acceptodds
Under review as a conference paper at ICLR 2027

OAVP-SPLAT: OBJECT-AWARE VIEW PLANNING FOR ACTIVE RECONSTRUCTION WITH SEMANTIC 3D GAUSSIAN SPLATTING

Abstract

Active reconstruction aims to recover scene geometry and semantic information through autonomous exploration. However, semantic maps reconstructed with 3D Gaussian Splatting (3DGS) can exhibit inconsistent class predictions within individual objects and gaps across otherwise continuous surfaces in semantic renderings. We propose OAVP-Splat, a framework that uses object-level observation value to guide two-stage autonomous view planning and refines semantic maps through semantic contribution reweighting. We perform object-level Gaussian clustering and estimate the value of additional observations from semantic uncertainty, temporal disagreement, and observation history. In the coverage stage, the planner builds an initial map through frontier exploration. In the refinement stage, it balances progress toward the next frontier target with the value of revisiting mapped regions and uses geometric support to locally refine viewing directions. After mapping, we reweight semantic contributions using per-Gaussian class probabilities and support from spatially connected 3D components of the same predicted class to improve semantic rendering. Experiments on Replica and Matterport3D demonstrate that OAVP-Splat improves semantic segmentation accuracy while maintaining favorable geometric accuracy and reconstruction quality, combining autonomous exploration with high-quality semantic reconstruction.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Active Reconstruction, 3D Gaussian Splatting, Autonomous View Planning, Object-Level Gaussian Clustering, Semantic Contribution Reweighting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_z9fQOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents OAVP-Splat, an active semantic reconstruction framework built on 3D Gaussian Splatting. The method extends semantic uncertainty-based view planning with object-level observation histories and an ObjectVOI score, together with a two-stage coverage/refinement planner, local pose refinement, and trajectory optimization. It further introduces a post-hoc semantic contribution reweighting scheme based on per-Gaussian response strength and connected-component support. Experiments on Replica and Matterport3D show improved semantic reconstruction over ActiveSGM while maintaining comparable geometric and rendering quality.

Critical strengths
  1. The paper addresses a meaningful limitation of entropy-based active semantic mapping.

Modeling observation history, temporal disagreement, and view novelty at the object level provides a reasonable mechanism for identifying complementary observations rather than repeatedly selecting views based only on current semantic uncertainty.


  1. The semantic reweighting component is isolated experimentally and provides consistent gains.

The fixed-map ablation shows that per-Gaussian response weighting and component support improve mIoU from 84.60 to 85.69 without changing the acquired observations or geometry.

Critical weaknesses
  1. The empirical contribution of ObjectVOI, which is the most conceptually novel component, is relatively small.

In Table 2, adding ObjectVOI improves mIoU by only 0.21, 0.22, and 0.80 points on the three reported scenes. In contrast, a substantial portion of the overall gain comes from semantic reweighting, which appears more incremental conceptually. This weakens the claim that object-aware view planning is the main source of improvement.


  1. The current planning ablations do not cleanly attribute the gains to individual planning components.

The “Frontier exploration” control simultaneously disables ObjectVOI, local pose refinement, B-spline processing, and semantic reweighting, while retaining an SGM-style score with frontier fallback. Therefore, the relatively large gap between this control and the full method cannot be directly interpreted as evidence for the proposed object-aware planner.

Questions for authors
  1. Could the authors provide a more controlled decomposition of the planning improvements, e.g. ActiveSGM-style planning -> + frontier completion -> + two-stage planning -> + ObjectVOI -> + local refinement, under the same observation budget and without semantic reweighting? This would clarify which planning component actually contributes to the performance gain.

  2. Given that ObjectVOI provides relatively modest gains in the reported ablation, while semantic reweighting contributes a larger improvement, could the authors clarify whether the main contribution should be interpreted primarily as object-aware active planning or as the combination of planning and improved semantic aggregation?

5 comments

We sincerely thank you for your constructive feedback on the step-by-step planning decomposition under Raw mode and the complementary roles of active view planning and Semantic Contribution Reweighting (SCR).

1. ObjectVOI-Guided Two-Stage Planning Without SCR (CW1 & CW2; Q1)

Table R1 reports an incremental planning ablation on room0 and room1 (equal-weight mean), under identical observation budgets in Raw mode (without SCR).

