acceptodds
Under review as a conference paper at ICLR 2027

OUR-GS: Online UAV Reconstruction with Adaptive Gaussian Mapping

Abstract

Online UAV reconstruction provides timely 3D spatial understanding during flight, supporting perception and decision-making as UAVs operate in previously unseen environments. Recent aerial reconstruction methods increasingly build on 3D Gaussian Splatting for high-quality, large-scale reconstruction, yet remain predominantly offline and rely on pre-collected imagery and extensive scene coverage. The online setting is substantially more challenging for these methods, as reconstruction must proceed from incomplete observations over wide depth ranges and uneven multi-view evidence within a limited optimization budget. This work presents OUR-GS, an online monocular Gaussian reconstruction framework for UAV streams that adapts Gaussian map construction and refinement to viewing distance and available multi-view evidence. OUR-GS combines multi-view depth reassessment, balanced Gaussian map expansion, and tangential-radial decomposed position updates to enhance cross-frame depth consistency, representation across depth ranges, and online refinement, thereby improving online reconstruction quality. Experiments on five outdoor UAV datasets demonstrate that OUR-GS consistently improves reconstruction quality over existing methods, with particularly pronounced gains on forward-facing large-scale sequences, while maintaining the lowest GPU memory requirement. Ablation studies further validate the contribution of each component in OUR-GS. The code will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
UAV, Online reconstruction, 3D Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_E7zYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes OUR-GS, an online Gaussian Splatting framework for UAV reconstruction that improves rendering quality by addressing cross-frame depth inconsistencies, unbalanced Gaussian allocation, and inaccurate position updates in distant regions.

Critical strengths

Strong empirical improvements over recent online Gaussian mapping methods across diverse UAV datasets.

Critical weaknesses

The paper claims to improve geometric consistency, and all 3 proposed modules are closely related to geometry: MDR aims to improve cross-frame depth consistency, BGME balances Gaussian allocation across different depth ranges, and TRDU improves Gaussian position updates, especially for distant regions.

However, the paper lacks direct quantitative evaluation of geometric accuracy. The authors mainly rely on rendering quality metrics such as PSNR, SSIM, and LPIPS, but improvements in these metrics do not necessarily mean more accurate 3D reconstruction. In particular, GS can achieve better rendering quality by adjusting opacity, scale, and appearance without necessarily recovering the correct underlying geometry.

More importantly, some previous baselines, such as MODP, have already evaluated depth or geometric accuracy in their original papers. I think the authors should include similar evaluations to better support their claims about improved geometric consistency.

Questions for authors

Could the authors provide quantitative evaluations of geometric accuracy, such as depth errors or 3D reconstruction errors, especially since some existing baselines, such as MODP, have already considered geometric evaluation? This would help clarify whether the improvements actually come from more accurate geometry rather than better appearance optimization.

1 comment
AuthorsOpenReview ↗

Response to Reviewer E7zY

We are grateful for the reviewer's thoughtful feedback and helpful suggestion to evaluate geometric accuracy. We have revised the manuscript accordingly, with updates highlighted in blue.


W1: The paper claims to ... However, the paper lacks direct quantitative evaluation of geometric accuracy ...

A: Please refer to the answer for Q1.

Q1: Could the authors provide quantitative evaluations of geometric accuracy, such as depth errors or 3D reconstruction errors ...

A: Thank you for your valuable comment. We have conducted additional evaluations of geometric accuracy, including (1) geometric evaluations of baseline comparisons, and (2) geometric evaluations of component ablations. Since ground-truth geometry is typically unavailable for real-world high-altitude UAV imagery, we evaluate geometric reconstruction in the synthetic scene from HorizonGS (Sec. 4.1, Datasets), which provides the reference depth and point cloud.

For depth evaluation, we compare rendered depth maps against reference depth using AbsRel, RMSE, and Acc.@1.25\mathrm{Acc.}@1.25 (indicating the percentage of valid pixels for which the ratio of rendered depth to ground-truth depth lies between 1/1.25 and 1.25). For 3D reconstruction evaluation, we back-project rendered depths to form a point cloud and compare it with the reference point cloud using Chamfer Distance (CD) and the F1 score with a threshold of 0.05.

(1) Geometric evaluations of baseline comparisons. Results are shown in the table below.

