Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Adaptive per-anchor feature aggregation improves monocular UAV 4D reconstruction without giving up real-time rendering.

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 ↗

Anchor-conditioned adaptive aggregation of shared spatiotemporal feature planes improves monocular UAV dynamic Gaussian reconstruction quality while keeping real-time rendering.

T0 review reviewed 2026-07-31 challenge →

load-bearing objection Solid incremental UAV dynamic-GS paper: ACFA is a clean, ablated fix for fixed multi-plane aggregation; cross-method gains are real but not fully capacity-matched. the 3 major comments →

arxiv 2607.28320 v1 pith:CN3K3H2H submitted 2026-07-30 cs.CV

AdaAnchor4D: Anchor-Conditioned Spatiotemporal Feature Aggregation for Monocular UAV 4D Reconstruction

Peiyi Xu , Junpeng Zhang , Guanbin Li , Ronghua Shang , Mingtao Feng , Le Dong , Weisheng Dong , Guangming Shi
show 1 more author
Jie Feng
This is my paper
classification cs.CV
keywords dynamic Gaussian splattingmonocular UAV reconstructionspatiotemporal feature planesanchor-based representation4D novel view synthesisadaptive feature aggregation
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

Monocular drone video of cities is full of uneven motion: some areas stay still, others move only briefly, and some keep changing in complicated ways. Standard dynamic Gaussian methods share one set of spatiotemporal feature planes and combine them the same way everywhere, which often leaves ghosting and blur on those uneven regions. AdaAnchor4D keeps the shared planes for efficiency but lets each local anchor reweight those planes using its own learned embedding plus time, so different places and moments pull different dynamic signals from the same compact field. It also splits anchor-level motion from fine Gaussian geometry updates and warps query coordinates to match how anchors cluster in space. On three UAV datasets the method raises novel-view quality over several strong dynamic Gaussian baselines while still rendering in real time.

Core claim

The paper claims that fixed plane-wise combination of shared spatiotemporal features is the main bottleneck for heterogeneous UAV dynamics, and that conditioning plane aggregation on per-anchor embeddings and time—together with decoupled local geometry deformation and density-adaptive coordinate warping—yields higher rendering quality than representative dynamic Gaussian methods on UAV-Arc4D, VisDrone, and UAVDT while preserving real-time FPS and a compact neural representation.

What carries the argument

Anchor-Conditioned Feature Aggregation (ACFA): each anchor carries a learnable aggregation embedding that, with normalized time, drives a lightweight predictor of channel-wise softmax weights over the six shared feature planes, so the same shared field is read differently per local unit and time step.

Load-bearing premise

Camera poses and the initial sparse point cloud from a full-sequence reconstruction are treated as fixed, trustworthy inputs; if those poses are badly wrong, the reported gains may not isolate the new representation.

What would settle it

On the same UAV sequences and splits, replace ACFA’s learned plane weights with the fixed product (or mean) aggregation used by the baseline while holding poses, initialization, and training budget fixed: if PSNR/SSIM/LPIPS no longer beat the listed dynamic Gaussian methods, the central claim fails.

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

If this is right

  • Shared multi-plane spatiotemporal fields can stay compact for wide-area UAV scenes if aggregation, not the field itself, is made local and time-dependent.
  • Splitting anchor-state deformation from intra-anchor Gaussian geometry should reduce blur and ghosting on moving vehicles and motion boundaries in aerial video.
  • Warping feature-query coordinates to axis-wise anchor density should better spend limited grid capacity on non-uniform urban sampling.
  • Real-time novel-view quality on monocular UAV traffic and urban road scenes can improve without large neural storage or offline-only rendering.

