Pith. sign in

REVIEW 5 major objections 5 minor 42 references

This paper argues that the core of asynchronous multi-view dynamic reconstruction is motion observability, and that aligning reconstructed 3D point trajectories with observed 2D motion tracks solves it where color matching fails.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 16:45 UTC pith:EVW3OZMS

load-bearing objection A genuinely useful idea for async dynamic reconstruction, but the core supervision signal is a black box and the paper's own ablation shows long tracks fail exactly where the headline results live. the 5 major comments →

arxiv 2608.02006 v1 pith:EVW3OZMS submitted 2026-08-03 cs.CV

ASTRA: Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment

Junyu Zhu , Hao Zhu , Xinzhuo Zhang , Hongdong Li , Zhan Ma , Xun Cao This is my paper
classification cs.CV
keywords dynamic 3D reconstructionasynchronous multi-view videostemporal synchronizationmotion trajectory alignmentGaussian splattingkinematic priorstrajectory maskingtemporal offset estimation
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 paper argues that photometric supervision is fundamentally the wrong signal for synchronizing unsynchronized multi-camera dynamic reconstruction: in low-texture regions the color gradient vanishes, and in deforming scenes the optimization hides temporal errors in distorted geometry. It proposes ASTRA, which instead anchors the reconstructed dynamic 3D representation to precomputed 2D motion trajectories — point tracks that follow physical scene motion — and jointly optimizes per-camera temporal offsets and Gaussian scene parameters by aligning the projected motion of reconstructed 3D points to those tracks. On dynamic Gaussian backbones, this yields roughly 1.4 dB PSNR improvement, 54% lower temporal-offset error, and nearly four times higher synchronization success rate at offsets up to 25 frames. The reader should care because asynchronous capture is the norm outside controlled studios, and this reframes the synchronization problem as a motion-observability problem rather than an appearance-matching problem.

Core claim

The central claim is that temporal alignment of multi-view dynamic scenes is reliably achieved by matching projected motion, not rendered color. ASTRA takes precomputed 2D point trajectories as texture-agnostic anchors and minimizes L_traj, the squared distance between the projected position of neighboring Gaussian centers queried at the offset-corrected time and the observed 2D track position. Because the trajectory loss carries Gaussian velocities through the projection Jacobian rather than through an appearance Jacobian, it retains gradient strength in textureless regions; and because it constrains many points across views and timestamps, it prevents deformation updates from absorbing tem

What carries the argument

The trajectory alignment loss L_traj (Eqn 6) with its spatio-temporal trajectory anchoring (Eqn 5): precomputed 2D point tracks, filtered by a dynamic mask and a certainty mask, are matched to the projected, blend-weighted positions of nearby time-varying Gaussian centers at offset-corrected physical times. This single objective couples all cameras through the shared Gaussian representation and supplies the gradient that drives both offset estimation and temporally coherent Gaussian trajectories.

Load-bearing premise

The load-bearing premise is that the precomputed 2D motion tracks reliably follow physical scene points under large offsets and weak texture; the paper's own conclusion concedes its simple reliability mask may underperform under severe occlusion and that rolling shutter, clock drift, and exposure variation are left unaddressed.

What would settle it

Record a multi-view sequence of a mostly textureless scene with a fast-moving object that is repeatedly occluded, with hardware-synchronized ground-truth offsets, and run ASTRA's track/mask pipeline: if the offset MAE degrades toward chance whenever the tracker drifts, the central claim that trajectory supervision suffices fails; a cleaner version replaces tracks with ground-truth optical flow and checks whether synchronization errors vanish.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Temporal offsets of up to about 25 frames (near three seconds of asynchrony) can be recovered well enough to nearly match reconstruction quality of a model given ground-truth offsets.
  • The trajectory term is backbone-agnostic: it improves both deformation-field-based and basis-expansion-based dynamic Gaussian reconstructions, so it can be added to existing pipelines.
  • Photometric-only synchronization is formally at risk of two failure modes—vanishing gradients in low texture and deformation compensation—so future async-reconstruction methods need either motion supervision or a reason appearance suffices.
  • Track length should scale with the expected offset: short tracks for small offsets, moderate tracks for large offsets, since overly long tracks lose reliability.

Where Pith is reading between the lines

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

  • If this line holds, hardware clock synchronization could be replaced by software trajectory alignment in casual multi-camera capture, and the same motion-anchoring principle should apply to other appearance-decoupled unknowns such as rolling-shutter rows or per-camera exposure.
  • A testable extension: substituting an occlusion-aware long-range tracker for the current simple certainty mask should extend the method beyond 25-frame offsets and into severe-occlusion scenes, where the paper itself flags potential weakness.
  • The gradient-observability argument suggests a general recipe: whenever jointly estimating time parameters and 3D structure, choose observables whose Jacobian is independent of appearance and cannot be mimicked by geometric deformation—motion tracks are one instance.