MethodAbsRel ↓RMSE ↓Acc.@1.25\mathrm{Acc.}@1.25 ↑CD ↓F1 ↑
Splat-SLAM0.0870.0870.4270.42795.9395.93\\%0.2960.2960.3840.384
HI-SLAM20.0610.0610.5080.508\underline{98.65\\%}0.1590.1590.7370.737
MODP0.038‾\underline{0.038}0.216‾\underline{0.216}98.4798.47\\%0.141‾\underline{0.141}0.748‾\underline{0.748}
OUR-GS0.030\mathbf{0.030}0.204\mathbf{0.204}\mathbf{98.87\\%}0.131\mathbf{0.131}0.760\mathbf{0.760}

Compared with the strong baseline MODP, the results reported in the table show that OUR-GS improves AbsRel from 0.038 to 0.030 and RMSE from 0.216 to 0.204. It also reduces CD from 0.141 to 0.131 and increases F1 from 0.748 to 0.760. These geometric evaluations show that, in addition to the gains in rendering quality reported in the paper (Sec. 4.2, Table 1, Fig. 3), OUR-GS also achieves higher geometric reconstruction accuracy than prior online reconstruction methods.

(2) Geometric evaluations of component ablations. Results are shown in the table below.

VariantAbsRel ↓RMSE ↓Acc.@1.25\mathrm{Acc.}@1.25 ↑CD ↓F1 ↑
w/o MDR0.0370.0370.2950.29597.2597.25\\%0.1380.1380.752‾\underline{0.752}
w/o BGME*0.036‾\underline{0.036}0.288‾\underline{0.288}\underline{98.43\\%}0.132‾\underline{0.132}0.7460.746
w/o TRDU0.0460.0460.3490.34996.7296.72\\%0.1440.1440.7230.723
OUR-GS0.030\mathbf{0.030}0.204\mathbf{0.204}\mathbf{98.87\\%}0.131\mathbf{0.131}0.760\mathbf{0.760}

As shown in the table, removing MDR, BGME's balanced sampling, or TRDU reduces geometric reconstruction accuracy. These performance drops verify the contributions of our proposed components to geometric reconstruction: MDR reassesses depth estimates (Sec. 3.1), BGME preserves representational capacity across different depth ranges (Sec. 3.2), and TRDU enables more effective position refinement within the limited online optimization budget (Sec. 3.3).

In the revised manuscript, we introduce these geometric evaluations in Sec. 4.2 (Lines 421-424) and Appendix A.1 (Tables 7 and 8).


Reviewer_3j5TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OUR-GS proposes an online monocular UAV reconstruction framework based on 3D Gaussian Splatting. The method addresses challenges in UAV reconstruction including inconsistent depth estimation across frames, unbalanced Gaussian allocation across different depth ranges, and inefficient position updates for distant Gaussians. Experiments on five UAV datasets show improvements in rendering quality and GPU memory usage compared to existing online reconstruction methods.

Critical strengths
  1. The proposed components are reasonably motivated, and the method provides specific formulation for depth correction and gaussian allocation.

  2. The experimental results show consistent improvement across five UAV datasets in rendering quality and lower GPU memory usage. The ablation studies also show that the proposed formulas contribute substantially to reconstruction performance.

Critical weaknesses
  1. The proposed components, MDR, BGME, and TRDU, address problems that have also been studied in previous Gaussian reconstruction methods. For example, OUSA-GS[1] considers depth scale consistency and Gaussian redundancy in online UAV reconstruction, and MVS-GS[2] uses multi-view geometric consistency for depth refinement. D2GS[3] also addresses insufficient Gaussian representation in distant regions. The authors do not sufficiently analyze how these differences improve reconstruction compared to existing approaches.

  2. MDR and TRDU aim to correct Gaussian positions using multi-view observations and distance-dependent updates, but their effectiveness is mainly evaluated through rendering metrics. It is unclear whether these methods actually improve 3D geometric accuracy.

[1] Zhou, Yan, et al. "OUSA-GS: real-time structurally-consistent 3D reconstruction of outdoor unbounded scenes via adaptive Gaussian splatting." Knowledge-Based Systems 331 (2026): 114834.

[2] Lee, Byeonggwon, et al. "Mvs-gs: High-quality 3d gaussian splatting mapping via online multi-view stereo." IEEE Access 13 (2025): 111441-111453.

