concise.Source: §5.3 · p. 5
  • Teleport pads followed training-camera viewpoints, and participants used the same pad layout and visit order in both conditions to keep paths consistent and views within reliable reconstruction angles.Source: §5.3 · p. 5
  • One bird appeared per pad after teleportation, outside the frontal ±65° zone; scene-specific elevation bands and geometry-scaled distances were used to keep search difficulty comparable.Elevation bands were [−5°, +45°] for Garden and Bear and [−15°, +40°] for Buddha and Tower; distances ranged from 35% to 85% of maximum inter-pad distance (3.0–12.0 m).Source: §5.3 · p. 5

5.4 Independent Variables

  • The 2 × 2 within-subjects design crossed Skybox versus DomeGS with panoramic versus close scene types, using identical foreground Gaussians, panorama, and viewpoints across technologies.Four outdoor reconstructions represented the two scene types: Buddha and Tower were panoramic, while Garden and Bear were close.Source: §5.4 · pp. 5–6
  • Under forward translation, DomeGS maintained background alignment in Garden and represented the Bear tree’s changing proximity, unlike the Skybox’s static background and failure to scale the tree.Source: Figure 6 · p. 6

5.5 Dependent Measures

The study combined in-task perceptual ratings, questionnaires, preference responses, and behavioral measures. These measures assessed presence, background consistency, experience, comfort, and task difficulty.

5.5 Dependent Measures

  • After each bird-search trial, two in-VR single-item probes rated immersion and presence (IMM) and perceived background consistency and stability during movement and teleportation (SKY).The probes used a 7-point Likert-like scale from −3 to +3 and yielded 288 observations per measure.Source: §5.5 · p. 6
  • After each technology condition, participants completed the 13-item IPQ-PL, covering general presence, spatial presence, involvement, experienced realism, and an overall score.Source: §5.5 · p. 6
  • Participants reported forced-choice or tied preferences across four dimensions, while pre- and post-study SSQ scores monitored cybersickness symptoms.The preference dimensions were overall immersion, close-range inspection, panoramic viewing, and willingness for prolonged stay.Source: §5.5 · p. 6
  • Bird-search duration, head trajectory length excluding teleportation jumps, and target discovery latency were logged to check whether task difficulty remained balanced across conditions.Source: §5.5 · p. 6

5.6 Hypotheses

The hypotheses predict that DomeGS will improve immersion, background consistency, presence, and preference over Skybox, while leaving task performance equivalent. The protocol counterbalances both technology and scene order and repeats the same scenes under each representation.

5.6 Hypotheses

  • DomeGS is hypothesized to improve immersion, background consistency under motion, presence, and user preference relative to Skybox.The behavioral hypothesis predicts equivalence in search duration, head path, and rotational exploration.Source: §5.6 · p. 6

5.7 Experimental Protocol

  • Participants completed four scenes under both representations, with technology order counterbalanced and scene order held constant across conditions.Target layouts were mirrored between conditions; each trial included observation, a three-pad search task, and an in-VR micro-probe.Source: §5.7 · p. 6
  • Sessions ended with a second presence questionnaire, a forced-choice technology preference survey, and a final simulator-sickness questionnaire.Sessions lasted approximately 30–40 minutes, including 15–25 minutes of HMD exposure.Source: §5.7 · p. 6

5.8 Statistical Analysis

The analyses use repeated-measures methods suited to the rating, preference, and behavioral outcomes. Equivalence testing checks whether behavioral differences remain within prespecified medium-effect bounds.

5.8 Statistical Analysis

  • Micro-probe ratings were analyzed with mixed-effects models including technology, scene, their interaction, and participant random intercepts.Models used REML; fixed effects used Satterthwaite degrees of freedom and 95% Wald confidence intervals, with Holm-adjusted post-hoc contrasts.Source: §5.8 · pp. 6–7, Table 2 · p. 7
  • IPQ subscale differences were tested with paired t-tests or Wilcoxon signed-rank tests according to normality, with Holm correction across five outcomes.The table reports estimated DomeGS−Skybox differences on the −3 to +3 scale, confidence intervals, adjusted p-values, and standardized effects.Source: §5.8 · p. 7, Table 2 · p. 7
  • Preference was tested against equal preference among non-tied responses, while behavioral equivalence used TOST with medium-effect bounds.Behavioral bounds were ±60 seconds, ±6 meters, and ±2600° for search duration, head path, and yaw-plus-pitch; paired tests and scene-difficulty and order checks were also reported.Source: §5.8 · p. 7

6 USER STUDY RESULTS

Across the user study, DomeGS improved immersion and motion-related background consistency, and showed selective gains in presence and realism. Behavioral load was equivalent across technologies, cybersickness remained mild, and automated visual-quality ratings favored DomeGS.

