acceptodds
Under review as a conference paper at ICLR 2027

Game2GS: Coverage-Aware Gaussianization for Large-Scale Free-View HDR Rendering

Abstract

Modern game worlds combine detailed geometry, materials, lighting, and engine-specific rendering effects, making their visual appearance costly to transfer across platforms. Distilling this appearance into 3D Gaussian representations offers a way to reuse game assets beyond the source engine, but requires faithful 6DoF rendering and exposure adaptation under complex lighting. Existing large-scale reconstruction pipelines are typically trained on images captured along prescribed street trajectories. Denser sampling along these trajectories increases acquisition costs without necessarily providing sufficiently resolved and directionally diverse observations of all relevant surfaces. This motivates selecting more informative views under a limited image budget. These views must also capture the world's bright and dark regions on a consistent radiometric scale to support runtime exposure adaptation. However, auto-exposed LDR capture introduces inconsistent brightness supervision across views, whereas fixed-exposure LDR capture can lose highlight or shadow detail. We introduce Game2GS, a framework combining coverage-aware view acquisition with HDR Gaussian reconstruction. Using known source geometry, we select camera poses derived from baked lighting probes to improve surface coverage through well-resolved observations from complementary directions and distinct positions. We then reconstruct an HDR Gaussian scene from fixed-exposure captures and enable real-time automatic exposure using an engine-exported compensation curve. Under matched image budgets, Game2GS improves PSNR by more than 3 dB over the strongest prescribed-view baseline across all evaluated scenes on independent 6DoF test views, and we will release our private game scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, View Acquisition, HDR Rendering, Game Environments, Novel View Synthesis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6rMGOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Game2GS, a framework for converting large-scale game environments into HDR Gaussian representations for free-view 6DoF rendering. The authors address two challenges: insufficient scene coverage caused by conventional trajectory-based image acquisition and inconsistent appearance supervision caused by automatic exposure.

The proposed framework uses source geometry and engine-native lighting probes to select informative camera views based on observation quality, directional coverage, and positional diversity. It also employs fixed-exposure HDR reconstruction with runtime exposure adaptation.

Critical strengths
  1. Optimizing camera acquisition under a limited image budget is a meaningful direction, especially for game environments where source geometry is available.
  2. The proposed method combines surface visibility, projected resolution, directional coverage, and distinct-position support to guide camera selection. Leveraging existing lighting probes and source geometry is a practical approach that avoids relying on an intermediate reconstruction for view selection.
Critical weaknesses
  1. Comparisons with simpler strategies, such as random sampling, uniform spatial sampling, or visibility-based greedy selection from the same candidate camera pool, would help determine whether the improvements come from the proposed optimization or primarily from broader spatial sampling.
  2. The framework targets large-scale scenes and real-time rendering, but the paper provides limited quantitative analysis of computational costs. In particular, the overhead of candidate generation, visibility evaluation, and greedy camera selection is not clearly reported.
  3. The view selection objective combines several components, including directional coverage, distinct-position support, spatial penalties, and height quotas. However, their individual contributions are not sufficiently evaluated.
Questions for authors

See the Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer 6rMG 1

We thank the reviewer for the constructive feedback and for recognizing the practical value of geometry-guided camera acquisition using engine-native lighting probes. We appreciate the suggestions regarding comparisons with simpler selection strategies, component-level ablations, and computational efficiency.

1. Comparison with Simpler View-Selection Strategies (Weakness 1)

We agree that comparisons against prescribed-view acquisition alone do not fully establish whether the improvements arise from the proposed joint objective or simply from broader spatial sampling.

Experiment 1: View Selection Strategy Comparison

To address this concern, we conducted controlled comparisons on Arena, using the same source mesh, validated candidate camera pool, image budget, fixed-exposure HDR reconstruction, training settings, and independent free-6DoF test views.

We compare our method against three simpler strategies:

  • Random probe sampling: Uniformly sampling camera poses at random from the same validated candidate pool.
  • Spatially uniform sampling (3D FPS): Selecting spatially distributed camera positions using farthest-point sampling, without explicitly optimizing surface observation quality.
  • Greedy surface coverage: Greedily maximizing the marginal gain in area-weighted surface coverage, retaining the highest-quality observation per surface without explicitly considering directional diversity or distinct-position support.

Table A. Comparison of view-selection strategies on Arena.

View Selection StrategyPSNR ↑SSIM ↑LPIPS ↓
Random probe sampling21.6360.61530.4945
Spatially uniform sampling (3D FPS)24.5580.72370.4372
Greedy surface coverage24.8630.72300.4366
Ours (Full Method)28.3680.83120.3354

Results and Discussion.

