arxiv: 2605.07181 · v1 · submitted 2026-05-08 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

SatSurfGS: Generalizable 2D Gaussian Splatting for Sparse-View Satellite Surface Reconstruction

Min Chen , Wei Guo , Bin Wang , Wen Li , Tong Fang , Jinbo Zhang , Junqi Zhao , Hong Kuang

show 4 more authors

Han Hu Xuming Ge Qing Zhu Bo Xu

Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:06 UTC · model grok-4.3

classification 💻 cs.CV

keywords satellite surface reconstruction2D Gaussian splattingsparse-view reconstructiongeneralizable 3D reconstructionconfidence-aware feature fusionmulti-view stereoheight map refinement

0 comments

The pith

SatSurfGS reconstructs satellite surfaces from sparse views by predicting 2D Gaussian attributes through three-level local geometric reliability modeling.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents SatSurfGS as a generalizable feed-forward method that reconstructs continuous surfaces from sparse satellite images using 2D Gaussian Splatting instead of per-scene optimization. It tackles the spatially uneven reliability of multi-view matches caused by photometric variations and weak textures by building a coarse-to-fine attribute prediction pipeline that explicitly tracks confidence at feature fusion, parameter refinement, and loss computation. A sympathetic reader cares because satellite surface mapping often operates under exactly these constraints yet still needs fast, accurate results that transfer across different datasets. The method claims to deliver higher rendering quality, better geometric accuracy, stronger cross-dataset generalization, and faster inference than both generalizable baselines and competitive scene-specific approaches.

Core claim

SatSurfGS constructs a coarse-to-fine 2D Gaussian Splatting pipeline for sparse-view satellite surface reconstruction that explicitly models local geometric reliability at three stages: a confidence-aware monocular multi-view feature fusion module that weights monocular priors against multi-view matches; a cross-stage self-consistency residual guidance module that stabilizes refinement using height-map residuals and confidence; and a confidence bidirectional routing loss that assigns differentiated geometric and appearance supervision. Experiments demonstrate that this yields improved rendering quality, surface reconstruction accuracy, cross-dataset generalization, and inference efficiency.

What carries the argument

The three-level confidence modeling framework that adaptively fuses monocular priors with multi-view residuals during feature learning, Gaussian parameter estimation, and bidirectional loss supervision.

If this is right

Rendering quality and geometric accuracy improve on satellite test sets relative to both generalizable and per-scene baselines.
Cross-dataset generalization holds without retraining or per-scene optimization.
Inference runs faster than methods that optimize Gaussians separately for each scene.
Sparse-view inputs become more practical for large-scale satellite surface mapping tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same reliability-modeling pattern could be tested on other sparse multi-view domains such as aerial or ground-level imagery with similar photometric inconsistencies.
If the confidence estimates prove stable, future pipelines might reduce the minimum number of required satellite views while maintaining accuracy.
Integration with streaming satellite feeds could enable near-real-time surface change detection without full re-optimization per frame.

Load-bearing premise

Local geometric reliability can be estimated reliably enough from monocular priors and multi-view residuals that the three-level modeling generalizes to new satellite datasets without introducing fresh failure modes.

What would settle it

Performance comparison on an unseen satellite dataset captured under different sensors or lighting conditions where the method shows no improvement in surface accuracy or cross-dataset transfer over strong baselines.

Figures

Figures reproduced from arXiv: 2605.07181 by Bin Wang, Bo Xu, Han Hu, Hong Kuang, Jinbo Zhang, Junqi Zhao, Min Chen, Qing Zhu, Tong Fang, Wei Guo, Wen Li, Xuming Ge.

