Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

VAD-GS recovers missing geometry in dynamic urban scenes through voxel-based visibility reasoning and multi-view stereo densification.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Visibility-aware densification with multi-view stereo fills missing geometry in dynamic urban 3D Gaussian splatting, improving reconstruction on Waymo and nuScenes.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The paper has a sensible incremental method and one big unmeasured claim; send it to review with a request for direct geometry evaluation and artifacts. the 5 major comments →

arxiv 2510.09364 v2 pith:XJBVJZ6M submitted 2025-10-10 cs.CV

VAD-GS: Visibility-Aware Densification for 3D Gaussian Splatting in Dynamic Urban Scenes

Yikang Zhang , Rui Fan This is my paper
classification cs.CV
keywords 3D Gaussian splattingdynamic urban scene reconstructionmulti-view stereovoxel-based visibility reasoningview selectionpoint cloud densificationnovel view synthesisdepth and normal maps
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Gaussian splatting reconstructs scenes from an initial point cloud, but in driving environments sparse LiDAR and limited camera overlap leave large regions with no points, and standard densification can only clone or split existing Gaussians. This paper claims those missing structures can be actively rebuilt: a voxel-based visibility stage maps pixels to their first-visible surface voxels, a diversity-aware view selector picks the most informative camera viewpoints, and patch-matching multi-view stereo generates new point geometry that is then turned into new Gaussian primitives. Because the new geometry is derived from image consistency rather than photometric gradient propagation alone, the paper argues that occluded or incomplete surfaces no longer receive corrupted updates. The result, demonstrated on two large autonomous-driving datasets, is higher rendering quality and, more importantly, cleaner depth and normal maps for both static scenery and moving vehicles.

Core claim

The paper's central claim is that the quality ceiling of 3D Gaussian splatting in unbounded urban scenes is set by the completeness of initialization, and that this ceiling can be lifted by explicitly detecting and filling missing geometry. VAD-GS voxelizes the initial point cloud, rasterizes visible voxel surfaces with z-buffering to obtain a dense depth and index map, and compares that voxel-derived depth with the depth rendered from current Gaussians. Where the Gaussian depth is absent or too large, the region is flagged as unreliable. A diversity score then selects supporting views from different cameras and times that maximize lateral baseline and angular diversity, and a patch-matching

What carries the argument

The central object is the visible-voxel index map, a per-pixel record of the first voxel surface intersected by each ray after z-buffered rasterization of the voxelized point cloud. This map does three jobs: it provides dense geometric supervision where raw points are sparse, it maps each segmented object's pixels back to the 3D voxels that actually explain them, and it flags unreliable geometry whenever the depth rendered from Gaussian primitives is missing or significantly larger than the voxel depth. The second mechanism is the diversity-aware view-selection score, which rewards view pairs with dense, close voxel coverage, large lateral translation, small longitudinal motion, and large or

Load-bearing premise

The method assumes the initial 3D points are trustworthy enough to tell complete from incomplete geometry, and that missing surfaces are flat enough and consistently textured to be matched across cameras — objects that are entirely unobserved or that bend while moving break both assumptions.

What would settle it

Take a sequence containing a thin, high, low-texture traffic sign and a fast-moving pedestrian, with the sign absent from the LiDAR point cloud. If VAD-GS does not reconstruct the sign's planar surface (or reconstructs the pedestrian with warped, inconsistent patches and degraded depth maps relative to a method that handles non-rigid motion), then the central claim that MVS-guided densification recovers missing dynamic-scene geometry is not general.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Regions with no initial point coverage—such as a traffic sign above LiDAR range—can gain new Gaussian primitives rather than being filled by distorted background Gaussians.
  • Dynamic vehicles can be densified by multi-view stereo when their moving observations are transformed into a local rigid coordinate frame and segmentation is accurate.
  • Depth and normal supervision from MVS patches, added during optimization, suppresses floaters and produces consistently cleaner rendered geometry, not just better RGB.
  • The method's densification is governed by explicit geometric priors instead of photometric gradients, reducing the case where occluded geometry receives erroneous updates.
  • On sparser LiDAR data (32-beam scans), the geometry-recovery gains are larger than on dense scans, because the initialization gaps are more severe.