[3] Song, Meixi, et al. "D 2^ 2 GS: Depth-and-Density Guided Gaussian Splatting for Stable and Accurate Sparse-View Reconstruction." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors
  1. How do the proposed MDR and BGME differ from related methods such as OUSA-GS and MVS-GS, and how much improvement comes from these specific formulations?
  2. Can the authors provide geometric accuracy measurements to evaluate whether MDR and TRDU improve 3D reconstruction?
  3. How does TRDU compared with simpler distant-dependent update strategies under the same optimization budget?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 3j5T 1/3

We sincerely thank the reviewer for the careful evaluation and constructive feedback. We provide our point-by-point responses below. All corresponding revisions have been incorporated into the revised paper and highlighted in blue.


W1: The proposed components, MDR, BGME, and TRDU, address problems that have also been studied in previous Gaussian reconstruction methods ... The authors do not sufficiently analyze how these differences improve reconstruction compared to existing approaches.

A: (Part 1/2) Thank you for raising this valuable question. We would like to clarify that OUR-GS accounts for and addresses an important yet underexplored issue in online UAV reconstruction: the substantial differences in available parallax evidence between nearby and distant regions during outdoor UAV flight (Sec. 1, Fig. 1). It uses viewing distance and available multi-view evidence to guide depth reassessment, Gaussian admission, and position refinement throughout online UAV reconstruction (Sec. 1, Sec. 3). Here, we first discuss how OUR-GS differs from the mentioned related methods, and then present experiments to demonstrate the improvements of our strategies over those used in prior work.

(1) Differences from related methods.

First, regarding depth refinement, OUSA-GS uses sparse depths from ORB-SLAM3 to fit a 3D scale grid for local calibration of monocular depth predictions. MVS-GS estimates depth through multi-view stereo, and then uses geometric relationships between different views to refine depth and remove incorrect depth estimates. Our proposed MDR instead uses depths triangulated from multi-frame matches as references and jointly estimates local depth correction fields for the frames (Sec. 3.1, Eqs. 1-4). More importantly, MDR explicitly propagates depth corrections updated with new observations to the corresponding existing Gaussians, allowing depth errors from their initialization to be corrected as new evidence becomes available (Sec. 3.1, Eq. 5).

Second, regarding Gaussian redundancy, OUSA-GS employs 3D-voxel-based primitive management to suppress redundant Gaussian initialization. Our BGME goes a step further by considering that, when 3D clues are weak, redundancy checks in 3D space may mistake depth errors for new structures (Sec. 3.2, Lines 266-267). It therefore designs separate branches for 3D and 2D redundancy checks. Weakly constrained distant candidates are deduplicated using bins in the current image space and log depth, reducing the impact of distant depth uncertainty on 3D redundancy detection (Sec. 3.2, Redundancy Checks).

Third, regarding distant Gaussian representation, D²GS alleviates far-field underfitting through additional supervision in distant regions and depth-aware Gaussian dropout, improving Gaussian representation through offline optimization. By contrast, our BGME directly balances the Gaussian initialization budget across different depth regions during map expansion, promoting more balanced representation under a limited online optimization budget without requiring additional far-field supervision (Sec. 3.2, Candidate Allocation).

In the revised manuscript, we expand the discussions of these related works and clarify how OUR-GS differs from these methods (Sec. 2, Lines 118-120; Sec. 3, Lines 239-241 and 264-266).

AuthorsOpenReview ↗

Response to Reviewer 3j5T 2/3

A: (Part 2/2, continued for W1)

(2) Improvements from our specific formulations.

To evaluate the benefits of our designs, we integrate the corresponding strategies from prior work into our online UAV reconstruction pipeline and conduct four component ablations:

  1. Replacing MDR with OUSA-GS's depth scale correction (MDR → OUSA-GS's TGSC): We use a 3D scale grid to align DepthCov predictions with sparse multi-view depths through locally interpolated scales, as described in OUSA-GS.
  2. Replacing BGME's redundancy handling with OUSA-GS's voxel filtering (BGME# → OUSA-GS's CAGP): We perform redundancy checks in occupied 3D voxels and initialize one anchor at the center of each newly occupied voxel, as described in OUSA-GS, while retaining our balanced candidate allocation.
  3. Replacing MDR with MVS-GS's depth refinement (MDR → MVS-GS's DR): We fuse depth predictions of neighboring keyframes and filter geometrically inconsistent estimates using relative depth differences after projection, as described in MVS-GS.
  4. Introducing D²GS's distant-region RGB supervision (OUR-GS + D²GS's DAFE): We add RGB L1 supervision to distant regions selected using Depth Anything V2, as described in D²GS, and apply this supervision during online reconstruction.