Spatially uniform sampling and greedy surface coverage achieve similar performance, suggesting that broader spatial sampling or maximizing single-best surface observation quality alone is insufficient for high-quality free-6DoF reconstruction.

In contrast, our full method improves PSNR by 6.73 dB, 3.81 dB, and 3.51 dB over random sampling, spatially uniform sampling, and greedy surface coverage, respectively, with consistent improvements in SSIM and LPIPS.

These results demonstrate the additional benefits of our joint selection strategy beyond broad spatial coverage alone, highlighting the importance of complementary observations from diverse viewing directions and distinct camera positions.

AuthorsOpenReview ↗

Response to Reviewer 6rMG 2

2. Computational Cost and Scalability (Weakness 2)

We thank the reviewer for highlighting the need for quantitative computational-cost analysis. Our geometry-guided acquisition is an offline preprocessing step, separate from reconstruction training and runtime Gaussian rendering.

Experiment 2A: Computational Cost Analysis

We profiled the complete offline acquisition pipeline, from loading precomputed lighting probes and DistanceMaps to exporting a final set of B_high.

Table B. Computational cost breakdown of geometry-guided view acquisition on MatrixCity Block A.

StageTime (min)
Initial and supplementary candidate generation25.45
Candidate organization and camera–surface relation construction44.46
Camera safety checks and mesh-based occlusion verification41.02
Coverage-gap seed and coarse selection6.89
Final greedy selection1.65
Surface sampling and other preprocessing12.63
Total recorded successful-stage time132.11

The timings cover processing from precomputed probes and DistanceMaps to exported camera poses, excluding probe baking, image capture, and 3DGS training.

Results and Discussion.

The entire offline acquisition preparation requires approximately 132 minutes, with most of the computational cost arising from candidate–surface relation construction, candidate generation, and geometric validation.

Importantly, the final greedy selection of B_high camera poses takes only 98.934 seconds (approximately 1.25% of the total), demonstrating that the main computational overhead lies in preparing and validating geometric observations rather than in the final view-selection optimization.

Our implementation uses sparse camera–surface relations and lazy priority-queue updates, as described in Appendix A.7, to reduce the computational cost of selection.

Experiment 2B: Runtime Rendering Efficiency

We additionally evaluated rendering efficiency on a large-scale scene containing approximately 280 million Gaussians, with up to 4 million Gaussians rendered per frame. The rendering performance reaches 132.67 FPS for HDR and 140.28 FPS for LDR.

These results provide additional evidence of practical real-time rendering performance. Furthermore, camera acquisition and view selection are performed entirely offline and introduce no additional per-frame view-selection overhead during rendering.

We will incorporate the computational breakdown and clarify the distinction between offline acquisition preprocessing and runtime rendering costs in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer 6rMG 3

3. Contribution of Individual View-Selection Components (Weakness 3)

We thank the reviewer for highlighting the need to further validate the contributions of the individual components in our view-selection pipeline. To further identify the contributions of individual components, we perform leave-one-component-out ablations on Arena, using the same validated candidate pool, image budget, HDR reconstruction settings, and free-6DoF test views.

Experiment 3: View-Selection Component Ablation

Our view-selection pipeline consists of the following components:

                 Validated Camera Candidates C
                              +
                   Source Mesh Surface Elements
                              |
                              v
                 Surface Observation Quality
           (Visibility, Range, Incidence, Resolution)
                              |
                +-------------+-------------+
                |                           |
                v                           v
 (1) Directional Coverage (F_dir)   (2) Distinct-Position Support (F_sup)
                |                           |
                +-------------+-------------+
                              |
                              v
                       Joint Objective
                 J(V) = F_dir(V) + λF_sup(V)
                              |
                              v
                 Greedy Selection Iteration
                 --------------------------
                 Candidate Eligibility:
                   - Positive marginal gain
                   - (4) Height Quotas
                   - (5) Redundancy Suppression

                 Candidate Scoring:
                   - (3) Spatial Penalty
                   - Score = s(c|V) × ΔJ(c|V)
                 --------------------------
                              |
                              v
                 Select Best View & Update V
                              |
                              v
          Repeat Until Budget B or No Eligible View

Table C. Leave-one-component-out ablation of the coverage-aware view-selection method on Arena.