**Figure 1.** Figure 1: Imaging challenges associated with sparse satellite views and the advantages of 2DGS for surface representation. (a) Multiview satellite images of the same area typically exhibit significant differences in phase and observation angle, resulting in marked variations in brightness and scale. (b) The surface representation of 2DGS is less prone to severe degradation than that of 3DGS, and is therefore more s… view at source ↗

**Figure 2.** Figure 2: Comparison with existing methods: PixelSplat (Charatan et al., 2024), MVSplat (Chen et al., 2024), DepthSplat (Xu et al., 2025a), HiSplat (Tang et al., 2024), Transplat (Zhang et al., 2025a), EOGS (Aira et al., 2025) and Sat-nerf (Marí et al., 2022). (a) Comparison of rendered images and estimated height maps;(b) Comparison of surface reconstruction results; (c) Generalization results on the DFC19 (Bosch e… view at source ↗

**Figure 3.** Figure 3: Overview of SatSurfGS: A Confidence-Guided Coarse-to-Fine Feed-Forward 2DGS Framework for Sparse-View Satellite Surface Reconstruction. The key designs include: (1) confidence-aware monocular–multiview feature fusion, (2) cross-stage selfcconsistency residual guidance, and (3) confidence bidirectional routing loss mechanism. 3. Methods 3.1. Overview Under sparse satellite observation conditions, scene sur… view at source ↗

**Figure 4.** Figure 4: Visualization and confidence-binned analysis of the proposed confidence. (a) Input satellite image. (b) Estimated height map. (c) Predicted confidence map. (d) Height error map. (e) Confidence-binned MAE and RMSE. (f) Confidence-binned 𝑃 𝐴𝐺2.5 and 𝑃 𝐴𝐺7.5 . The spatial comparison between (c) and (d) shows that low-confidence regions generally coincide with regions of larger geometric error. The binned stat… view at source ↗

**Figure 5.** Figure 5: The illustration of confidence-aware monocular–multiview feature fusion module. The module uses confidence-guided routing to adaptively fuse multi-view and monocular features, improving robustness in ambiguous regions. branches, and further constructs the attention logits of the two branches by incorporating confidence-dependent bias terms: 𝐿 𝑗 s,mvs = ⟨ 𝑄 𝑗 𝑠 , 𝐾𝑗 𝑠,mvs⟩ √ 𝑑 + 𝛽(𝐶𝑠−1), 𝐿 𝑗 𝑠,𝑚𝑜𝑛𝑜 = ⟨ 𝑄 𝑗 … view at source ↗

**Figure 6.** Figure 6: The illustration of cross-stage self-consistency residual guidance module. The module exploits cross-stage depth residuals and confidence to guide adaptive Gaussian refinement and improve geometric consistency across stages. Where ⊙ denotes element-wise multiplication, and 𝐹̃ 𝑗 𝑠 represents the fused feature volume. According to this formulation, 𝑉 𝑗 𝑠,𝑚𝑣𝑠 is treated as the primary feature, while the monoc… view at source ↗

**Figure 7.** Figure 7: Novel view synthesis results of different methods on the DFC19 dataset. (a) PixelSplat; (b) MVSplat; (c) DepthSplat; (d) HiSplat; (e) Transplat; (f) Ours; (g) Ground truth. scenarios, existing methods generally suffer from varying degrees of blurring, structural stretching, edge diffusion, and color drifting, which are particularly evident in slender buildings, sloped roofs, and locally occluded regions. I… view at source ↗

**Figure 8.** Figure 8: Novel-view height estimation results of different methods on the DFC19 dataset. (a) PixelSplat; (b) MVSplat; (c) DepthSplat; (d) HiSplat; (e) Transplat; (f) Ours; (g) Ground truth. local misalignment, and excessive smoothing. These issues lead to inconsistent internal heights within building regions, diffused structural contours, and even difficulty in distinguishing the true height differences between adj… view at source ↗

**Figure 9.** Figure 9: Novel view synthesis results of different methods on the MVS3D dataset. (a) PixelSplat; (b) MVSplat; (c) DepthSplat; (d) HiSplat; (e) Transplat; (f) Ours; (g) Ground truth. Overall, the cross-dataset experiments indicate that the proposed method is not limited to the statistical characteristics of the source-domain data, but instead learns more transferable geometric and appearance priors for satellite sc… view at source ↗

**Figure 10.** Figure 10: Novel-view height estimation results of different methods on the MVS3D dataset. (a) PixelSplat; (b) MVSplat; (c) DepthSplat; (d) HiSplat; (e) Transplat; (f) Ours; (g) Ground truth. feed-forward inference; rather, by learning cross-scene priors, it can directly produce competitive or even superior reconstruction results with extremely low time overhead. From the overall trend, per-scene optimization method… view at source ↗