Where Pith is reading between the lines

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

  • The same conditioned-readout idea may transfer to other wide-area dynamic settings where motion is sparse and region-dependent, not only UAV video.
  • Joint pose refinement with ACFA-style aggregation is a natural next stress test, given the paper’s own note on pose sensitivity.
  • Visualizing learned plane weights could become a practical diagnostic for where a scene is static versus intermittently dynamic.
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

3 major / 5 minor

Summary. AdaAnchor4D is an anchor-based canonical-space deformation framework for monocular UAV dynamic novel-view synthesis. It keeps a compact multi-plane shared spatiotemporal field but replaces fixed plane aggregation with Anchor-Conditioned Feature Aggregation (ACFA), which predicts channel-wise plane weights from a per-anchor embedding and time (Eqs. 4–6). Decoupled Local Geometry Deformation (DLGD) separates ACFA-driven anchor position/feature residuals from dedicated offset/scale residual branches (Eqs. 7–10), and Density-Adaptive Coordinate Warping (DACW) freezes axis-wise nonlinear query maps from coarse-stage anchor histograms (Eqs. 11–14). On UAV-Arc4D, VisDrone, and UAVDT, the method reports higher PSNR/SSIM and lower LPIPS than D3DGS, 4D-GS, MoDec-GS, 4D-SFGS, SpeeDe3DGS, and MoRel while retaining real-time FPS and modest neural size (Tables 1–2), with module and aggregation ablations (Tables 3–4) and a weight visualization (Fig. 4).

Significance. Heterogeneous local dynamics under monocular UAV capture are a genuine and under-served setting relative to object-centric dynamic Gaussian work. Conditioning shared multi-plane readout on anchor- and time-specific weights is a clear, implementable idea that preserves compactness while targeting ghosting and blur. The paper provides consistent gains on three datasets, structured ablations isolating ACFA/DLGD/DACW, an aggregation-strategy study, and a qualitative weight analysis aligned with static vs. dynamic regions. Public code is promised. If the quality ranking holds under better-matched capacity and reporting, the work is a useful systems contribution for aerial 4D reconstruction rather than a purely incremental deformation tweak.

major comments (3)
  1. [§5.1, Tables 1–2] §5.1 states that all methods share camera poses, sparse initialization, split, images, and main-stage iteration count, but does not match or report representation capacity (anchor/Gaussian count, densification budget, shared-field width, or deformation MLP size). Tables 1–2 show Neural Size ranging from ~0.11 MB (4D-SFGS) to ~30 MB (MoDec-GS), with Ours at 7.6–9.4 MB. Several baselines are non-anchor or differently factored (D3DGS, 4D-GS, SpeeDe3DGS), while anchor peers use different deformation pathways. The headline attribution of cross-method gains to ACFA (abstract; §5.2) is therefore not isolated from backbone/capacity differences. Please report per-method primitive counts and, where feasible, capacity-controlled or matched-backbone comparisons (e.g., fixed aggregation vs. ACFA on the same anchor+multi-plane stack already used in Tables 3–4, extended to the full baseline set or to M
  2. [§5.2–5.3, Tables 1–4] Tables 1–4 report single-run point estimates with no seed variance, scene-wise standard deviation, or significance testing, despite free parameters in ACFA/DLGD/DACW and a two-stage schedule (§5.1). The claimed margins (e.g., +1.69 dB PSNR on UAV-Arc4D over the second-best method) are large enough to be plausible, but without multi-seed or leave-one-scene statistics it is hard to judge stability—especially on the self-collected UAV-Arc4D set and on smaller VisDrone/UAVDT subsets. Please add scene-level breakdowns and at least multi-seed or repeated-split variance for the full model and the main ablation row.
  3. [§5.1; Conclusion] §5.1 runs COLMAP on the complete video sequence to obtain poses and the initial sparse cloud for all methods; the Conclusion correctly notes reliance on pre-estimated poses. For monocular dynamic evaluation this is common, but it couples test-time geometry to full-sequence SfM and weakens claims that quality gaps isolate the dynamic representation under realistic UAV pose noise. Please clarify whether test frames participate in SfM, quantify pose sensitivity (e.g., controlled pose noise or train-only COLMAP), and keep relative rankings only as representation comparisons under shared poses rather than as end-to-end aerial robustness.
minor comments (5)
  1. [§5.1] UAV-Arc4D is central to the strongest quantitative claim but is only briefly introduced in §5.1 (‘10 urban road and traffic scenes’), with details deferred to the supplement. A short main-text description of capture altitude, resolution, motion types, train/test split protocol, and licensing/release plan would help readers assess generality.
  2. [§5.3, Fig. 4] Fig. 4 supports the intended static-vs-dynamic weighting story, but the main text does not define how ‘spatial-plane’ vs. ‘spatiotemporal-plane’ contributions are aggregated from the six planes and C channels. A one-line definition would make the figure falsifiable.
  3. [§4.4, §5.1] DACW hyperparameters γ, λ_min, λ_max, and bin count B are only partially specified in §5.1 (B=64); sensitivity of final quality to freezing maps after the coarse stage is not shown. A small sensitivity table in the supplement would strengthen §4.4.
  4. [§2.1] Related work (§2.1–2.2) is generally appropriate; briefly contrasting ACFA’s channel-wise softmax over planes with Grid4D’s directional attention (already cited) in one sentence would sharpen the novelty claim.
  5. [§3–4.2] Minor notation/clarity: in §3 the product aggregation (Eq. 2) is the baseline fixed rule, while ACFA uses a weighted sum (Eq. 6)—state explicitly that this changes both adaptivity and the aggregator family. Also ensure consistent spelling of ‘spatiotemporal’ and clean spacing in the camera-ready text (PDF extraction shows many concatenated words).