ConfigurationDirectional CoveragePosition SupportSpatial PenaltyHeight QuotasRedundancy SuppressionPSNR ↑SSIM ↑LPIPS ↓
w/o Directional Coverage—✓✓✓✓23.6360.63670.5201
w/o Position Support✓—✓✓✓25.6390.75240.4224
w/o Spatial Penalty✓✓—✓✓27.1330.80130.3824
w/o Height Quotas✓✓✓—✓25.3740.74230.4233
w/o Redundancy Suppression✓✓✓✓—28.0420.81250.3994
Ours (Full Method)✓✓✓✓✓28.3680.83120.3354

Results and Discussion.

The ablation results demonstrate the importance of the individual components in our view-selection pipeline. Removing directional coverage causes the largest performance degradation (−4.73 dB PSNR), highlighting the importance of observing scene surfaces from complementary viewing directions. Removing height quotas and distinct-position support also leads to substantial drops (−2.99 dB and −2.73 dB, respectively), supporting the need for balanced height coverage and qualified observations from multiple camera positions.

Spatial penalties and redundancy suppression provide further improvements, with the latter showing a particularly noticeable effect on LPIPS. Overall, these results confirm that high-quality free-6DoF reconstruction requires not only broad surface coverage, but also complementary viewing directions, multi-position support, and balanced spatial allocation.

Reviewer_wNcROpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Game2GS, a pipeline for distilling game-engine scenes into HDR 3D Gaussian representations. It has two parts: (i) geometry-guided view acquisition, and (ii) HDR reconstruction with runtime exposure. For (i), camera candidates are proposed from baked GI probes and their distance maps. A budget of B views is then selected greedily with a coverage objective that combines directional coverage, distinct-position support, height quotas, a spatial penalty, and redundancy suppression. For (ii), the method trains on fixed-exposure scene-linear HDR images with a luminance–chromaticity Gaussian parameterization (following LCD-GS) and a μ-law loss. At runtime it applies an engine-exported exposure-compensation curve. The method is integrated into Horizon-GS and Urban-GS and evaluated on two benchmark scenes (MatrixCity Block A, Elvenruins) and two private scenes. The paper reports >3 dB PSNR gains on “free-6DoF” test views.

Critical strengths

The problem framing is useful and practical: when source geometry is available, view acquisition can be planned before reconstruction rather than relying on prescribed trajectories. Adding a free-6DoF evaluation protocol on top of trajectory-near evaluation is a valuable point, and Table 2 convincingly shows that trajectory-near metrics can hide off-manifold failures. Separating scene-linear HDR capture from display exposure is a clean design choice for game assets. The qualitative indoor/outdoor example in Fig. 3 illustrates a real failure mode of auto-exposed LDR training.

Critical weaknesses
  1. The headline comparison is confounded, and the selection objective itself is never isolated. -In Table 1, the baselines use prescribed-view LDR while Game2GS uses coverage-aware views and fixed-exposure HDR, so the 3.5–5.8 dB gain cannot be attributed to either contribution. -Table 2 isolates acquisition but compares only against prescribed street views. It omits the obvious controls at the same budget: random or uniform sampling from the same probe-derived candidate set C, uniform sampling over Ω, and prior selection methods.
  • The free-6DoF test views are drawn from the same navigation domain from which the method’s candidates are drawn, while baselines are confined to street level. A large part of the gain may simply come from covering the test distribution, not from the proposed objective.
  • None of the many selection components (F_dir vs. F_sup, λ, τ, R, L, height quotas, spatial factor s, redundancy suppression) are ablated.
  1. The radiometric ablation is unfair and internally inconsistent. -The HDR model is rendered with the recorded GT exposure, while the appearance-embedding baseline uses the mean training embedding without the standard test-time embedding optimization. This gives the proposed method oracle information that the baseline is denied. -The text says fixed-exposure LDR has slightly higher SSIM, but Table 3 shows the opposite. -The table column is labeled “Farmland” while the text says Arena, and the text refers to “Figure 1” where Fig. 3 is meant. -The runtime auto-exposure claim, a core contribution, is never evaluated quantitatively against the engine’s actual exposure behavior, and no FPS, memory, or Gaussian counts are reported despite the real-time claim.
Questions for authors
  1. What is the PSNR at budget B when views are sampled uniformly at random from the same validated candidate set C, and when only F_dir or only F_sup is used?
  2. How are free-6DoF test poses sampled? What is the minimum position and orientation distance to the selected training poses, and to the probe locations?
  3. How are LDR-trained baselines scored against HDR ground truth in Table 1? Is the GT for baselines auto-exposed LDR or μ-law-mapped HDR?
  4. Table 1 “MatrixCity” for Game2GS (25.767/0.8176/0.3315) is identical to the Table 2 entry described as “part region of Block A.” Is Table 1 also evaluated on a sub-region?
  5. In Table 2, why is dense prescribed sampling (B_high) worse than sparse (B_low) on free-6DoF? How many training runs are reported, and what is the variance?
  6. How does the method apply to engines or scenes without MagicDawn probes? What are B, the selection time, and the reconstruction cost?
11 comments
AuthorsOpenReview ↗