**Figure 11.** Figure 11: Comparison of novel-view synthesis results between the proposed method and per-scene optimization methods across different scenes. (a) Sat-NeRF; (b) EOGS; (c) Ours; (d) Ground truth [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗

**Figure 12.** Figure 12: Comparison of novel-view height map estimation results between the proposed method and per-scene optimization methods across different scenes. (a) Sat-NeRF; (b) EOGS; (c) Ours; (d) Ground truth [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗

**Figure 13.** Figure 13: Qualitative comparison of mesh models extracted by different methods. (a) Sat-NeRF; (b) EOGS; (c) HiSplat; (d) Ours; (e) corresponding Image. First, when CMMF and CSRG are removed simultaneously, the model exhibits the most pronounced performance degradation, indicating that these two core modules are essential to the overall performance improvement. Furthermore, when Naive Fusion is used to replace CMMF … view at source ↗

read the original abstract

Sparse-view satellite image surface reconstruction remains highly challenging, fundamentally because the reliability of multi-view matching under satellite imaging conditions is strongly spatially heterogeneous. Affected by large photometric differences, weak textures, and repetitive textures, multi-view geometric constraints are often sparse, unevenly distributed, and locally unreliable. Although 2D Gaussian Splatting (2DGS) is more suitable than 3D Gaussian Splatting (3DGS) for the explicit representation of continuous surfaces, research on generalizable feed-forward 2DGS frameworks for sparse-view satellite surface reconstruction is still lacking. To address this issue, we propose SatSurfGS, a generalizable sparse-view surface reconstruction method for satellite imagery based on 2DGS. The proposed method builds a coarse-to-fine Gaussian attribute prediction framework and explicitly models local geometric reliability at three levels: feature learning, Gaussian parameter estimation, and training optimization. Specifically, we propose a confidence-aware monocular multi-view feature fusion module to adaptively integrate monocular priors and multi-view matching features according to local confidence; a cross-stage self-consistency residual guidance module to stabilize stage-wise Gaussian parameter refinement using the residual between the rendered height map from the previous stage and the current-stage MVS height map, together with confidence information; and a confidence bidirectional routing loss to achieve differentiated allocation of geometric and appearance supervision. Experiments on satellite datasets show that the proposed method achieves improved rendering quality, surface reconstruction accuracy, cross-dataset generalization, and inference efficiency compared with representative generalizable baselines and competitive per-scene optimization methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SatSurfGS adds three confidence modules to a generalizable 2DGS pipeline for sparse satellite views, but the abstract shows no numbers and the cross-stage MVS residual step risks feeding unreliable geometry back into the model.

read the letter

The paper's core move is a coarse-to-fine 2D Gaussian Splatting network that predicts Gaussian attributes from sparse satellite images while trying to handle the uneven reliability of multi-view matches. It introduces a monocular-multi-view feature fusion that weights inputs by local , a cross-stage residual guidance that refines parameters using the difference between the prior stage's rendered height map and the current MVS height map, and a bidirectional routing loss that splits geometric and appearance supervision. These pieces are new in combination for the satellite setting, and the authors correctly flag the photometric and texture problems that make standard 2DGS or 3DGS struggle here. That focus on explicit local reliability at feature, parameter, and loss stages is a reasonable way to adapt the method to earth-observation data where views are few and lighting varies a lot. The claimed gains in rendering quality, surface accuracy, cross-dataset generalization, and inference speed are the right metrics to track for this use case. The abstract, however, gives none of the actual numbers, baseline names, or ablation tables, so it is impossible to tell whether the improvements are meaningful or sensitive to dataset selection. The bigger structural worry is the cross-stage module itself. It computes residuals against MVS height maps produced under the same weak-texture and photometric conditions the paper says make multi-view constraints unreliable. The three-level confidence is meant to down-weight bad regions, but nothing in the abstract demonstrates that the monocular priors or bidirectional loss can separate MVS artifacts from real geometry at the scale needed for stable refinement. If the guidance mostly propagates errors, the whole pipeline could be less robust than claimed. This work is aimed at remote-sensing researchers who already use Gaussian splatting or MVS pipelines and want a feed-forward option for large-area mapping. A reader in that group could extract useful module ideas even if the experiments need tightening. The paper is coherent on its own terms and engages the literature, so it deserves a serious referee rather than a desk reject, though the authors should expect requests for full metrics, ablations, and failure-case analysis on the MVS dependency.