  • A (Standard view planning): Ordinary view planning without ObjectVOI guidance.
  • B (A + ObjectVOI-guided view planning): Adds ObjectVOI guidance to the view-planning stage of A.
  • C (B + Pose refinement & B-spline): Further adds geometry-guided local pose refinement and B-spline trajectory optimization.

Table R1. Incremental planning ablation under Raw mode.

ConfigurationRaw mIoU (%) ↑Raw F1 (%) ↑Step / Cumulative mIoU Gain
A: Standard view planning83.94388.011—
B: A + ObjectVOI-guided view planning86.73790.693+2.794 pp / +2.794 pp
C: B + Local pose refinement & B-spline87.00791.015+0.270 pp / +3.064 pp

Raw mIoU improves monotonically as ObjectVOI guidance and then pose/trajectory refinement are added to standard view planning, yielding total gains of +3.064 pp mIoU and +3.004 pp F1, entirely without SCR:

  1. Adding ObjectVOI-guided view planning to A (A → B): Incorporating ObjectVOI guidance into standard view planning improves Raw mIoU by +2.794 pp and F1 by +2.682 pp. ObjectVOI guides the acquisition of complementary observations for ambiguous and under-observed objects, improving mapping quality without semantic post-processing.

  2. Adding local pose refinement & B-spline to B (B → C): Geometry-guided pose refinement and trajectory optimization further improve Raw mIoU by +0.270 pp and F1 by +0.322 pp.

Summary: Under Raw mode, adding ObjectVOI-guided view planning to A improves mIoU by +2.794 pp, while pose/trajectory refinement brings the cumulative gain to +3.064 pp, independently of SCR.


2. Interpreting the Contributions of Planning and SCR(CW1 & Q2)

We applied identical SCR to ActiveSGM's final maps across all eight Replica scenes. Table R2 presents the 2 × 2 comparison (Planner × Readout Mode) under Raw and + SCR settings.

Table R2. Eight-scene Replica averages under Raw and matched SCR settings.

MethodReadout ModemIoU (%) ↑F1 (%) ↑Depth L1 (cm) ↓
ActiveSGMRaw84.10887.821.31
ActiveSGM+ SCR85.10488.881.31
OAVP-Splat (Ours)Raw84.60188.281.15
OAVP-Splat (Ours)+ SCR85.69589.451.15
  1. Consistent acquisition advantage from object-aware view planning: In Raw mode, OAVP-Splat outperforms ActiveSGM by +0.493 pp mIoU (84.601% vs. 84.108%) and +0.46 pp F1 (88.28% vs. 87.82%), while reducing Depth L1 by 12.2% (1.31 cm → 1.15 cm). With identical SCR, the lead widens to +0.591 pp mIoU (85.695% vs. 85.104%) and +0.57 pp F1 (89.45% vs. 88.88%).
  2. Complementary gain from SCR: SCR improves eight-scene mean mIoU by +0.996 pp on ActiveSGM and +1.094 pp on OAVP-Splat (+1.06 pp and +1.17 pp F1, respectively).

Summary: Our main contribution is the complementary combination of object-aware active planning and SCR. Active planning improves 3D geometry (−12.2% Depth L1) and multi-view semantic evidence (+0.493 pp Raw and +0.591 pp with SCR in mIoU over ActiveSGM), while SCR exploits 3D connected-component support to further improve semantic predictions (+1.094 pp mIoU).

Thank you very much again for your thoughtful and constructive review. We are sincerely grateful for your careful reading of the paper and for the valuable questions you raised.

We posted our rebuttal several days ago, and we would like to very respectfully follow up. As the discussion period is drawing to a close, we would be most grateful if you might have a chance to let us know whether our response has helped address your concerns.

For ease of reference, our rebuttal mainly addressed the following points:

  • Individual contributions of the planning components: We conducted an incremental planning ablation under identical observation budgets with SCR disabled. In this setting, adding ObjectVOI-guided planning improves mIoU by +2.794 pp, and further adding local pose refinement and B-spline trajectory optimization increases the cumulative improvement to +3.064 pp. We hoped this would help more clearly separate the contributions of the planning components.

  • Separating planning gains from semantic reweighting: We also compared OAVP-Splat and ActiveSGM under both Raw and identical SCR settings across all eight Replica scenes. OAVP-Splat achieves +0.493 pp mIoU without SCR and +0.591 pp with matched SCR, together with a 12.2% reduction in Depth L1 error, which we hoped would clarify that the gains are not solely explained by semantic reweighting.