Where Pith is reading between the lines

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

  • A testable extension is to apply the same voxel-visibility flagging to non-rigid objects by decomposing them into locally rigid parts; if that works, the main limitation named in the paper would be directly addressed.
  • The view-selection score could be reused as a general sampling strategy in other multi-camera reconstruction pipelines, since it explicitly balances overlap and triangulation quality.
  • The pipeline's reliance on instance segmentation and voxel geometry suggests that errors in either—especially for transparent or reflective surfaces—will be the first failure mode to appear in wider deployment, a prediction a reader could check by stress-testing on wire fences and glass.
  • One could test whether the improved depth and normal maps translate to downstream autonomy tasks such as occupancy prediction or planning, which would make the geometric gains practically load-bearing.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes VAD-GS, a 3D Gaussian splatting framework for dynamic urban scenes. The method combines voxel-based visibility reasoning over the initialized point cloud, a diversity-aware view-selection score, and patch-match MVS reconstruction to densify Gaussians in regions where the initial geometry is incomplete or unreliable. The recovered MVS depth and normal priors are used both to initialize new Gaussian primitives and to supervise the optimization through additional geometric losses. The authors evaluate on the Waymo Open and nuScenes datasets, reporting PSNR/SSIM/LPIPS gains over prior dynamic-scene 3DGS methods and qualitatively cleaner depth and normal maps, and claim significant improvement in reconstructed geometry for both static and dynamic objects.

Significance. If the central claim is substantiated, VAD-GS would be a useful step toward making 3DGS robust to incomplete LiDAR/SfM initialization in unbounded dynamic scenes, and the extension of MVS-style reconstruction to dynamic multi-camera driving settings is a plausible and valuable direction. The paper also introduces a concrete, if heuristic, view-selection criterion and demonstrates a working integration of MVS priors into Gaussian densification. However, the strongest advertised result — improved geometric quality — is currently not directly measured, and several experimental and methodological issues leave the evidence base weaker than the claims. The contribution is potentially significant for the urban NVS community, but the manuscript needs substantial revision before the claims can be accepted at face value.

major comments (5)
  1. [Experiments — Ablation Study (Table 3)] The abstract and conclusion claim that VAD-GS 'significantly improves the quality of reconstructed geometry,' but no direct geometric metric is reported. The only quantitative support is photometric (PSNR/SSIM/LPIPS) plus qualitative depth/normal images. In Table 3, removing all geometric losses improves PSNR (24.59 vs. 24.51) and LPIPS (0.194 vs. 0.199) relative to the complete model, while the complete model wins only on PSNR* and SSIM. This means the reported photometric metrics do not track the claimed geometric improvement. Please add a direct geometry evaluation (e.g., Chamfer distance or F-score against held-out LiDAR or accumulated points, depth error against LiDAR range images) and discuss why the geometric losses do not improve photometric metrics.
  2. [MVS Reconstruction from Selected Views / Loss Function (Eq. 5)] There is a bootstrapping loop in the pipeline: patch hypotheses for MVS are initialized using 'real-world images and Gaussian-rendered results,' and the resulting MVS depth/normal maps are then used to supervise and densify the same Gaussian model through L_normal, L_hard, and L_soft. If the current Gaussian renderings contain missing geometry, floaters, or distorted surfaces, the loop can reinforce these errors rather than correct them. No experiment isolates this effect or validates the MVS output against independent geometry (e.g., held-out LiDAR scans). Please provide a non-circular validation, such as initializing MVS from a separately trained model or comparing MVS-produced depth against ground-truth range data, and quantify how much the densified points deviate from the real surfaces.
  3. [Voxel-Based Visibility Reasoning] The incomplete-region flagging mechanism compares Gaussian-rendered depth with the z-buffered voxel depth computed from the initialized point cloud. In exactly the motivating cases — objects above LiDAR range, sparse 32-beam nuScenes scans, fast-moving targets — the initial point cloud may have no voxels at all in the region that needs densification. The paper does not state how voxel-empty regions are detected or why they would be flagged for MVS reconstruction. Without such a mechanism, the method cannot 'actively evaluate structural completeness' in the hardest settings. Please clarify how the voxel index map handles uninitialized areas and provide statistics on how many MVS-densified regions had no initial voxel support.
  4. [Experiments — Table 2] Table 2 contains impossible metric values: the EmerNeRF entry for PSNR* is 0.346 and the NSG entry for LPIPS is 21.67; the MARS LPIPS value 0.905 is also implausible. These are outside the valid ranges for these metrics and strongly suggest a formatting or column-alignment error. Since Table 2 is the primary Waymo Open quantitative comparison, the current evidence for the 'consistently outperforms' claim on Waymo is unreliable. Please correct the table and re-run the comparison if needed.
  5. [Experiments and Limitations (Table 1, Limitations section)] The claim that VAD-GS improves geometry 'for both static and dynamic objects' is weakened by the paper's own limitation statement that MVS-based reconstruction assumes rigidity and 'degrades when handling deformable or non-rigid objects such as pedestrians.' The nuScenes results in Table 1 show that in several scenes, VAD-GS does not achieve the lowest LPIPS precisely because the dynamic objects are mostly pedestrians. Since pedestrians are a major component of urban dynamic scenes and nuScenes is one of only two benchmarks, the scope of the dynamic-object claim is narrower than stated. Please report object-class-conditional results or clearly restrict the claim to rigid dynamic objects such as vehicles.
minor comments (6)
  1. [Eq. (3)] The view-diversity score formula is ambiguous: the placement of N, the vector dot product, and the terms sqrt(tx^2+ty^2), |tz|, and sin(theta) lacks parentheses. Please define the dimensions of d_R and d_S and state whether higher or lower |tz| is preferred, matching the prose.
  2. [Supplementary Material] Several references are left as '(?)' placeholders, including 'prior methods such as (?)' in the supplement and '4DGS (?)' and 'SC-GS (?)' in the Limitations section. These must be completed.
  3. [Loss Function (Eq. 5)] The geometric loss weights lambda_normal, lambda_hard, and lambda_soft are said to follow Yan et al. (2024), but the actual values are not provided. Please report exact settings for reproducibility.
  4. [Ablation Study (Table 3)] The 'w/o voxel visibility reasoning' variant disables all downstream components (view selection and MVS), so it does not isolate the contribution of visibility reasoning itself. Consider an ablation that keeps view selection and MVS but removes only the visibility-based flagging, to separate the effects.
  5. [Experimental Setup] No error bars or multiple seeds are reported. With only 6 nuScenes scenes and 8 Waymo sequences, the per-scene comparisons would be more convincing with variance or paired tests. Also, the exclusion of nuScenes Scenes 02 and 07–09 should be justified as pre-registered criteria rather than post hoc selection.
  6. [Related Work / Baselines] The paper builds directly on GaussianPro and DNGaussian, but neither is included as a baseline. A comparison against these MVS/depth-guided densification methods would help isolate the contribution of the dynamic-scene components and clarify the claimed advantage.