Referee Report

2 major / 2 minor

Summary. The paper proposes SatSurfGS, a generalizable feed-forward 2D Gaussian Splatting framework for sparse-view satellite surface reconstruction. It introduces a coarse-to-fine Gaussian attribute prediction pipeline with explicit three-level confidence modeling to handle spatially heterogeneous multi-view reliability caused by photometric differences, weak textures, and repetitive patterns. The key modules are a confidence-aware monocular multi-view feature fusion module, a cross-stage self-consistency residual guidance module that uses residuals between prior-stage rendered height maps and current MVS height maps, and a confidence bidirectional routing loss for differentiated geometric and appearance supervision. Experiments on satellite datasets are claimed to show gains in rendering quality, surface accuracy, cross-dataset generalization, and inference speed over generalizable baselines and per-scene optimization methods.

Significance. If the central claims hold under rigorous validation, the work would be significant for satellite 3D reconstruction, where sparse views and imaging artifacts make standard multi-view stereo unreliable. The shift to a generalizable 2DGS approach with multi-stage confidence modeling offers a practical alternative to slow per-scene optimization and directly targets the core problem of locally unreliable geometric constraints. Strengths include the explicit handling of monocular priors alongside multi-view features and the focus on surface representation via 2DGS.

major comments (2)

[Abstract / §3] Abstract and §3 (cross-stage self-consistency residual guidance module): the module computes residuals between the previous-stage rendered height map and the current-stage MVS height map to refine Gaussian parameters. Because MVS height maps are generated under the same photometric differences, weak/repetitive textures, and sparse correspondences that the paper identifies as making multi-view constraints locally unreliable, any systematic error in those maps risks being propagated rather than corrected. The three-level confidence modeling is intended to down-weight unreliable regions, but the manuscript provides no ablation, error propagation analysis, or visualization demonstrating that monocular priors and the bidirectional loss can reliably distinguish MVS artifacts from true geometry at the scale needed for stable refinement.
[§4] §4 (Experiments): the abstract states that the method achieves improved rendering quality, surface reconstruction accuracy, cross-dataset generalization, and inference efficiency, yet the provided description contains no numerical metrics, specific baseline implementations, ablation results on the individual confidence components, or error analysis. Without these, it is impossible to determine whether the reported gains are robust or sensitive to dataset selection and metric choice, which directly affects the load-bearing claim of superior performance.

minor comments (2)

[Abstract] The abstract would be strengthened by including at least one or two key quantitative results (e.g., PSNR or Chamfer distance deltas) to support the performance claims.
[§3] Notation for the three-level confidence (feature fusion, parameter estimation, bidirectional loss) should be introduced consistently with symbols in the method section to avoid ambiguity when reading the loss formulation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We sincerely thank the referee for the constructive and detailed feedback. We address each major comment below and describe the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract / §3] Abstract and §3 (cross-stage self-consistency residual guidance module): the module computes residuals between the previous-stage rendered height map and the current-stage MVS height map to refine Gaussian parameters. Because MVS height maps are generated under the same photometric differences, weak/repetitive textures, and sparse correspondences that the paper identifies as making multi-view constraints locally unreliable, any systematic error in those maps risks being propagated rather than corrected. The three-level confidence modeling is intended to down-weight unreliable regions, but the manuscript provides no ablation, error propagation analysis, or visualization demonstrating that monocular priors and the bidirectional loss can reliably distinguish MVS artifacts from true geometry at the scale needed for stable refinement.