6.1 Immersion and Background Consistency (H1)

  • DomeGS increased perceived immersion (Δ=+0.44, Holm p<0.001) and background consistency under motion (Δ=+1.24, Holm p<0.001) relative to Skybox.The consistency advantage varied by scene: it was significant in Buddha, Bear, and Garden, but not Tower; immersion did not significantly vary by scene.Source: Figure 7 · p. 7, §6.1 · p. 7

6.2 Preconceived Presence and Preference (H2)

  • Spatial presence (1.73 vs. 1.41, Δ=+0.32, Holm p=0.042) and experienced realism (0.36 vs. −0.07, Δ=+0.44, Holm p=0.042) were higher with DomeGS.DomeGS was preferred for panoramic scenes and overall immersion; 75.0% chose it for longer VR visits versus 8.3% for Skybox, while the close-scene preference contrast was not significant.Source: §6.2 · pp. 7–8, Figure 8 · p. 8

6.3 Task Performance and Manipulation Checks (H3)

  • Search duration, head path, and rotational exploration were equivalent within the prespecified medium-effect bounds across DomeGS and Skybox.Participants explored broadly in both conditions, and task metrics detected greater movement and search demands in close than panoramic scenes.Source: §6.3 · p. 8, Table 3 · p. 8

6.4 Cybersickness and Comfort

  • Cybersickness scores did not differ significantly between DomeGS and Skybox, and total scores remained in the mild range.Total SSQ scores were 16.5 versus 17.9, respectively (Δ=−1.35, Holm p=1.000); pre-to-post total accumulation was also not significant.Source: §6.4 · p. 8

6.5 Automated Vision-Language Quality Evaluation

  • Automated vision-language evaluation rated DomeGS above Skybox across datasets and visual-quality metrics, with pairwise comparisons also favoring DomeGS.The evaluation assessed boundary continuity, lighting coherence, and perspective alignment using absolute scores and position-swapped pairwise comparisons.Source: §6.5 · p. 8, Table 4 · p. 8

7 DISCUSSION

The controlled comparison isolates the far-field representation: DomeGS unifies the near-field scene and skydome in a finite 3DGS asset, unlike an infinite skybox. The discussion links the gains to motion parallax and notes that benefits depend on scene content and depth cues.

7 DISCUSSION

  • The study changed the far-field representation while holding the rest of the comparison to DomeGS versus an infinite skybox, isolating that representation’s effect.Prior approaches either suppress the sky or generate entire panoramic worlds, changing image content and representation together.Source: §7 · pp. 8–9
  • Background-consistency gains were largest in Buddha and nonsignificant in Tower, despite both being panoramic scenes.The authors tentatively attribute the difference to Buddha’s closer terrain and rock formations, which an infinite-distance skybox cannot represent as well as DomeGS.Source: §7 · p. 9
  • Only spatial presence and experienced realism were significant among IPQ subscales, while close-scene preference was nonsignificant and had the most tied responses.The discussion relates this pattern to depth-cue redundancy in cluttered near-field scenes and the greater salience of background coherence in open panoramic scenes.Source: §7 · p. 9
  • The perceptual gains came without a significant technology difference in SSQ scores, which remained mild.The authors connect this result to skydome parallax being geometrically compatible with user translation, while describing the relation to simulator sickness as consistent with that account.Source: §7 · p. 9

8 CONCLUSION

DomeGS integrates 360°-aware generative outpainting into a Gaussian-based skydome to complete 3DGS scenes while preserving depth cues and realism. In a within-subjects study, it outperformed the skybox baseline on background consistency under motion, spatial presence, and experienced realism.

8 CONCLUSION

  • DomeGS integrates 360°-aware generative outpainting into a Gaussian-based skydome for coherent background completion without degrading the primary reconstructed scene.The pipeline is designed to preserve depth cues and realism in immersive VR.Source: §8 · p. 9
  • The within-subjects study found significantly greater background consistency under motion for DomeGS than for the classical skybox baseline.The comparison included 36 participants.Source: §8 · p. 9
  • DomeGS significantly improved spatial presence and experienced realism compared with the classical skybox baseline.Source: §8 · p. 9
Loading 2610.37172v1…