Circularity Check

1 steps flagged

MVS geometric priors are seeded from the model's own renderings and then used to supervise the same model, so the claimed geometry recovery is partly a self-consistency loop rather than an independent measurement.

specific steps
  1. fitted input called prediction [Methodology > MVS Reconstruction from Selected Views; Eq. (5)]
    "Patch hypotheses are initialized using real-world images and Gaussian-rendered results, and are iteratively refined by propagating candidates from neighboring pixels based on the assumption of local hypothesis similarity. ... Lnormal quantifies the angular deviations between the rendered surface normals and those obtained via patch matching, and Lhard and Lsoft quantify depth errors under hard and soft Gaussian opacity settings."

    The MVS depth/normal maps are the target of the geometric losses in Eq. (5), but their patch hypotheses are initialized from the very Gaussian-rendered results that those losses are supposed to correct. Thus the 'geometric priors' are not independent of the optimized model: errors such as missing geometry, floaters, or distorted normals present in the current render can seed the MVS estimates and then be reinforced by Lnormal, Lhard, and Lsoft. The paper reports improved rendered depth/normal maps, but no held-out geometry metric (e.g., comparing the new MVS points or rendered depth against LiDAR) isolates the MVS contribution; the loop can therefore masquerade as geometry recovery even when it only re-imitates the current render.

full rationale

The paper's rendering-quality claim is grounded in held-out image comparisons against external baselines on Waymo and nuScenes, so it is not circular. The voxel-visibility reasoning and view-selection components are heuristics evaluated on public datasets and do not reduce to a fitted parameter or self-citation. However, the geometry-improvement claim is weakened by a genuine bootstrapping loop: MVS patch hypotheses are explicitly initialized using Gaussian-rendered results, and the resulting depth/normal priors are then used as supervision in Eq. (5) for the same Gaussian model. This makes the 'geometric priors' partly a function of the current model rather than an independent ground truth. The paper's own ablation (Table 3) shows that removing the geometric losses improves PSNR and LPIPS, underscoring that the photometric metrics do not directly validate the geometry claim. No load-bearing self-citation or imported uniqueness theorem was found; the circularity is partial, not total, because MVS also uses real-world images and multi-view consistency checks. Overall score 4 reflects a partially circular geometry-validation loop while acknowledging the independent photometric core.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The central claim depends on treating initial point clouds as a reliable geometric scaffold, on MVS's planarity/rigidity assumptions, and on several unreported hyperparameters (voxel size, loss weights, view count). No new physical entities are introduced. The method's own limitations confirm the fragility of these assumptions.