As OUSA-GS’s official code is not currently available, we reproduce its strategies based on the paper for experiments 1 and 2. Experiments 3 and 4 are adapted from the open-source implementations of MVS-GS and D²GS. The table below reports the ablation results on the 360DVO dataset.

VariantPSNR ↑SSIM ↑LPIPS ↓FPS ↑VRAM ↓
MDR → OUSA-GS's TGSC26.2726.270.8510.8510.2330.2331.52\mathbf{1.52}6.92‾\underline{6.92} G
BGME# → OUSA-GS's CAGP26.7626.760.8440.8440.2230.2231.381.387.097.09 G
MDR → MVS-GS's DR23.0923.090.8410.8410.3290.3291.241.245.47\mathbf{5.47} G
OUR-GS + D²GS's DAFE27.10‾\underline{27.10}0.862‾\underline{0.862}0.206‾\underline{0.206}1.351.357.247.24 G
OUR-GS27.28\mathbf{27.28}0.865\mathbf{0.865}0.202\mathbf{0.202}1.39‾\underline{1.39}7.057.05 G

As shown in the table, (i) regarding depth refinement, MDR improves PSNR by 1.01 dB over OUSA-GS's depth scale correction and 4.19 dB over MVS-GS's depth refinement; (ii) regarding Gaussian redundancy, BGME's redundancy handling improves PSNR by 0.52 dB over OUSA-GS's voxel filtering; (iii) regarding distant Gaussian representation, adding D²GS's distant-region RGB supervision decreases PSNR by 0.18 dB. These results support the benefits of our specific depth reassessment and redundancy handling designs over the strategies adapted from prior methods, without requiring additional distant-region supervision.

In the revised manuscript, these additional ablations comparing our designs with strategies adapted from existing methods are included in Appendix A.3.

AuthorsOpenReview ↗

Response to Reviewer 3j5T 3/3


W2: MDR and TRDU aim to correct ... It is unclear whether these methods actually improve 3D geometric accuracy.

A: Thank you for this valuable comment. Here, we conduct ablation experiments to evaluate the effects of MDR and TRDU on geometric accuracy. Since real-world high-altitude UAV image streams typically do not provide ground-truth geometry, we conduct additional geometric accuracy evaluations on the synthetic aerial sequence from the HorizonGS dataset (Sec. 4.1, Datasets), which provides ground-truth depth. Specifically, we evaluated depth maps rendered from the online reconstructions of Full OUR-GS, w/o MDR and w/o TRDU, comparing them against the ground-truth depth using three metrics including Absolute Relative Error (AbsRel), Root Mean Squared Error (RMSE), and the fraction within a depth-ratio threshold of 1.25 (Acc.@1.25\mathrm{Acc.}@1.25).

The following table presents the geometric accuracy ablation results.

VariantAbsRel ↓RMSE ↓Acc.@1.25\mathrm{Acc.}@1.25 ↑
w/o MDR0.037‾\underline{0.037}0.295‾\underline{0.295}\underline{97.25\\%}
w/o TRDU0.0460.0460.3490.34996.7296.72\\%
Full OUR-GS0.030\mathbf{0.030}0.204\mathbf{0.204}\mathbf{98.87\\%}

Removing MDR increases AbsRel from 0.030 to 0.037 and RMSE from 0.204 to 0.295, while removing TRDU raises them to 0.046 and 0.349, respectively. These results demonstrate that both the proposed MDR and TRDU improve the geometric accuracy of online reconstruction.

In the revised version, we add these ablations on geometric accuracy in Appendix A.1, Table 8.


Q1: How do the proposed MDR and BGME differ from related methods such as OUSA-GS and MVS-GS, and how much improvement comes from these specific formulations?

A: Please refer to the answer for W1.


Q2: Can the authors provide geometric accuracy measurements to evaluate whether MDR and TRDU improve 3D reconstruction?

A: Please refer to the answer for W2.

Q3: How does TRDU compare with simpler distance-dependent update strategies under the same optimization budget?