Authors: We appreciate this important observation on the risk of error propagation in the cross-stage residual guidance module. The three-level confidence modeling (feature, parameter, and loss stages) together with monocular priors and the bidirectional routing loss are intended to down-weight unreliable MVS regions and provide complementary geometric cues. However, we acknowledge that the current manuscript does not include dedicated ablations, error-propagation analysis, or visualizations to empirically demonstrate stable refinement. In the revised version we will add: (i) ablation tables isolating the residual guidance module with and without confidence weighting, (ii) visualizations of per-stage confidence maps overlaid on MVS residual errors, and (iii) quantitative metrics tracking refinement stability across stages. These additions will directly address the concern. revision: yes
Referee: [§4] §4 (Experiments): the abstract states that the method achieves improved rendering quality, surface reconstruction accuracy, cross-dataset generalization, and inference efficiency, yet the provided description contains no numerical metrics, specific baseline implementations, ablation results on the individual confidence components, or error analysis. Without these, it is impossible to determine whether the reported gains are robust or sensitive to dataset selection and metric choice, which directly affects the load-bearing claim of superior performance.

Authors: We agree that detailed quantitative validation is essential. While the full manuscript contains Section 4 with comparative tables (PSNR/SSIM for rendering, surface RMSE for accuracy, runtime for efficiency, and cross-dataset results), we recognize that the current presentation lacks exhaustive per-component ablations, explicit baseline implementation details, and sensitivity/error analysis. In the revision we will expand Section 4 to include: comprehensive ablation tables for each proposed module, precise descriptions of baseline re-implementations, additional error-distribution plots, and sensitivity tests across dataset subsets and metric choices. This will make the performance claims fully substantiated and reproducible. revision: yes

Circularity Check

0 steps flagged

No significant circularity; novel architecture with external empirical validation.

full rationale

The paper introduces a new coarse-to-fine 2DGS framework with three explicitly proposed modules (confidence-aware feature fusion, cross-stage residual guidance, and bidirectional routing loss) for sparse-view satellite reconstruction. These are architectural and loss-design choices whose claimed benefits are measured via rendering and reconstruction metrics on held-out satellite datasets against independent baselines and per-scene optimizers. No equation reduces a predicted quantity to a fitted parameter defined from the same data, no load-bearing premise rests solely on self-citation, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the result. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of the proposed modules rather than on new mathematical axioms or invented physical entities. No free parameters beyond standard network weights and loss coefficients are introduced in the abstract. No new particles, forces, or dimensions are postulated.

pith-pipeline@v0.9.0 · 5610 in / 1190 out tokens · 29901 ms · 2026-05-11T01:06:13.794349+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

we propose a confidence-aware monocular multi-view feature fusion module... cross-stage self-consistency residual guidance module... confidence bidirectional routing loss
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

the cost-volume confidence is used as a proxy for local geometric reliability

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

14 extracted references · 14 canonical work pages

[1]

5470–5479

Mip-nerf 360: Unbounded anti-aliased neural radiance fields, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5470–5479. Barron,J.T.,Mildenhall,B.,Verbin,D.,Srinivasan,P.P.,Hedman,P.,2023. Zip-nerf:Anti-aliasedgrid-basedneuralradiancefields,in:Proceedings of the IEEE/CVF International Conference on Computer Vision...

work page 2023
[2]

Semantic stereo for incidental satellite images, in: 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), IEEE. pp. 1524–1532. Bosch, M., Kurtz, Z., Hagstrom, S., Brown, M.,

work page 2019
[3]

A multiple view stereo benchmark for satellite imagery, in: 2016 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), IEEE. pp. 1–9. Charatan, D., Li, S.L., Tagliasacchi, A., Sitzmann, V.,

work page 2016
[4]

21590–21599

Splatter-360: Generalizable 360 gaussian splatting for wide-baseline panoramic images, in: Proceedings of the Computer Vision and Pattern Recognition Conference, pp. 21590–21599. Cheng,S.,Xu,Z.,Zhu,S.,Li,Z.,Li,L.E.,Ramamoorthi,R.,Su,H.,2020. Deepstereousingadaptivethinvolumerepresentationwithuncertainty awareness, in: Proceedings of the IEEE/CVF conferenc...

work page 2020
[5]

ISPRS Journal of Photogrammetry and Remote Sensing 232, 124–137

3d building reconstruction from monocular remote sensing imagery via diffusion models and geometric priors. ISPRS Journal of Photogrammetry and Remote Sensing 232, 124–137. Huang,B.,Yu,Z.,Chen,A.,Geiger,A.,Gao,S.,2024. 2dgaussiansplattingforgeometricallyaccurateradiancefields,in:ACMSIGGRAPH2024 conference papers, pp. 1–11. Huang, X., Liu, X., Wan, Y., Zhe...