Response to Reviewer wNcR 1

We thank the reviewer for the detailed and constructive feedback. We appreciate the recognition of our geometry-guided acquisition strategy, free-6DoF evaluation protocol, and HDR reconstruction framework. We agree that further controlled comparisons and clearer evaluation details would strengthen the experimental evidence.

1. Acquisition and Radiometric Contributions (Weakness 1)

We agree that Table 1 presents a full-system comparison, where both camera acquisition and radiometric reconstruction differ from the baselines. Therefore, the reported 3–5 dB improvements should not be attributed solely to the proposed view-selection objective.

Our original manuscript provides two complementary controlled comparisons:

  • Table 2 isolates the effect of camera acquisition on MatrixCity Block A. Under identical fixed-exposure HDR reconstruction settings, coverage-aware acquisition improves free-6DoF PSNR over prescribed-view acquisition by 5.21 dB and 6.08 dB at the two image budgets, respectively.
  • Table 3 isolates the effect of radiometric representation on Arena under coverage-aware acquisition, using Horizon-GS as the reconstruction backbone. Fixed-exposure HDR achieves 27.790 dB, compared with 25.981 dB for fixed-exposure LDR and 23.646 dB for auto-exposed LDR.

We acknowledge that these factors were previously evaluated separately rather than through a unified controlled comparison across multiple scenes.

Experiment 1: Acquisition × Radiometric Representation Ablation

To further isolate these contributions, we have extended the controlled comparisons to Elvenruins, Rome, and Arena, using a full 2×2 design with two acquisition strategies (prescribed-view and coverage-aware) and two radiometric configurations (fixed-exposure LDR and fixed-exposure HDR).

Within each scene, we maintain the same reconstruction backbone, matched image budgets, depth-supervision settings, training schedules, and independent free-6DoF test views, varying only the factors under investigation.

Table A. Controlled acquisition × radiometric-representation comparison on independent free-6DoF test views.

SceneAcquisitionRepresentationPSNR ↑SSIM ↑LPIPS ↓
ElvenruinsPrescribedLDR19.1590.54380.4517
ElvenruinsPrescribedHDR20.4670.56330.4337
ElvenruinsCoverage-awareLDR22.1860.58360.4026
ElvenruinsCoverage-awareHDR23.2840.67280.3741
RomePrescribedLDR23.5000.69000.3850
RomePrescribedHDR24.7860.70450.3812
RomeCoverage-awareLDR25.7530.72560.3795
RomeCoverage-awareHDR26.9680.74190.3761
ArenaPrescribedLDR23.4430.66630.4522
ArenaPrescribedHDR24.6630.73670.4215
ArenaCoverage-awareLDR26.9670.78530.3769
ArenaCoverage-awareHDR28.3680.83120.3354

Results and Discussion.

The results consistently demonstrate the benefits of both components across all three scenes.

Acquisition contribution: Under identical HDR reconstruction settings, coverage-aware acquisition improves free-6DoF PSNR over prescribed-view acquisition by 2.82 dB, 2.18 dB, and 3.71 dB on Elvenruins, Rome, and Arena, respectively. This confirms that camera acquisition itself contributes substantially to the overall improvements, rather than the gains arising solely from HDR reconstruction.

Radiometric contribution: Under each fixed acquisition strategy, HDR consistently achieves higher reported PSNR than LDR across all three scenes, with differences ranging from 1.10 to 1.40 dB. We clarify the representation-specific evaluation protocols separately in our response to Q3.

The full Game2GS configuration achieves the best PSNR, SSIM, and LPIPS across all three scenes. Together with the existing results on MatrixCity Block A, these findings provide stronger cross-scene evidence for the complementary contributions of coverage-aware acquisition and HDR reconstruction.

AuthorsOpenReview ↗

Response to Reviewer wNcR 2

2. Comparison with Simpler View-Selection Strategies (Weakness 1, Q1)

We thank the reviewer for highlighting the need to further validate the effectiveness of our joint coverage objective. While Table 2 demonstrates the benefits of coverage-aware acquisition over prescribed-view sampling, we agree that comparisons with simpler selection strategies using the same candidate pool would help better isolate the contribution of our proposed objective.