  • Interpretation of the main contributions: In our rebuttal, we also clarified that we view object-aware planning and SCR as complementary. In particular, planning improves the acquired observations and reconstructed geometry, while SCR improves semantic aggregation on top of the resulting maps.

We truly appreciate the opportunity for constructive discussion, and if convenient, we would be very grateful to hear whether these clarifications have been helpful. If there are still aspects that remain unclear, we would be more than happy to further explain them while the discussion period remains open.

Thank you once again for your time and consideration.

Thank you for the additional ablations. I have two follow-up questions regarding Table R1:

  1. Could the authors report the per-scene Raw mIoU/F1 results for room0 and room1, rather than only their equal-weight mean?

  2. Could the authors clarify exactly what “A: Standard view planning” corresponds to? Is configuration A intended to be equivalent to ActiveSGM, or is it an internal baseline based on the authors’ planning framework? If it is not ActiveSGM, please explain the differences and preferably include ActiveSGM under the same room0/room1 Raw evaluation setting in Table R1 for reference.

Reviewer_1JRVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents OAVP-Splat for active reconstruction with semantic 3D Gaussian Splatting. The method maintains object-level semantic states and observation histories and uses uncertainty and temporal disagreement to estimate the value of additional observations. The authors propose two-stage planning, including a coverage stage and a refinement stage. After mapping, semantic contribution reweighting adjusts Gaussian contributions. Experiments on Replica and Matterport3D show semantic improvements with comparable reconstruction quality.

Critical strengths

I considered the problem worthwhile. Deciding where to acquire additional evidence and how to aggregate existing semantic predictions is a relevant problem under limited observation budgets.

Critical weaknesses
  1. The methodological advance over ActiveSGM appears incremental. The paper needs stronger evidence that its particular combination of temporal disagreement, observation deficit, and view novelty offers an advance.

  2. The paper explains the history cues used in ObjectVOI. Using observation history to decide which objects to revisit makes sense, but the current ablations do not show which history cues actually help. The ablation only removes the whole ObjectVOI in Table 2.

Questions for authors
  1. The paper does not compare runtime or memory usage with previous methods. Maintaining object histories, scoring additional views, optimizing trajectories, and applying semantic reweighting introduce extra computation, but the overhead is not quantified.

  2. The four costs in Eq. (7) are intuitively motivated, but the experiments do not establish the contribution of their combinations. Could the authors provide an ablation of the objective terms?

  3. How does ActiveSGM perform when equipped with the identical semantic reweighting?

5 comments

We sincerely thank the reviewer for the constructive feedback. Below, we provide finer-grained evidence on Object Value of Information (ObjectVOI) history cues, matched Semantic Contribution Reweighting (SCR), runtime/memory, and trajectory optimization.

1. ObjectVOI History-Cue Ablation (CW1 & CW2)

Table R1 ablates Eqs. (1)–(2) on room0/room1 in isolated Raw mode (without frontier/pose/B-spline/SCR), progressively adding semantic entropy (H), temporal disagreement (HT), observation deficit (HTD), and view novelty (HTDN). Differences are in percentage points (pp).

Table R1. ObjectVOI cue ablation (Raw; equal-weight scene mean).

Variantroom0 mIoU / F1room1 mIoU / F1Mean mIoU (%) ↑Mean F1 (%) ↑
H (Entropy)83.848 / 87.84788.681 / 92.73386.26490.290
HT (+ Disagreement)84.145 / 88.19188.941 / 93.00086.54390.596
HTD (+ Deficit)84.480 / 88.42389.117 / 93.11886.79990.771
HTDN (+ View novelty)85.692 / 89.84089.484 / 93.46087.58891.650

Finding: HTDN gains +1.324 pp mIoU and +1.360 pp F1 over entropy alone, showing the complementary value of instability, coverage, and viewpoint cues before SCR is applied.

2. ActiveSGM with Matched SCR (Q3)

We applied identical SCR to the final maps of both methods across all eight Replica scenes, comparing Raw and + SCR modes (Table R2).

Table R2. Eight-scene Replica averages, with and without SCR.