work page 2024
[6]

Skyfall-gs: Synthesiz- ing immersive 3d urban scenes from satellite imagery.arXiv preprint arXiv:2510.15869, 2025

Gaussian entropy fields: Driving adaptive sparsity in 3d gaussian optimization. ISPRS Journal of Photogrammetry and Remote Sensing 236, 273–285. Lee,J.Y.,Liu,Y.R.,Tsai,S.R.,Chang,W.C.,Wu,C.H.,Chan,J.,Zhao,Z.,Lin,C.H.,Liu,Y.L.,2025. Skyfall-gs:Synthesizingimmersive3durban scenes from satellite imagery. arXiv preprint arXiv:2510.15869 . Li, Z., Yao, S., Wu,...

work page arXiv 2025
[7]

ISPRSJournalofPhotogrammetryandRemoteSensing230,861–880

Ulsr-gs: Urban large-scale surface reconstructiongaussiansplattingwithmulti-viewgeometricconsistency. ISPRSJournalofPhotogrammetryandRemoteSensing230,861–880. Liu,Z.,Niu,S.,Qiu,X.,Peng,L.,Shang,Y.,Zhong,L.,Ding,C.,2026. Adifferentiablemethodfornovelviewsarimagegenerationvia3dgaussian splatting. ISPRS Journal of Photogrammetry and Remote Sensing 231, 167–1...

work page 2026
[8]

arXiv preprint arXiv:2505.22279

Learning fine-grained geometry for sparse-view splatting via cascade depth loss. arXiv preprint arXiv:2505.22279 . Marí, R., Facciolo, G., Ehret, T.,

work page arXiv
[9]

1311–1321

Sat-nerf: Learning multi-view satellite photogrammetry with transient objects and shadow modeling using rpc cameras, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 1311–1321. Marí,R.,Facciolo,G.,Ehret,T.,2023. Multi-dateearthobservationnerf:Thedetailisintheshadows,in:ProceedingsoftheIEEE/CVFConference on Compute...

work page 2023
[10]

Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians.arXiv preprint arXiv:2403.17898, 2024

Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG) 41, 1–15. Newcombe,R.A.,Izadi,S.,Hilliges,O.,Molyneaux,D.,Kim,D.,Davison,A.J.,Kohi,P.,Shotton,J.,Hodges,S.,Fitzgibbon,A.,2011.Kinectfusion: Real-timedensesurfacemappingandtracking,in:201110thIEEEinternationalsymposiumonmixedandaugmentedreality,Ieee....

work page arXiv 2011
[11]

Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction.arXiv preprint arXiv:2410.06245, 2024

Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245 . Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.,

work page arXiv
[12]

ISPRS Journal of Photogrammetry and Remote Sensing 231, 288–306

Arsgaussian: 3d gaussian splatting with lidar for aerial remote sensing novel view synthesis. ISPRS Journal of Photogrammetry and Remote Sensing 231, 288–306. Ye,Z.,Li,W.,Liu,S.,Qiao,P.,Dou,Y.,2024. Absgs:Recoveringfinedetailsin3dgaussiansplatting,in:Proceedingsofthe32ndACMinternational conference on multimedia, pp. 1053–1061. Yu, Z., Sattler, T., Geiger, A.,

work page 2024
[13]

ACM Transactions on Graphics (ToG) 43, 1–13

Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (ToG) 43, 1–13. Zhang,C.,Zou,Y.,Li,Z.,Yi,M.,Wang,H.,2025a. Transplat:Generalizable3dgaussiansplattingfromsparsemulti-viewimageswithtransformers, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 9869–9877. Zhang,K.,Snav...

work page 2019
[14]

arXiv preprint arXiv:2309.00277

Sparsesat-nerf: Dense depth supervised neural radiance fields for sparse satellite images. arXiv preprint arXiv:2309.00277 . Page 25 of 26 SatSurfGS: Generalizable 2D Gaussian Splatting for Sparse-View Satellite Surface Reconstruction Zhang, Q., Wysocki, O., Jutzi, B., 2025b. Gs4buildings: Prior-guided gaussian splatting for 3d building reconstruction. ar...

work page arXiv