free parameters (6)
  • geometric loss weights λ_normal, λ_hard, λ_soft = not reported; set following Yan et al. 2024
    Control the balance of normal and depth losses in Eq. (5); no values or sensitivity analysis are given.
  • voxel resolution for visibility voxelization = not reported
    Determines the fidelity of rasterized depth/index maps used to flag incomplete geometry; critical and unreported.
  • number of supporting views k = not reported
    Subset size in the view-selection objective in supplement Eq. (1); affects MVS reliability and cost.
  • view-sampling noise scale ε = not reported
    Introduced in supplement Eq. (1) to randomize view selection; no value is given.
  • depth-comparison thresholds for flagging incomplete instances = not reported
    Determines when Gaussian-rendered depth is considered 'significantly larger' than voxel depth; central to triggering densification.
  • densification interval = every 5 complete sampling cycles
    Schedule for visibility-based densification; chosen by hand and not ablated.
axioms (6)
  • domain assumption Initialized point cloud (LiDAR/SfM) provides a sufficient base for voxelization and visibility reasoning.
    Invoked in 'Voxel-Based Visibility Reasoning'; if initial points are missing in target regions, the visibility flagging may miss or misidentify the geometry needing densification.
  • domain assumption Scene surfaces are locally piecewise planar, and photometric consistency is a valid matching signal.
    Required by the patch-matching MVS in 'MVS Reconstruction from Selected Views'; fails on non-planar, reflective, or transparent surfaces.
  • domain assumption Dynamic objects are rigid in their local coordinate frame.
    Patch matching is restricted within dynamic regions by transforming views into the object frame; the Limitations section admits pedestrians violate this assumption.
  • domain assumption Offline instance segmentation produces sufficiently accurate masks for static/dynamic separation.
    Used to extract instances and restrict MVS; the paper itself notes segmentation is 'highly sensitive to the quality of input prompts'.
  • domain assumption Z-buffered rasterization of voxelized points yields correct surface visibility and depth.
    The method rasterizes visible voxel distances/index maps to detect incomplete geometry; assumes voxel resolution and point density are adequate, which is questionable for sparse LiDAR.
  • standard math 3DGS alpha-blending and covariance parameterization are valid rendering models.
    Background rendering model from Kerbl et al. 2023, used without modification in Eq. (2).

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of VAD-GS: Visibility-Aware Densification for 3D Gaussian Splatting in Dynamic Urban Scenes." pith.science (2026). https://pith.science/paper/XJBVJZ6M

@misc{pith2026251009364,
  author       = {Pith},
  title        = {Pith review of: VAD-GS: Visibility-Aware Densification for 3D Gaussian Splatting in Dynamic Urban Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XJBVJZ6M}},
  note         = {Machine review of arXiv:2510.09364}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

3D Gaussian splatting (3DGS) has demonstrated impressive performance in synthesizing high-fidelity novel views. Nonetheless, its effectiveness critically depends on the quality of the initialized point cloud. Specifically, achieving uniform and complete point coverage over the underlying scene structure requires overlapping observation frustums, an assumption that is often violated in unbounded, dynamic urban environments. Training Gaussian models with partially initialized point clouds often leads to distortions and artifacts, as camera rays may fail to intersect valid surfaces, resulting in incorrect gradient propagation to Gaussian primitives associated with occluded or invisible geometry. Additionally, existing densification strategies simply clone and split Gaussian primitives from existing ones, incapable of reconstructing geometry from missing structures. To address these limitations, we propose VAD-GS, a 3DGS framework tailored for geometry recovery in challenging urban scenes. Our method identifies unreliable geometry structures via voxel-based visibility reasoning, selects informative supporting views through diversity-aware view selection, and recovers missing structures via multi-view stereo reconstruction. This design enables the generation of new Gaussian primitives guided by reliable geometric priors, even in regions lacking initial points. Extensive experiments on the Waymo and nuScenes datasets demonstrate that VAD-GS outperforms state-of-the-art 3DGS approaches and significantly improves the quality of reconstructed geometry for both static and dynamic objects. Our project webpage is at mias.group/VAD-GS.