MethodReadout ModemIoU (%) ↑F1 (%) ↑Depth L1 (cm) ↓
ActiveSGMRaw84.10887.821.31
ActiveSGM+ SCR85.10488.881.31
OAVP-Splat (Ours)Raw84.60188.281.15
OAVP-Splat (Ours)+ SCR85.69589.451.15
  • Raw: OAVP-Splat leads by +0.493 pp mIoU / +0.46 pp F1, with 12.2% lower Depth L1.
  • Matched SCR: Both improve (ActiveSGM: +0.996 pp mIoU; ours: +1.094 pp mIoU), while our lead widens to +0.591 pp mIoU / +0.57 pp F1.

Finding: Our gains persist under identical SCR, confirming an independent benefit from object-aware planning and multi-view observations.

3. Runtime and Sampled Peak Memory (Q1)

Table R3 reports end-to-end runtime breakdowns and sampled peak VRAM for OAVP-Splat (Full) and Original ActiveSGM on the same RTX 4090 GPU (room0).

Table R3. End-to-end runtime breakdown (s) and peak GPU memory (room0).

Stage / ComponentOAVP-Splat (Full)Original ActiveSGM
Stage 0 + Stage 1 Mapping / SLAM2,019.592,142.46
Post Refinement / SLAM6,746.857,752.23
Stage 0 + Post Planning Calls145.18143.38
Stage 1 Planning Calls224.70131.01
Initial Object Map Construction111.91—
Final SCR Application88.99—
Simulation & Misc Overhead197.61193.64
Total End-to-End Runtime (s)9,534.8310,362.73
Sampled Peak VRAM (MiB / GiB)23,677 (23.12)23,675 (23.12)

Finding: Added modules are lightweight: initial object mapping takes 111.91 s (history updated during SLAM), Stage 1 planning (ObjectVOI, pose/B-spline, and evaluation) takes 224.70 s, and final SCR takes 88.99 s (no retraining). Overall, OAVP-Splat is 827.90 s (~8.0%) faster than ActiveSGM, with <0.01% extra VRAM (+2 MiB).

4. Trajectory Objective Ablation: Eq. (7) (Q2)

Eq. (7) optimizes a cubic B-spline using smoothness (JsmoothJ_{\text{smooth}}), tracking (JtrackJ_{\text{track}}), clearance (JclearanceJ_{\text{clearance}}), and orientation (JorientationJ_{\text{orientation}}). Table R4 ablates each term on room0 (Raw).

Table R4.Leave-one-out Eq. (7) ablation (room0, Raw).

VariantmIoU (%) ↑F1 (%) ↑Δ mIoU / Δ F1 vs. Full
w/o JsmoothJ_{\text{smooth}}83.46787.505−2.432 / −2.616 pp
w/o JclearanceJ_{\text{clearance}}84.17688.280−1.723 / −1.841 pp
w/o JtrackJ_{\text{track}}84.27788.285−1.622 / −1.836 pp
w/o JorientationJ_{\text{orientation}}84.55788.588−1.342 / −1.533 pp
Full Eq. (7) (Ours)85.89990.121—
  • No smoothness: mIoU and F1 decreasing by 2.432 and 2.616 pp, respectively.
  • No clearance: Removing the clearance term reduces obstacle avoidance effectiveness, decreasing mIoU and F1 by 1.723 and 1.841 pp.
  • No tracking: Anchor drift and high-frequency motion degrade mapping performance, reducing mIoU and F1 by 1.622 and 1.836 pp.
  • No orientation: Removing orientation optimization decreases mIoU and F1 by 1.342 and 1.533 pp.

Finding: Removing any individual objective term consistently degrades semantic mapping performance, confirming that all four terms contribute to the effectiveness of the proposed trajectory optimization.

Thank you very much again for your thoughtful and constructive review. We are sincerely grateful for the detailed feedback and suggestions you provided, which were very helpful in guiding our rebuttal.

We posted our rebuttal several days ago, and we would like to very respectfully follow up. As the discussion period is approaching its end, we would be most grateful if you might have a chance to let us know whether our response has helped address your concerns.

For ease of reference, our rebuttal mainly focused on the following points:

  • Contribution of historical information in ObjectVOI: We provided a progressive ablation of temporal disagreement, observation deficit, and view novelty. These results show that combining the three components improves mIoU by +1.324 pp and F1 by +1.360 pp over entropy-only scoring, which we hoped would better clarify their contributions.