Experiment 2A: View Selection Strategy Comparison

We conducted controlled comparisons on Arena, using the same source mesh, validated candidate camera pool, image budget, fixed-exposure HDR reconstruction, training settings, and independent free-6DoF test views.

We compare our method against three simpler strategies:

  • Random probe sampling: Uniformly sampling camera poses at random from the same validated candidate pool.
  • Spatially uniform sampling (3D FPS): Selecting spatially distributed camera positions using farthest-point sampling, without explicitly optimizing surface observation quality.
  • Greedy surface coverage: Greedily maximizing the marginal gain in area-weighted surface coverage, retaining the highest-quality observation per surface without explicitly considering directional diversity or distinct-position support.

Table B. Comparison of view-selection strategies on Arena.

View Selection StrategyPSNR ↑SSIM ↑LPIPS ↓
Random probe sampling21.6360.61530.4945
Spatially uniform sampling (3D FPS)24.5580.72370.4372
Greedy surface coverage24.8630.72300.4366
Ours (Full Method)28.3680.83120.3354

Results and Discussion.

Spatially uniform sampling and greedy surface coverage achieve similar performance, with PSNR values also close to the prescribed-view HDR baseline (24.663 dB in Table A). This suggests that broader spatial sampling or maximizing single-best surface observation quality alone is insufficient for high-quality free-6DoF reconstruction.

In contrast, our full method improves PSNR by 6.73 dB, 3.81 dB, and 3.51 dB over random sampling, spatially uniform sampling, and greedy surface coverage, respectively, with consistent improvements in SSIM and LPIPS.

These results demonstrate the additional benefits of our joint selection strategy beyond broad spatial coverage alone, highlighting the importance of complementary, high-quality observations from diverse viewing directions and distinct camera positions.

AuthorsOpenReview ↗

Response to Reviewer wNcR 3

3. View-Selection Component Ablation and Hyperparameter Settings (Weakness 1, Q1)

To further identify the contributions of individual components, we conducted leave-one-component-out ablations on Arena, using the same validated candidate pool, image budget, HDR reconstruction settings, and free-6DoF test views.

Specifically, directional coverage and distinct-position support constitute our joint objective, while spatial penalties adjust the greedy selection scores. Height quotas and redundancy suppression constrain candidate eligibility.

Table C. Leave-one-component-out ablation of the coverage-aware view-selection method on Arena.

ConfigurationDirectional CoveragePosition SupportSpatial PenaltyHeight QuotasRedundancy SuppressionPSNR ↑SSIM ↑LPIPS ↓
w/o Directional Coverage—✓✓✓✓23.6360.63670.5201
w/o Position Support✓—✓✓✓25.6390.75240.4224
w/o Spatial Penalty✓✓—✓✓27.1330.80130.3824
w/o Height Quotas✓✓✓—✓25.3740.74230.4233
w/o Redundancy Suppression✓✓✓✓—28.0420.81250.3994
Ours (Full Method)✓✓✓✓✓28.3680.83120.3354

For the first two ablations, we retain only one coverage term in the joint objective (FdirF_{\mathrm{dir}} or FsupF_{\mathrm{sup}}), while keeping the other selection mechanisms unchanged.

Direct response to Q1.

Under the same image budget (BB), the random probe sampling baseline achieves 21.636 dB PSNR, while the FdirF_{\mathrm{dir}}-only and FsupF_{\mathrm{sup}}-only objective configurations achieve 25.639 dB and 23.636 dB, respectively. The full method achieves 28.368 dB.

These results directly demonstrate that neither random sampling nor either individual coverage term alone achieves the performance of our joint selection strategy.

Results and Discussion.

The ablation results demonstrate the importance of the individual components in our view-selection pipeline.

Removing directional coverage causes the largest performance degradation (−4.73 dB PSNR), highlighting the importance of observing scene surfaces from complementary viewing directions.

Removing height quotas and distinct-position support also leads to substantial drops (−2.99 dB and −2.73 dB, respectively), supporting the need for balanced height coverage and qualified observations from multiple camera positions.

Spatial penalties and redundancy suppression provide further improvements, with the latter showing a particularly noticeable effect on LPIPS.

Overall, these results support the importance of combining directional diversity, multi-position support, and balanced spatial allocation, rather than relying solely on broad surface coverage.

Hyperparameter Settings

We also clarify the values and intended functions of the parameters highlighted by the reviewer.