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 / 5 minor

Summary. The paper proposes ASTRA, a framework for dynamic 3D reconstruction from unsynchronized multi-view videos. It augments the standard photometric loss with a trajectory alignment loss (Eq. 6) that compares the projected motion of reconstructed Gaussian centers with precomputed 2D motion tracks, using dynamic and certainty masks to select reliable anchors. The authors claim this 'texture-agnostic' kinematic supervision decouples temporal synchronization from appearance and deformation, and they report consistent improvements over 4DGS and EDGS baselines for temporal offsets up to 25 frames, including higher PSNR, better motion consistency, and lower temporal-offset MAE. The paper also provides an analytical gradient comparison between photometric and trajectory supervision and presents ablations on track length, masks, and loss weights.

Significance. If the external 2D tracks are reliable, the core idea is significant: using motion trajectories as explicit synchronization cues is a plausible and under-explored direction for asynchronous multi-view capture. The paper evaluates on multiple datasets and two dynamic-Gaussian backbones, and the reported gains are large and consistent. The gradient comparison (Eqs. 8-11) is a useful pedagogical heuristic, and the mask and weight ablations are informative. However, the central empirical claim depends on an external supervision signal whose quality is never measured, and the theoretical claims about convexity and observability are not substantiated. These gaps currently limit the paper's reproducibility and generalizability claims.

major comments (5)
  1. [Sec. 3.2-3.3, Eqns (5)-(6)] The central supervision signal—the precomputed 2D tracks and the masks—is never characterized. No tracker is named, no thresholds for the dynamic/certainty masks are given, no count of surviving tracks per scene is reported, and no tracking-accuracy metric is provided. Because L_traj is minimized directly against these tracks, drift or an occlusion jump misleads the offset and geometry optimization. The paper's own Sec. 6 concedes that the reliability mask may underperform under severe occlusion. This is load-bearing: the claimed texture-agnostic advantage depends entirely on the quality of an external, unvalidated input. Please specify the tracker, report track survival/accuracy, and include a sensitivity study to track noise (e.g., corrupting a fraction of anchors).
  2. [Sec. 5.1, Table 3] The track-length ablation undercuts the claimed robustness of multi-frame trajectory supervision. On Deer at offset 15, Length 15 gives synchronization MAE 0.5190 and SR@1 0.0769, whereas Length 5 gives MAE 0.0339 and SR@1 0.3846. Yet the paper advertises ASTRA-Full (multi-frame) mainly for large offsets (20-25). The configuration of ASTRA-Full (track length, support-set size) is not specified in Sec. 4.1 or Table 3. Without stating which track length is used at offsets 20-25 and why it does not suffer the collapse shown in Table 3, the central large-offset claim is not fully supported.
  3. [Sec. 3.4, Eqns (9)-(10)] Eqn. (9) omits the time-dependence of the covariance. Eqn. (2) defines Sigma_i(t) through a time-dependent rotation R_i(t), and rendered color depends on projected covariance as well as position. The chain rule should include a term (partial hat I / partial Sigma_i) * d Sigma_i/dt. Without it, the 'analytical gradient' is incomplete and the comparison of temporal sensitivity in Sec. 3.4.1 is not exact. Please correct the derivation or state that covariance effects are neglected and justify that they are small.
  4. [Sec. 3.4.2 and Sec. 4.4, Eqn. (12), Figs. 4-5] The observability analysis is heuristic. Eqn. (12) is stated but never evaluated; no rank computation for J_G and J_Delta_t is provided for either objective. Fig. 4 shows 1D slices of the loss at a converged solution; such slices do not establish a 'distinctly convex loss landscape' (Sec. 4.4) or global convexity. Please either provide a concrete computation of the rank condition (or a Hessian check) for a simplified model, or rephrase the claim to 'locally sharper and more consistent minima'.
  5. [Eqn. (5)] The support sets N_{k,j} and interpolation weights w_{k,j,i} are not defined. If these weights depend on the Gaussian centers (e.g., inverse-distance weighting at time t), then the gradient in Eqn. (11) must include d w_{k,j,i}/dt terms; as written, the weights are treated as constants. Please specify the construction of the support sets and weights and provide the full gradient.
minor comments (5)
  1. [Sec. 3.4.1] Spacing issue: 'T exture-induced collapse' appears twice. Please fix the LaTeX/formatting.
  2. [Table 1] Header 'Meet Room Datasets' should be 'MeetRoom Datasets' to match the dataset name used elsewhere.
  3. [Sec. 4.1] The temporal-slice evaluation protocol is attributed to prior work [40]; please describe the exact slice construction and how SSIM/LPIPS are computed on them so the metric is self-contained.
  4. [Fig. 1] The caption says 'Radial Values (5~25) denote Offsets' but the radial axis values are not visible in the text version. Please ensure the figure clearly shows offset values.
  5. [Sec. 5.4] The overhead numbers are useful, but the comparison between 4DGS and EDGS would be clearer if the table also reported whether ASTRA-Full uses the same effective track length in both backbones.