  • Fair comparison with identical SCR: We applied the same semantic reweighting module to both ActiveSGM and OAVP-Splat across all eight Replica scenes. Under this matched setting, OAVP-Splat remains better than ActiveSGM both without SCR (+0.493 pp mIoU) and with identical SCR (+0.591 pp mIoU).

  • Runtime and memory overhead: We also included a runtime and GPU memory breakdown on the same RTX 4090. In this comparison, OAVP-Splat is approximately 8.0% faster end-to-end, with only 2 MiB additional sampled peak VRAM, which we hoped would help clarify the practical overhead.

  • Trajectory optimization objectives: Finally, we provided separate ablations of the smoothness, tracking, clearance, and orientation terms in Eq. (7). Removing any one of these terms leads to reduced semantic mapping performance, which we hoped would better support their individual contributions.

We truly appreciate the opportunity for constructive discussion, and if convenient, we would be very thankful to hear whether these additions and clarifications have been helpful. If any concerns remain, we would be more than happy to continue the discussion and provide any further clarification that may be useful.

Thank you once again for your time and consideration.

We sincerely thank you again for your thoughtful review and constructive suggestions.

As the discussion period is drawing to a close, we would like to very respectfully follow up on our previous response. We would be truly grateful if you could kindly let us know whether our additional experiments and clarifications have helped address your concerns. If any questions or concerns remain, we would be more than happy to provide further clarification.

Thank you once again for your time and consideration!

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

This paper proposes OAVP-Splat, an active semantic mapping method built on semantic 3D Gaussian Splatting. It scores candidate views by object-level semantic uncertainty, temporal instability, and under-observation and reweights each Gaussian’s semantic vote at readout by its amplitude and object-component support. Compared with ActiveSGM on Replica and Matterport3D, it reports modest gains in semantic segmentation with geometry and rendering quality roughly unchanged.

Critical strengths

The method achieves good semantic improvements compared with ActiveSGM on 8 Replica scenes and 4 Matterport3D scenes. And the benefit of semantic reweighting is clearly demonstrated.

Critical weaknesses
  1. The evidence for the object-aware planner’s benefit is relatively weak. Removing ObjectVOI reduces mIoU by only 0.21, 0.22, and 0.80 points on the 3 ablation scenes, while semantic reweighting contributes substantially to the full system’s gains. Applying the same reweighting to ActiveSGM’s maps would help establish how much improvement specifically requires the proposed acquisition strategy.
  2. Generalization of the semantic reweighting to independently chosen query views remains unclear. Component-support weights are computed using rendered depth at the same 36 camera poses used for evaluation. Although no evaluation RGB or GT labels are accessed, this makes the readout depend on the evaluation-view distribution. Using independently sampled probe poses and evaluating separate query views would test whether the gains persist more generally.
Questions for authors

Please see weaknesses.

4 comments

We sincerely thank you for your insightful and constructive suggestions on evaluating both methods (ours and ActiveSGM) under identical Semantic Contribution Reweighting (SCR) settings and verifying SCR generalization with independent visibility probe poses.

1. ActiveSGM with Identical Semantic Contribution Reweighting (CW1)

Following your suggestion, we applied our Semantic Contribution Reweighting (SCR) module to the final maps reconstructed by ActiveSGM across all eight Replica scenes. Table R1 compares the two acquisition strategies under Raw (without SCR) and + SCR settings. SCR is applied post-hoc after mapping and does not alter the planned trajectories or reconstructed 3D geometry.

Table R1. Eight-scene Replica averages with and without SCR.

MethodSCR SettingmIoU (%) ↑F1 (%) ↑Depth L1 (cm) ↓
ActiveSGMRaw (w/o SCR)84.10887.821.31
ActiveSGM+ SCR85.10488.881.31
OAVP-Splat (Ours)Raw (w/o SCR)84.60188.281.15
OAVP-Splat (Ours)+ SCR85.69589.451.15
  • Pure planning advantage (Raw, without SCR): OAVP-Splat outperforms ActiveSGM across the eight-scene average by +0.493 pp in mIoU (84.601% vs. 84.108%) and +0.46 pp in F1 (88.28% vs. 87.82%), while reducing 3D geometric Depth L1 error by 12.2% (1.31 cm → 1.15 cm).
  • Advantage under identical SCR: Equipping ActiveSGM with SCR improves its mIoU by +0.996 pp, while OAVP-Splat gains +1.094 pp, widening our lead over ActiveSGM to +0.591 pp in mIoU (85.695% vs. 85.104%) and +0.57 pp in F1 (89.45% vs. 88.88%).