Figures

Figures reproduced from arXiv: 2510.09364 by Rui Fan, Yikang Zhang.

Figure 1
Figure 1. Figure 1: A comparison between VAD-GS and Street￾Gaussians. While both methods achieve comparable render￾ing quality, VAD-GS demonstrates superior recovery of in￾complete or unreliable scene geometry, as evidenced by no￾table improvements in the rendered depth and normal maps. anisotropic 3D Gaussian primitives with learnable geometry and appearance attributes. These primitives are jointly opti￾mized to align with t… view at source ↗
Figure 2
Figure 2. Figure 2: VAD-GS pipeline. For each static or dynamic instance with incomplete geometry, VAD-GS first performs voxel-based visibility reasoning to identify a set of potential observation views. It then incrementally selects diverse supporting views to perform MVS reconstruction. The resulting geometric priors are subsequently used for Gaussian densification and optimization. Gaussian ellipsoids onto the 2D image pla… view at source ↗
Figure 3
Figure 3. Figure 3: Voxel-based visibility reasoning. (a) Red points are visible, whereas blue points, captured from other views, are invisible in the reference view. (b) The invisibility of blue points may result from occlusions or insufficient sampling rays in the reference view. (c) Rasterizing the distances and indices of visible voxels (in green) yields dense depth maps and accurate pixel-voxel mapping. to-back α-blendin… view at source ↗
Figure 4
Figure 4. Figure 4: View Selection and MVS Reconstruction. Image patches are warped across views to check the consistency of depth, normal, and color. Only consistently matched patches (in red) are considered valid for MVS reconstruction, while inconsistent ones (in blue) are discarded. The reconstructed geometry is then used to guide Gaussian densification. the existence of both dynamic objects and continuous ego￾motion. The… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparisons between VAD-GS and other SoTA approaches on the nuScenes dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative ablation study results. geometry, which effectively suppresses photometric distor￾tions in erroneously exposed Gaussians. VAD-GS also out￾performs the second-best method in terms of SSIM and LPIPS by 0.012 and 0.032, respectively, owing to its more complete and high-fidelity reconstruction of geometry and appearance, which in turn enhances photometric consis￾tency. However, these performance ga… view at source ↗
Figure 1
Figure 1. Figure 1: An illustration of multi-camera, cross-frame views. For both static and dynamic objects, informative ob￾servation views are typically captured by different cameras at different timestamps. Implementation Details An Illustration of Multi-Camera, Cross-Frame Views As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Additional qualitative results on the Waymo Open dataset. Due to the single-camera configuration, test views captured by the forward-facing camera exhibit substantial overlap with the training views. While all methods achieve high￾fidelity rendering results under this setting, such performance may not reliably indicate the quality of the underlying geometry. configuration that utilizes cameras 0, 1, and 2 … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison between VAD-GS and other SoTA methods on the Waymo Open dataset when a multi￾camera configuration is used. Given that our objective is to recover geometry in com￾plex urban scenes, the presence of walking pedestrians is inevitable. Nonetheless, these non-rigid objects violate the rigidity assumption required by MVS-based reconstruction. Future work will explore the integration of sta… view at source ↗
Figure 4
Figure 4. Figure 4: Additional qualitative ablation study results on the nuScenes dataset. The rendered RGB images, depth maps, and normal maps are visualized in (a), (b), and (c), respectively [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Uncertainty-driven 3D Gaussian Splatting Active Mapping via Anisotropic Visibility Field

    cs.CV 2026-05 unverdicted novelty 6.0

    GAVIS quantifies per-particle anisotropic visibility in 3DGS via spherical harmonics, integrates it into a Bayesian rasterizer for real-time uncertainty, and uses maximum information gain for active mapping.

Reference graph

Works this paper leans on

38 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Aan s, H.; et al. 2016. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120: 153--168

    2016

  4. [4]

    Bao, Y.; et al. 2025. 3D Gaussian splatting: Survey, technologies, challenges, and opportunities. IEEE Transactions on Circuits and Systems for Video Technology, 35: 6832--6852

    2025

  5. [5]

    T.; et al

    Barron, J. T.; et al. 2021. Mip-NeRF: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 5855--5864

    2021

  6. [6]

    T.; et al

    Barron, J. T.; et al. 2022. Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 5470--5479

    2022

  7. [7]

    Caesar, H.; et al. 2020. nuScenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11621--11631

    2020

  8. [8]

    Chen, Y.; et al. 2024. Periodic vibration Gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv:2311.18561