A: Sorry for this confusion. We test two simpler distance-dependent update strategies, including (1) Distance scaling only: directly scaling the position displacement produced by the optimizer according to viewing distance, without treating the radial and tangential components differently (i.e., gi=1g_i=1 in Eq. 11); and (2) Radial damping only: only attenuating the radial displacement component after decomposition (i.e., γi=1\gamma_i=1 in Eq. 11). Together with the ablation of w/o TRDU (Sec. 4.3), which directly applies the Adam updates from the optimization, these variants are evaluated on 360DVO as described in Sec. 4.3 under the same optimization budget. The results are shown in the following table.

VariantPSNR ↑SSIM ↑LPIPS ↓FPS ↑VRAM ↓
Distance scaling only (gi=1g_i=1)27.13‾\underline{27.13}0.860‾\underline{0.860}0.209‾\underline{0.209}1.51‾\underline{1.51}6.956.95 G
Radial damping only (γi=1\gamma_i=1)26.0726.070.8460.8460.2140.2141.441.446.89‾\underline{6.89} G
w/o TRDU26.1126.110.8520.8520.2130.2131.58\mathbf{1.58}6.80\mathbf{6.80} G
OUR-GS27.28\mathbf{27.28}0.865\mathbf{0.865}0.202\mathbf{0.202}1.391.397.057.05 G

Distance scaling alone improves PSNR from 26.11 to 27.13 dB, while incorporating radial attenuation further increases it to 27.28 dB. These results show that our TRDU improves reconstruction quality compared with the simpler distance-dependent update strategies, verifying the effectiveness of our update formulation based on tangential–radial decomposition.

In the revised version, these ablations of TRDU's scaling strategies are included in Sec. 4.3, Table 6, Lines 502-513.


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

OUR-GS proposes an online Gaussian mapper for monocular UAV video, targeting large depth ranges and limited parallax. Building on existing tracking and depth estimation, it corrects depth and existing map geometry (MDR), balances Gaussian allocation across depths while filtering redundancy (BGME), and adapts position updates for distant structure (TRDU). Evaluation on five UAV datasets reports 0.85-6.65 dB PSNR gains over MODP and the lowest GPU memory use among compared methods. Component ablations and comparisons with identical inputs from three tracking systems further support the mapping contributions.

Critical strengths
  1. Substantial mapping gains with low memory use. OUR-GS improves over MODP by 0.85-6.65 dB across five UAV datasets while using the least VRAM (Tables 1-2). On GaussianPro, the gains persist with identical poses and sparse points under three tracking systems (Table 6). This is the strongest reason for acceptance: the method improves a practical online mapping system, with evidence that the gains come from the mapper.

  2. A coherent geometric design with meaningful component evidence. MDR, BGME and TRDU address depth inconsistency, uneven allocation and distance-dependent refinement. Removing MDR or TRDU costs 1.39 or 1.17 dB on 360DVO (Table 4); redundancy handling reduces Gaussians and memory at comparable quality (Table 5). These results support the technical choices behind the system's gains.

Critical weaknesses
  1. The depth-adaptive explanation is not quantitatively established across the settings with the largest gains. Figure 4 and Appendix A provide qualitative support, but quantitative component ablations cover only 360DVO and aggregate across depth. TRDU's scaling and radial damping are also removed together. This limits what readers learn about when the proposed mechanisms matter, and is the main reason I stop at marginal acceptance rather than a stronger recommendation.

  2. Uncertainty and reproducibility are incompletely reported. Table 1's error bars are undefined and absent from the component ablations. Code release is promised, but Wild-UAV release is unspecified. These gaps limit confidence in the quantitative comparisons and independent verification of one of the largest gains, although four public data sources provide separate supporting evidence.

Questions for authors

Clarification (Weakness 2): What do the error bars measure, over how many runs and which sources of variation? Will the Wild-UAV evaluation data be released? This would clarify the reliability of the reported gains and whether the Wild-UAV results can be independently reproduced. Additional evidence (Weakness 1): Can you substantiate how the component benefits vary with depth on one dataset with the largest gains? This would clarify when the proposed depth-adaptive mechanisms matter, particularly on datasets with the largest gains.

4 comments
AuthorsOpenReview ↗

Response to Reviewer VXjH 1/2

We sincerely thank the reviewer for the careful evaluation and constructive suggestions. We provide our point-by-point responses below and have revised the manuscript accordingly (please see the updated PDF with changes highlighted in blue).