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper with held-out rendering metrics and architectural ablations, not a self-referential derivation.

full rationale

AdaAnchor4D is a monocular UAV dynamic Gaussian reconstruction method whose central claims are empirical: higher PSNR/SSIM/LPIPS than named baselines on UAV-Arc4D, VisDrone, and UAVDT while retaining real-time FPS (Tables 1–2), attributed to ACFA (anchor- and time-conditioned plane weights), DLGD, and DACW. The method is trained end-to-end for novel-view synthesis; quality is measured on held-out views with standard external image metrics under matched poses, sparse init, splits, images, and main-stage iterations (§5.1). Ablations (Tables 3–4) compare aggregation operators and module combinations inside the same stack; Fig. 4 visualizes learned weights but does not treat them as predictions of fitted targets. Zero-init of Φw/Φo/Φs and freezing DACW maps after the coarse stage are optimization stabilizers, not definitional identities. Related-work citations (4D-GS, Scaffold-GS, MoDec-GS, etc.) supply prior art, not load-bearing uniqueness theorems or self-cited ansätze that force the reported gains. There is no step where a claimed first-principles result reduces by construction to its inputs. Fairness concerns about unmatched baseline capacity are comparative-evaluation issues, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

Load-bearing content is architectural and empirical, not axiomatic physics. The claim rests on standard differentiable rendering assumptions, COLMAP geometry, the multi-plane factorization borrowed from prior work, and several hand-chosen network/warping hyperparameters. Invented pieces are learned modules (aggregation embeddings, geometry condition embeddings, warping maps), not new physical entities.

free parameters (5)
  • ACFA aggregation embedding dim and weight-MLP hidden size = 32 / 256
    Set to 32 and 256; control capacity of plane-weight prediction and thus the main adaptive mechanism.
  • DLGD geometry condition embedding dim and branch hidden size = 32 / 128
    Set to 32 and 128; govern local offset/scale residual pathways.
  • DACW histogram bins B and density sensitivity γ, λmin, λmax = B=64; γ, λ range in supplement
    B=64 stated; γ and λ bounds control how strongly non-uniform anchors warp query coordinates and therefore feature capacity allocation.
  • Coarse/fine iteration schedule = 3000 / 30000
    3000 coarse + 30000 fine iterations fix when DACW maps freeze and how long adaptive weights train.
  • Number of local Gaussians K per anchor and plane feature dimension C
    Inherited scaffold/multi-plane hyperparameters that set representation capacity; exact values partly deferred to supplement.
axioms (5)
  • domain assumption Differentiable 3D Gaussian splatting with photometric supervision yields a valid proxy for dynamic scene fidelity under novel views.
    Standard 3DGS/dynamic-GS evaluation premise used throughout §5.
  • domain assumption COLMAP poses and sparse points from the full monocular sequence are accurate enough to compare dynamic representations fairly.
    §5.1 applies the same COLMAP outputs to all methods; Conclusion flags pose fragility.
  • domain assumption Scene dynamics can be modeled as deformation of canonical anchors/Gaussians queried from a six-plane spatiotemporal factorization.
    Preliminaries §3 and method overview inherit 4D-GS-style planes rather than re-deriving them.
  • ad hoc to paper Axis-wise marginal anchor histograms after the coarse stage are a stable enough density signal to freeze DACW maps for the fine stage.
    §4.4 constructs mappings once and keeps them fixed through densification.
  • ad hoc to paper Softmax channel-wise plane weights conditioned on a per-anchor embedding and scalar time suffice to capture heterogeneous local dynamics without per-region fields.
    Core ACFA design choice in §4.2; supported empirically but not theoretically necessary.
invented entities (3)
  • Anchor-specific aggregation embedding e_acfa_i no independent evidence
    purpose: Condition plane aggregation weights so each anchor reads the shared field differently over time.
    Learned latent introduced in §4.2; only justified by ablations and weight plots inside this paper.
  • Geometry condition embeddings e_o_i and e_s_i no independent evidence
    purpose: Provide dedicated pathways for local offset and scale residuals separate from ACFA-driven anchor state.
    Introduced in §4.3 DLGD; no external measurement identifies them outside rendering loss.
  • Density-Adaptive Coordinate Warping maps from axis-wise anchor histograms no independent evidence
    purpose: Reallocate uniform grid query capacity toward non-uniform UAV anchor distributions.
    Constructed in §4.4 from coarse-stage anchors; utility shown only via ablation PSNR.