Summary: These controlled 2 × 2 results demonstrate that OAVP-Splat's improvements stem from both ObjectVOI and SCR, rather than SCR alone. Even without SCR, our planner acquires more accurate 3D geometry (1.15 cm vs. 1.31 cm Depth L1) and stronger multi-view semantic evidence; when both methods use identical SCR, OAVP-Splat maintains a clear and even wider margin over ActiveSGM.


2. Generalization of Semantic Contribution Reweighting with Independent Probe Poses (CW2)

To verify that SCR generalizes to novel query views without using the 36 evaluation camera poses to compute component visibility counts, we tested two strictly independent visibility probe sets on the frozen final maps of four representative Replica scenes (room0, room2, office0, and office4):

  1. A36 (Executed-Trajectory Probes): We uniformly sampled 36 camera poses directly from the agent's own executed exploration trajectory during active mapping.
  2. B36 (Discovered Free-Space Probes): We randomly sampled 36 camera poses within the discovered navigable free-space voxel map.

Both A36 and B36 have zero overlap with the 36 novel evaluation query views, and all variants are evaluated on the exact same query views (Table R2).

Table R2. Four-scene equal-weight averages (room0, room2, office0, office4) across SCR visibility-probe settings.

SCR ConfigurationVisibility Probe SourcemIoU (%) ↑F1 (%) ↑Gain over Raw (mIoU / F1)
Raw (w/o SCR)None82.39286.041—
Original defaultEvaluation poses (36 poses)84.01987.746+1.627 / +1.705 pp
A36 (revised default)Executed trajectory (36 poses)83.73587.511+1.342 / +1.469 pp
B36Discovered free space (36 poses)83.48087.267+1.088 / +1.225 pp

Relative to Raw (without SCR), calibrating SCR component weights with independent executed-trajectory probes (A36) improves four-scene mean mIoU by +1.342 pp (82.392% → 83.735%) and F1 by +1.469 pp (86.041% → 87.511%), retaining 82.5% of the mIoU gain and 86.2% of the F1 gain achieved by the original-default evaluation probes. Similarly, independent free-space probes (B36) achieve substantial improvements of +1.088 pp in mIoU and +1.225 pp in F1.

Summary: These results demonstrate that Semantic Contribution Reweighting (SCR) consistently improves semantic mapping performance across all three visibility-probe sampling strategies (Original default, A36, and B36), highlighting its robustness and general applicability across different probe selection strategies.

Thank you very much again for your thoughtful review and helpful suggestions. We are sincerely grateful for the time and care you have devoted to evaluating our work.

We posted our rebuttal several days ago, and we would like to very respectfully follow up. As the discussion period is drawing to a close, we would be most grateful if you might have a chance to let us know whether our response has helped address your concerns.

For ease of reference, our rebuttal mainly addressed the following two points:

  • Object-aware planning versus semantic reweighting: We conducted a controlled comparison by applying the same Semantic Contribution Reweighting (SCR) module to both ActiveSGM and OAVP-Splat across all eight Replica scenes. Under this matched setting, OAVP-Splat still maintains an advantage of +0.493 pp in mIoU without SCR and +0.591 pp with identical SCR, together with a 12.2% lower Depth L1 error. We hoped this comparison would help clarify that the gains are not solely due to semantic reweighting.

  • Generalization of semantic reweighting: We also evaluated SCR using two independent visibility-probe strategies (A36 and B36), both with zero overlap with the evaluation views. These experiments showed consistent improvements of +1.342 pp and +1.088 pp in mIoU, respectively, which we hoped would provide additional evidence that the SCR benefit generalizes beyond evaluation-view-dependent probe settings.

We truly value the opportunity for constructive discussion, and if convenient, we would greatly appreciate hearing whether these clarifications have been helpful. If any concerns remain, we would be very happy to further clarify them while the discussion period is still open.

Thank you once again for your time and consideration.

Thanks for the additional experiments, my major concerns are addressed.

We sincerely thank the reviewer for your thoughtful questions and constructive feedback throughout the discussion. We are very glad that our additional experiments and clarifications have addressed your major concerns. We also greatly appreciate your positive reassessment of our work and your willingness to raise the score. Your valuable suggestions have helped us further strengthen our paper. Thank you again for your time, consideration, and support!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.