W1: The depth-adaptive explanation is not quantitatively established across the settings with the largest gains ... TRDU's scaling and radial damping are also removed together ...

A: Thank you for this insightful comment. We extend the ablation studies to PanoAir (Sec. 4.3, Appendix A.2) and Wild-UAV, on which OUR-GS respectively achieves PSNR improvements of 6.65 dB and 4.55 dB over the strong baseline MODP (Sec. 4.2, Table 1), to show how these contributions vary across depth. We also ablate TRDU's scaling and radial damping separately.

(1) Component benefits across depth.

Since ground-truth depth is unavailable for the real-captured high-altitude UAV imagery in PanoAir and Wild-UAV, we use a pretrained Depth Anything V2 model to define the depth regions for evaluation. For each RGB image, we exclude sky pixels from evaluation to prevent them from inflating far-region PSNR. Then we sort all non-sky pixels by their estimated depth and assign the nearest third to the near region, the middle third to the middle region, and the farthest third to the far region. The partition is shared by all variants. We compute PSNR within each region for each frame and average it across frames in the following ablation table, and overall PSNR is computed separately over all non-sky regions.

VariantPanoAir\qquadWild-UAV
NearMiddleFarOverallNearMiddleFarOverall
w/o MDR30.6130.6129.66‾\underline{29.66}24.63‾\underline{24.63}27.10‾\underline{27.10}27.2027.2027.87‾\underline{27.87}26.5026.5026.7926.79
w/o BGME*30.68\mathbf{30.68}29.5829.5824.5224.5227.0627.0627.28\mathbf{27.28}27.7327.7326.62‾\underline{26.62}26.85‾\underline{26.85}
w/o TRDU30.4530.4528.7528.7523.2023.2026.0426.0427.1027.1027.4127.4125.9825.9826.3226.32
OUR-GS30.65‾\underline{30.65}29.72\mathbf{29.72}24.97\mathbf{24.97}27.46\mathbf{27.46}27.21‾\underline{27.21}27.99\mathbf{27.99}27.11\mathbf{27.11}27.27\mathbf{27.27}

From the results across depth regions, we observe that (i) MDR benefits the overall reconstruction; (ii) BGME's balanced sampling addresses the tendency of Gaussian representation to be overly concentrated in nearby regions; and (iii) the benefit of TRDU is concentrated in more distant regions. Compared with w/o TRDU, OUR-GS improves PSNR by 0.20, 0.97, and 1.77 dB in the near, middle, and far regions on PanoAir, respectively. On Wild-UAV, the corresponding PSNR changes are 0.11, 0.58, and 1.13 dB. These increasing gains with depth support the motivation for compensating position updates at greater viewing distances.

(2) Contributions of distance scaling and radial damping.

The following table also ablates TRDU's scaling and radial damping and shows their benefits across depth.

VariantPanoAir\qquadWild-UAV
NearMiddleFarOverallNearMiddleFarOverall
w/o TRDU30.4530.4528.7528.7523.2023.2026.0426.0427.1027.1027.4127.4125.9825.9826.3226.32
Distance scaling only (gi=1g_i = 1)30.70\mathbf{30.70}29.49‾\underline{29.49}24.82‾\underline{24.82}27.28‾\underline{27.28}27.14‾\underline{27.14}27.85‾\underline{27.85}26.83‾\underline{26.83}27.00‾\underline{27.00}
Radial damping only (γi=1\gamma_i = 1)30.5230.5228.5628.5623.0123.0125.9525.9527.0727.0727.3027.3025.8425.8426.2926.29
OUR-GS30.65‾\underline{30.65}29.72\mathbf{29.72}24.97\mathbf{24.97}27.46\mathbf{27.46}27.21\mathbf{27.21}27.99\mathbf{27.99}27.11\mathbf{27.11}27.27\mathbf{27.27}

The results show that distance scaling contributes most of TRDU's improvement (far region from 23.20 to 24.82 dB and overall from 26.04 to 27.28 dB on PanoAir), while combining it with radial damping further yields better far-region (24.82 to 24.97 dB on PanoAir) and overall (27.28 to 27.46 dB on PanoAir) rendering quality. The same trend holds on Wild-UAV, verifying the benefit of our TRDU formulation.