We use λ = 0.5, τ = 0.2, R = 4, and L = 2 across all four evaluated scenes, without scene-specific tuning.

ParameterValueFunction
λ\lambda0.5Weight of distinct-position support
τ\tau0.2Qualifying observation-quality threshold
RR4Position-support saturation target
LL2Directional-quality saturation target

Here, RR and LL define reward-saturation targets rather than hard requirements on the number of observations or guarantees of reconstruction sufficiency.

These parameters are interdependent within our joint optimization: varying one may affect the relative contributions of directional coverage and position support.

Rather than tuning them separately for individual scenes, we adopt one fixed configuration across all experiments. The favorable results across four scenes, together with our selection-strategy comparisons and component ablations, provide evidence of the effectiveness of the proposed formulation under this shared configuration.

We cannot claim that these parameter values are globally optimal or that performance is insensitive to their variations. We will clarify their settings and functions in the revised manuscript.

Reviewer_zqVYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Game2GS proposes a game-world Gaussianization pipeline that selects camera poses from engine-native lighting probes using known source geometry, then reconstructs a fixed-exposure HDR Gaussian representation. View selection rewards surface observation quality, viewing-direction diversity, and support from distinct camera positions; runtime exposure follows an engine-exported compensation curve.

Critical strengths
  1. Meaningful free-6DoF evaluation: The distinction between trajectory-near test views and independently sampled valid 6DoF viewpoints targets an important limitation of conventional large-scene reconstruction benchmarks. This evaluation protocol is a useful contribution for game worlds intended to be explored beyond the original capture paths.

  2. Evidence that camera selection matters: Under matched HDR capture, image budget and reconstruction configuration, Table 2 shows substantially better free-6DoF PSNR from coverage-aware views than prescribed street-view sampling, even though accuracy near the prescribed trajectory is lower. This result supports the paper’s acquisition motivation.

Critical weaknesses
  1. Table 1 compares prescribed-view LDR baselines with the full system, which changes both camera acquisition and radiometric reconstruction. The reported 3–5 dB gains therefore cannot be assigned to the proposed selection objective. Table 2 controls acquisition on a partial region of MatrixCity Block A, but corresponding isolation is not shown on the other main scenes.

  2. The coverage objective combines familiar visibility, projected resolution, directional diversity, spatial distribution and position-support heuristics. Although the paper discusses existing view-selection work, it does not adequately isolate whether the specific joint objective improves upon simpler choices using the same mesh and candidate pool. This limits the demonstrated methodological novelty.

Questions for authors

Q1: Can the authors reproduce Table 2-style acquisition-only comparisons on the other three scenes, with the same HDR capture, depth supervision, backbone, training budget and view count?

Q2: How does the proposed objective compare with random probe selection, spatially uniform or height-stratified sampling, and a simple greedy visible-surface/projected-resolution coverage objective using the identical camera candidates? Which terms contribute most?

Q3: Can the authors provide a joint acquisition × radiometric-representation ablation under matched training settings, separating fixed HDR supervision from geometry-guided view selection across more than the currently isolated examples?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zqVY 1

We thank the reviewer for the constructive feedback and for recognizing the value of our free-6DoF evaluation protocol and the importance of camera acquisition. We agree that further controlled experiments would strengthen the attribution of performance improvements and better establish the benefits of our joint view-selection objective.

1. Isolating Acquisition and Radiometric Contributions (Weakness 1, Q1 and Q3)

We agree that Table 1 presents a full-system comparison, where both camera acquisition and radiometric reconstruction differ from the baselines. Therefore, the reported 3–5 dB improvements should not be attributed solely to the proposed view-selection objective.

Our original manuscript already provides two complementary controlled comparisons:

  • Table 2 isolates the effect of camera acquisition on MatrixCity Block A. Under identical fixed-exposure HDR reconstruction settings, coverage-aware acquisition improves free-6DoF PSNR over prescribed-view acquisition by 5.21 dB and 6.08 dB at the two image budgets, respectively.
  • Table 3 isolates the effect of radiometric representation on Arena under coverage-aware acquisition, using Horizon-GS as the reconstruction backbone. Fixed-exposure HDR achieves 27.790 dB, compared with 25.981 dB for fixed-exposure LDR and 23.646 dB for auto-exposed LDR.

We acknowledge that these factors were previously evaluated separately rather than through a unified controlled comparison across multiple scenes.