reviewed 2026-07-31 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AdaAnchor4D: Anchor-Conditioned Spatiotemporal Feature Aggregation for Monocular UAV 4D Reconstruction." pith.science (2026). https://pith.science/paper/CN3K3H2H

@misc{pith2026260728320,
  author       = {Pith},
  title        = {Pith review of: AdaAnchor4D: Anchor-Conditioned Spatiotemporal Feature Aggregation for Monocular UAV 4D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CN3K3H2H}},
  note         = {Machine review of arXiv:2607.28320}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Monocular UAV videos provide valuable observations for dynamic reconstruction of complex urban scenes. However, such scenes exhibit pronounced spatiotemporal heterogeneity: different regions follow distinct temporal activity patterns, while the motion states of some dynamic regions may further evolve over time. Although dynamic Gaussian methods based on decomposed shared spatiotemporal feature fields have achieved efficient and accurate reconstruction in object-centric or relatively compact scenes, their commonly adopted fixed plane-wise feature combination mechanisms are less suited to the heterogeneous local dynamics of UAV scenes, often leading to ghosting artifacts and blurred dynamic details. To address this challenge, we propose AdaAnchor4D, an adaptive anchor deformation framework for monocular UAV dynamic scene reconstruction. At its core, Anchor-Conditioned Feature Aggregation (ACFA) adaptively aggregates shared spatiotemporal features using anchor-specific aggregation embeddings and temporal information, allowing different local units to obtain dynamic representations tailored to their local and temporal states. Decoupled Local Geometry Deformation (DLGD) separates anchor-state deformation from local Gaussian geometry deformation, while Density-Adaptive Coordinate Warping (DACW) reparameterizes feature-query coordinates according to the axis-wise anchor distributions, alleviating the mismatch between non-uniform geometric sampling and uniform grid parameterization. Experiments on UAV-Arc4D, VisDrone, and UAVDT show that AdaAnchor4D achieves higher rendering quality than representative dynamic Gaussian methods while maintaining real-time rendering performance. The code will be made publicly available.

Figures

Figures reproduced from arXiv: 2607.28320 by Guanbin Li, Guangming Shi, Jie Feng, Junpeng Zhang, Le Dong, Mingtao Feng, Peiyi Xu, Ronghua Shang, Weisheng Dong.

Figure 1
Figure 1. Figure 1: Monocular UAV scenes exhibit strong spatiotempo [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of AdaAnchor4D. Anchor-Conditioned Feature Aggregation performs anchor- and time-dependent feature [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparisons on UAV-Arc4D and VisDrone. The top three rows correspond to UAV-Arc4D, while the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the learned ACFA aggregation [PITH_FULL_IMAGE:figures/full_fig_p007_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.