    Pith/arXiv arXiv 2024

  9. [9]

    Chen, Z.; et al. 2025. OmniRe: Omni urban scene reconstruction. In Proceedings of the International Conference on Learning Representations (ICLR), 85508--85527

    2025

  10. [10]

    Cheng, K.; et al. 2024. GaussianPro: 3D Gaussian splatting with progressive propagation. In Proceedings of the International Conference on Machine Learning (ICML), 8123--8140

    2024

  11. [11]

    Cheng, S.; et al. 2020. Deep stereo using adaptive thin volume representation with uncertainty awareness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2524--2534

    2020

  12. [12]

    Dai, A.; et al. 2017. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 5828--5839

    2017

  13. [13]

    Dosovitskiy, A.; et al. 2017. CARLA: An open urban driving simulator. In Conference on Robot Learning (CoRL), 1--16

    2017

  14. [14]

    Hou, Y.; Kannala, J.; and Solin, A. 2019. Multi-view stereo by temporal nonparametric fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2651--2660

    2019

  15. [15]

    Jiang, H.; et al. 2024. GeoTexDensifier: Geometry-texture-aware densification for high-quality photorealistic 3D Gaussian splatting. arXiv:2412.16809

    arXiv 2024

  16. [16]

    Kerbl, B.; et al. 2023. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4): 1--14

    2023

  17. [17]

    Kerbl, B.; et al. 2024. A hierarchical 3D Gaussian representation for real-time rendering of very large datasets. ACM Transactions on Graphics, 43(4): 1--15

    2024

  18. [18]

    Kirillov, A.; et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 4015--4026

    2023

  19. [19]

    Li, J.; et al. 2024. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20775--20785

    2024

  20. [20]

    Mildenhall, B.; et al. 2021. NeRF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  21. [21]

    M \"u ller, T.; et al. 2022. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41(4): 1--15

    2022

  22. [22]

    Ost, J.; et al. 2021. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2856--2865

    2021

  23. [23]

    Ranftl, R.; et al. 2021. Vision Transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 12179--12188

    2021

  24. [24]

    L.; and Frahm, J.-M

    Schonberger, J. L.; and Frahm, J.-M. 2016. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 4104--4113

    2016

  25. [25]

    Schops, T.; et al. 2017. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 3260--3269

    2017

  26. [26]

    Sun, P.; et al. 2020. Scalability in perception for autonomous driving: Waymo Open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2446--2454

    2020

  27. [27]

    Wang, F.; et al. 2021. PatchmatchNet: Learned multi-view patchmatch stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14194--14203

    2021

  28. [28]

    Wang, F.; et al. 2024. Learning-based Multi-View Stereo: A Survey. arXiv:2408.15235

    arXiv 2024

  29. [29]

    Wei, D.; Li, Z.; et al. 2025. Omni-Scene: Omni-Gaussian representation for ego-centric sparse-view scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 22317--22327

    2025

  30. [30]

    Wu, G.; et al. 2024. 4D Gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20310--20320

    2024

  31. [31]

    Wu, Z.; et al. 2023. MARS: An instance-aware, modular and realistic simulator for autonomous driving. In Proceedings of the CAAI International Conference on Artificial Intelligence (CICAI), 3--15. Springer

    2023

  32. [32]

    Xu, Q.; and Tao, W. 2020. Planar prior assisted patchmatch multi-view stereo. In Proceedings of the AAAI conference on artificial intelligence (AAAI), 12516--12523

    2020

  33. [33]

    Yan, Y.; et al. 2024. Street Gaussians: Modeling dynamic urban scenes with Gaussian splatting. In Proceedings of the European Conference on Computer Vision (ECCV), 156--173. Springer

    2024

  34. [34]

    Yang, J.; et al. 2022. Non-parametric depth distribution modelling based depth inference for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8626--8634

    2022

  35. [35]

    Yang, J.; et al. 2023. EmerNeRF: Emergent spatial-temporal scene decomposition via self-supervision. arXiv:2311.02077

    Pith/arXiv arXiv 2023

  36. [36]

    Yao, Y.; et al. 2018. MVSNet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), 767--783

    2018

  37. [37]

    Zhang, Z.; et al. 2024. Pixel-GS: Density control with pixel-aware gradient for 3D Gaussian splatting. In Proceedings of the European Conference on Computer Vision (ECCV), 326--342. Springer

    2024

  38. [38]

    Zhou, X.; et al. 2024. DrivingGaussian: Composite Gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 21634--21643

    2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.