Experiment 1: Acquisition × Radiometric Representation Ablation

To address Q1 and Q3, we have extended the controlled comparisons to Elvenruins, Rome, and Arena, using a full 2×2 design with two acquisition strategies (prescribed-view and coverage-aware) and two radiometric configurations (fixed-exposure LDR and fixed-exposure HDR).

Within each scene, we maintain the same reconstruction backbone, matched image budgets, depth-supervision settings, training schedules, and independent free-6DoF test views, varying only the factors under investigation.

Table A. Controlled acquisition × radiometric-representation comparison on independent free-6DoF test views.

SceneAcquisitionRepresentationPSNR ↑SSIM ↑LPIPS ↓
ElvenruinsPrescribedLDR19.1590.54380.4517
ElvenruinsPrescribedHDR20.4670.56330.4337
ElvenruinsCoverage-awareLDR22.1860.58360.4026
ElvenruinsCoverage-awareHDR23.2840.67280.3741
RomePrescribedLDR23.5000.69000.3850
RomePrescribedHDR24.7860.70450.3812
RomeCoverage-awareLDR25.7530.72560.3795
RomeCoverage-awareHDR26.9680.74190.3761
ArenaPrescribedLDR23.4430.66630.4522
ArenaPrescribedHDR24.6630.73670.4215
ArenaCoverage-awareLDR26.9670.78530.3769
ArenaCoverage-awareHDR28.3680.83120.3354

Results and Discussion.

The results consistently demonstrate the independent benefits of both components across all three scenes. These controlled comparisons directly address Q1 and Q3:

  • Q1 — Acquisition contribution: Under identical HDR reconstruction settings, coverage-aware acquisition improves free-6DoF PSNR over prescribed-view acquisition by 2.82 dB, 2.18 dB, and 3.71 dB on Elvenruins, Rome, and Arena, respectively. This confirms that the acquisition strategy itself contributes substantially to the overall improvements, rather than the gains arising solely from HDR reconstruction.

  • Q3 — Radiometric contribution: Under each fixed acquisition strategy, HDR consistently achieves higher PSNR than LDR across all three scenes, with improvements ranging from 1.10 to 1.40 dB.

The full Game2GS configuration achieves the best PSNR, SSIM, and LPIPS across all three scenes. Together with the existing results on MatrixCity Block A, these findings provide stronger cross-scene evidence for the complementary contributions of coverage-aware acquisition and HDR reconstruction.

AuthorsOpenReview ↗

Response to Reviewer zqVY 2

2. Evaluating the Joint Coverage Objective (Weakness 2 and Q2)

We thank the reviewer for highlighting the need to further validate the effectiveness of our joint coverage objective. While Table 2 demonstrates the benefits of coverage-aware acquisition over prescribed-view sampling, we agree that comparisons with simpler selection strategies using the same candidate pool would help better isolate the contribution of our proposed objective.

Experiment 2A: View Selection Strategy Comparison

We conducted controlled comparisons on Arena using the same source mesh, validated candidate camera pool, image budget, fixed-exposure HDR reconstruction, and independent free-6DoF test views for all methods.

We compare the proposed method against three simpler strategies:

Random probe sampling: Uniformly sampling camera poses at random from the same validated candidate pool.

Spatially uniform sampling (3D FPS): Selecting spatially distributed camera positions using farthest-point sampling, without explicitly optimizing surface observation quality.

Greedy surface coverage: Greedily maximizing the marginal gain in area-weighted surface coverage, retaining the highest-quality observation per surface without explicitly considering directional diversity or distinct-position support.

Table B. Comparison of view-selection strategies on Arena.

View Selection StrategyPSNR ↑SSIM ↑LPIPS ↓
Random probe sampling21.6360.61530.4945
Spatially uniform sampling24.5580.72370.4372
Greedy surface coverage24.8630.72300.4366
Ours (Full Method)28.3680.83120.3354

Results and Discussion.

Spatially uniform sampling and greedy surface coverage achieve similar performance, with PSNR values also close to the prescribed-view HDR baseline. This suggests that broader spatial sampling or maximizing single-best surface observation quality alone is insufficient for high-quality free-6DoF reconstruction.

In contrast, our joint coverage objective improves PSNR by 3.81 dB and 3.51 dB over spatially uniform sampling and greedy surface coverage, respectively, with substantial improvements in SSIM and LPIPS.

These results highlight the importance of not only achieving spatial and surface coverage, but also selecting complementary, high-quality observations from diverse viewing directions and distinct camera positions. The contributions of individual components are further examined in Experiment 2B.