Reference graph

Works this paper leans on

46 extracted references · 5 linked inside Pith

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    4d scaffold gaussian splatting with dynamic-aware anchor growing for efficient and high-fidelity dynamic scene reconstruction , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  13. [13]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =. 2023 , url =

    2023

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Haif-gs: Hierarchical and induced flow-guided gaussian splatting for dynamic scene , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    4d gaussian splatting for real-time dynamic scene rendering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [16]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scaffold-gs: Structured 3d gaussians for view-adaptive rendering , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Modec-gs: Global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3d gaussian splatting , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  18. [18]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spacetime gaussian feature splatting for real-time dynamic view synthesis , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  20. [20]

    European Conference on Computer Vision , pages=

    Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splatting , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  21. [21]

    arXiv preprint arXiv:2308.09713 , year=

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis , author=. arXiv preprint arXiv:2308.09713 , year=

    Pith/arXiv arXiv

  22. [22]

    ACM SIGGRAPH 2024 Conference Papers , pages=

    4d-rotor gaussian splatting: towards efficient novel view synthesis for dynamic scenes , author=. ACM SIGGRAPH 2024 Conference Papers , pages=

    2024

  23. [23]

    International Conference on Learning Representations , volume=

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting , author=. International Conference on Learning Representations , volume=

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Fully explicit dynamic gaussian splatting , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    arXiv preprint arXiv:2604.04063 , year=

    4C4D: 4 Camera 4D Gaussian Splatting , author=. arXiv preprint arXiv:2604.04063 , year=

    Pith/arXiv arXiv

  26. [26]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Tu, Allen and Ying, Haiyang and Hanson, Alex and Lee, Yonghan and Goldstein, Tom and Zwicker, Matthias , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2026 , pages =

    2026

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    K-planes: Explicit radiance fields in space, time, and appearance , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  29. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Degauss: Dynamic-static decomposition with gaussian splatting for distractor-free 3d reconstruction , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [30]

    arXiv preprint arXiv:2507.07465 , year=

    SD-GS: Structured Deformable 3D Gaussians for Efficient Dynamic Scene Reconstruction , author=. arXiv preprint arXiv:2507.07465 , year=

    Pith/arXiv arXiv

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    4d gaussian splatting in the wild with uncertainty-aware regularization , author=. Advances in Neural Information Processing Systems , volume=

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Grid4d: 4d decomposed hash encoding for high-fidelity dynamic gaussian splatting , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    arXiv preprint arXiv:2408.13972 , year=

    Dynasurfgs: Dynamic surface reconstruction with planar-based gaussian splatting , author=. arXiv preprint arXiv:2408.13972 , year=

    Pith/arXiv arXiv

  34. [34]

    Image and Vision Computing , volume=

    Drone-NeRF: Efficient NeRF based 3D scene reconstruction for large-scale drone survey , author=. Image and Vision Computing , volume=. 2024 , publisher=

    2024

  35. [35]

    arXiv preprint arXiv:2409.00381 , year=

    3d gaussian splatting for large-scale surface reconstruction from aerial images , author=. arXiv preprint arXiv:2409.00381 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vastgaussian: Vast 3d gaussians for large scene reconstruction , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  37. [37]

    arXiv preprint arXiv:2602.02089 , year=

    UrbanGS: A Scalable and Efficient Architecture for Geometrically Accurate Large-Scene Reconstruction , author=. arXiv preprint arXiv:2602.02089 , year=

    arXiv

  38. [38]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    AeroGS: Scale-Aware Gaussian Splatting for Pose-Free Dynamic UAV Scene Reconstruction , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  39. [39]

    arXiv preprint arXiv:2602.22376 , year=

    AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D Reconstruction , author=. arXiv preprint arXiv:2602.22376 , year=

    arXiv

  40. [40]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Uav4d: Dynamic neural rendering of human-centric uav imagery using gaussian splatting , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  41. [41]

    2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    TK-planes: Tiered K-planes with high dimensional feature vectors for dynamic UAV-based scenes , author=. 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2025 , organization=

    2025

  42. [42]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Detection and tracking meet drones challenge , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2021 , publisher=

    2021

  43. [43]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    The unmanned aerial vehicle benchmark: Object detection and tracking , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  44. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    MoRel: Long-Range Flicker-Free 4D Motion Modeling via Anchor Relay-based Bidirectional Blending with Hierarchical Densification , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  45. [45]

    Structure-from-Motion Revisited , booktitle=

    Sch\". Structure-from-Motion Revisited , booktitle=

  46. [46]

    Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=

    Sch\". Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=

This paper was first reviewed by grok-4.5 on July 31, 2026.