In the revised version, we add the ablations of TRDU's scaling and radial damping and evaluations of component benefits across depth in Sec. 4.3 (Scaling Strategies in TRDU) and Appendix A.2.


AuthorsOpenReview ↗

Response to Reviewer VXjH 2/2

W2: Uncertainty and reproducibility are incompletely reported. Table 1's error bars are undefined and absent from the component ablations. Code release is promised, but Wild-UAV release is unspecified.

A: Sorry for the confusions.

(1) The error bars in Table 1 represent the standard deviations of OUR-GS results across five independent runs with different random seeds under identical experimental settings (Sec. 4.2, Lines 363-364), and the reported OUR-GS results are the means of the five runs. We have also added error bars to the ablation results in the revised manuscript (Sec. 4.3, Tables 4, 5, 6).

(2) The Wild-UAV dataset will be publicly released to support reproduction of the reported evaluation and to support future research on online reconstruction for UAVs, as discussed in Sec. 4.1 (Line 350).


Q1: What do the error bars measure, over how many runs and which sources of variation?

A: Please refer to the answer for W2, (1).


Q2: Will the Wild-UAV evaluation data be released?

A: Please refer to the answer for W2, (2).


Q3: Can you substantiate how the component benefits vary with depth on one dataset with the largest gains?

A: Please refer to the answer for W1, (1).


Thank you for the additional ablations and clarifications. They answer my question about how TRDU's gains vary with depth and clarify the contribution of distance scaling.

Thank you again for your feedback and kind review. We are glad that our responses and the revised paper have addressed your concerns. We truly appreciate your constructive suggestions, which have helped us refine this work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviews and Responses

We sincerely thank the Area Chair and all reviewers for their time and insightful feedback. In response, we have provided point-by-point clarifications and revised the manuscript accordingly.


Summary of This Work

This work addresses the challenges of wide depth ranges and limited parallax in online UAV reconstruction. It presents OUR-GS, an online monocular Gaussian reconstruction framework that adapts 3DGS map construction and refinement to viewing distance and available multi-view evidence. OUR-GS reassesses depths and revises existing Gaussians; balances Gaussian allocation across depth ranges while filtering redundancy; and rescales position updates and constrains radial displacements. Experiments across five UAV datasets demonstrate improved reconstruction and rendering quality with lower GPU memory usage.


Summary of Reviewers’ Comments and Our Responses

All three reviewers give a positive rating of 3. They acknowledge the motivation, technical design, and effectiveness of this paper. Their comments highlight its coherent geometric design and meaningful component evidence (Reviewer VXjH), reasonable motivation of the proposed components (Reviewer 3j5T), and strong empirical improvements across diverse UAV datasets (Reviewer VXjH, 3j5T, and E7zY). In addition, their concerns and our responses are summarized below.

Reviewer VXjH raises concerns about (i) the component benefits across depth; (ii) the individual contributions of TRDU’s distance scaling and radial damping; (iii) the meaning of the error bars and the public release of Wild-UAV. In response, we (i) add near-, middle-, and far-region component evaluations on PanoAir and Wild-UAV (Appendix A.2); (ii) separately ablate distance scaling and radial damping in TRDU (Sec. 4.3, Table 6, Appendix A.2); and (iii) clarify the error bars and specify the planned release of Wild-UAV (Sec. 4.1, Sec. 4.2, Sec. 4.3).

Reviewer 3j5T raises concerns about (i) the differences from existing methods and improvements from our specific formulations; (ii) the effects of MDR and TRDU on geometric accuracy; and (iii) TRDU’s benefits over simpler distance-dependent updates. In response, we (i) clarify the distinctions and present four component ablations integrating strategies from previous methods into our online pipeline (Sec. 2, Sec. 3, Appendix A.3); (ii) provide geometric ablations of MDR and TRDU (Appendix A.1); and (iii) compare TRDU with simpler update strategies including distance scaling alone and radial damping alone (Sec. 4.3, Table 6).

Reviewer E7zY raises concerns about quantitative evaluation of geometric accuracy. In response, we introduce geometric evaluations of both baseline comparisons and component ablations on HorizonGS, measuring depth errors and 3D reconstruction errors (Sec. 4.2, Appendix A.1).


Detailed explanations are provided in the point-by-point responses, and all corresponding revisions are included in the updated manuscript. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.