AuthorsOpenReview ↗

Response to Reviewer zqVY 3

2. Evaluating the Joint Coverage Objective (Weakness 2 and Q2)

Experiment 2B: View-Selection Component Ablation

To further identify the contributions of individual components, we perform leave-one-component-out ablations on Arena, using the same validated candidate pool, image budget, HDR reconstruction settings, and free-6DoF test views.

Our view-selection pipeline consists of the following components:

                 Validated Camera Candidates C
                              +
                   Source Mesh Surface Elements
                              |
                              v
                 Surface Observation Quality
           (Visibility, Range, Incidence, Resolution)
                              |
                +-------------+-------------+
                |                           |
                v                           v
 (1) Directional Coverage (F_dir)   (2) Distinct-Position Support (F_sup)
                |                           |
                +-------------+-------------+
                              |
                              v
                       Joint Objective
                 J(V) = F_dir(V) + λF_sup(V)
                              |
                              v
                 Greedy Selection Iteration
                 --------------------------
                 Candidate Eligibility:
                   - Positive marginal gain
                   - (4) Height Quotas
                   - (5) Redundancy Suppression

                 Candidate Scoring:
                   - (3) Spatial Penalty
                   - Score = s(c|V) × ΔJ(c|V)
                 --------------------------
                              |
                              v
                 Select Best View & Update V
                              |
                              v
          Repeat Until Budget B or No Eligible View

Table C. Leave-one-component-out ablation of the coverage-aware view-selection method on Arena.

ConfigurationDirectional CoveragePosition SupportSpatial PenaltyHeight QuotasRedundancy SuppressionPSNR ↑SSIM ↑LPIPS ↓
w/o Directional Coverage—✓✓✓✓23.6360.63670.5201
w/o Position Support✓—✓✓✓25.6390.75240.4224
w/o Spatial Penalty✓✓—✓✓27.1330.80130.3824
w/o Height Quotas✓✓✓—✓25.3740.74230.4233
w/o Redundancy Suppression✓✓✓✓—28.0420.81250.3994
Ours (Full Method)✓✓✓✓✓28.3680.83120.3354

Results and Discussion.

The ablation results demonstrate the importance of the individual components in our view-selection pipeline. Removing directional coverage causes the largest performance degradation (−4.73 dB PSNR), highlighting the importance of observing scene surfaces from complementary viewing directions. Removing height quotas and distinct-position support also leads to substantial drops (−2.99 dB and −2.73 dB, respectively), supporting the need for balanced height coverage and qualified observations from multiple camera positions.

Spatial penalties and redundancy suppression provide further improvements, with the latter showing a particularly noticeable effect on LPIPS. Overall, these results confirm that high-quality free-6DoF reconstruction requires not only broad surface coverage, but also complementary viewing directions, multi-position support, and balanced spatial allocation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

We thank all reviewers for their constructive feedback and recognition of the value of geometry-guided acquisition and free-6DoF evaluation. We have addressed the main concerns through additional controlled experiments, component ablations, computational analysis, and methodological clarifications.

Reviewer zqVY. We conducted cross-scene 2×2 acquisition × radiometric ablations on Elvenruins, Rome, and Arena, demonstrating the independent benefits of coverage-aware acquisition under matched HDR reconstruction settings. We further compared our joint objective against random, spatially uniform, and greedy surface-coverage sampling, and evaluated the contributions of its individual components.

Reviewer 6rMG. We provided controlled view-selection comparisons and component-level ablations to establish the effectiveness of the proposed selection strategy. We additionally profiled the offline acquisition pipeline, reporting candidate generation, geometric validation, and greedy selection costs, together with runtime rendering performance.

Reviewer wNcR. Beyond the shared ablation experiments, we clarified the free-6DoF test-view distribution through training–test pose statistics, the LDR/HDR evaluation protocols, and the exposure-compensation mechanism. We also added quantitative runtime auto-exposure rendering results, clarified the MatrixCity evaluation scope and image budgets, and discussed computational costs and the applicability of probe-based acquisition.

The manuscript revisions focus on three areas:

  1. Controlled experimental validation: Cross-scene contribution isolation, simpler selection baselines, and component-level ablations.
  2. Evaluation transparency: Clearer dataset and test-view protocols, radiometric evaluation settings, runtime exposure evaluation, and correction of presentation inconsistencies.
  3. Practical efficiency and reproducibility: Acquisition time breakdown, reconstruction and rendering costs, hyperparameter settings, and implementation details.

These additions strengthen the empirical evidence for the effectiveness of Game2GS while preserving the original framework and its core methodology. We will incorporate the corresponding results and clarifications into the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.