Circularity Check

0 steps flagged

No significant circularity: the trajectory supervision is external to the model, and the minor self-citation in the evaluation protocol is not load-bearing.

full rationale

The central derivation chain is not circular. The temporal offsets are optimized by minimizing Eq. 6, which compares the projected Gaussian motion (Eq. 5) with precomputed 2D tracks u_k,j(t_j). These tracks are external observations of the input videos; they do not depend on the reconstructed Gaussians or on the learned offsets, so the offset estimate is not fitted to a quantity that already encodes the answer. The gradient analysis in Sec. 3.4 is a chain-rule expansion of the two losses and does not reduce either objective to the other. The ASTRA-Base/ASTRA-Full comparison is an internal ablation, not a predicted-against-input construction. The only self-citation found is Ref. [40] (the authors' own Scale-NeRF paper) for the temporal-slice evaluation protocol: "This protocol is inspired by the temporal-slice evaluation method utilized in Scale-NeRF [40]." This is an evaluation-metric choice, not a load-bearing premise of the method, and it does not make the reconstruction or synchronization result depend on the cited work. The paper's own limitation statement — "the framework relies on a simple reliability mask that may underperform under severe occlusion" — and the unspecified 2D tracker in Sec. 3.2 highlight external-validity and robustness risks, not circular reductions; per the review rules those concerns belong to correctness risk, not to the circularity score. The derivation is otherwise self-contained.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests primarily on domain assumptions about the quality of external 2D tracking and the validity of the projection association (Eqn 5); no new physical entities are introduced. The loss weights and track length selection are fitted parameters that influence the reported results.

free parameters (4)
  • lambda_traj (trajectory loss weight) = 0.1
    Chosen by ablation on the Blender Deer sequence (Tab. 5); used for all reported results.
  • lambda_dt (temporal offset regularization weight) = 0.0005
    Chosen by ablation on the Blender Deer sequence (Tab. 5); used for all reported results.
  • Track length (ASTRA-Base vs ASTRA-Full) = 1 (Base) vs multi-frame (Full, up to 15 in ablation)
    The paper reports two variants and effectively selects the better one per offset regime; ablation (Tab. 3) shows the optimal track length varies with offset, so this is a tuned parameter.
  • Dynamic and certainty mask thresholds = not specified
    The masks are described qualitatively in Sec 3.2; their construction or thresholds are chosen by the authors but not reported, making them implicit free parameters.
axioms (5)
  • domain assumption The scene is representable as time-varying 3D Gaussians with a deformation field (Eqns 1-2).
    This is the backbone representation adopted from 4DGS and EDGS; the entire method operates on these primitives.
  • domain assumption Each camera has a constant temporal offset Delta_t_j over the whole sequence (Sec 3.1).
    The model optimizes a single scalar offset per camera; clock drift and rolling shutter are excluded and listed as future work.
  • domain assumption The precomputed 2D tracks u_{k,j}(t_j) are projections of physical 3D points at the recorded timestamps and stay reliable under large offsets and motion blur (Sec 3.2).
    The trajectory loss in Eqn 6 treats the tracks as ground truth; if the tracker fails, the supervision is corrupted.
  • ad hoc to paper The support sets N_{k,j} and interpolation weights w_{k,j,i} in Eqn 5 yield a valid projection of the Gaussian centers to the 2D anchor position.
    The paper does not define how these are computed; this association is a modeling choice needed to connect continuous Gaussian centers to discrete track points.
  • domain assumption The dynamic and certainty masks can separate reliable moving tracks from unreliable ones without ground truth.
    The masks are described verbally in Sec 3.2 and ablated in Tab 4, but their operation is assumed to improve rather than harm the trajectory signal.

pith-pipeline@v1.3.0-daily-deepseek · 24192 in / 14667 out tokens · 147695 ms · 2026-08-04T16:45:59.046040+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of ASTRA: Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment." pith.science (2026). https://pith.science/paper/EVW3OZMS

@misc{pith2026260802006,
  author       = {Pith},
  title        = {Pith review of: ASTRA: Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EVW3OZMS}},
  note         = {Machine review of arXiv:2608.02006}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Dynamic 3D scene reconstruction has achieved remarkable success under the assumption of strictly synchronized multi-camera inputs. However, in real-world scenarios, temporal asynchrony among capturing devices remains a critical challenge, leading to severe motion blur and geometric artifacts. Existing asynchronous reconstruction methods typically estimate temporal offsets through photometric supervision, but appearance matching provides weak temporal cues under large offsets and complex motions. We attribute this limitation to two major bottlenecks: texture-induced collapse, where low-texture regions provide nearly vanishing alignment signals, and deformation-induced coupling, where temporal errors are absorbed into distorted geometry or motion rather than being explicitly corrected. To address these issues, we propose ASTRA (Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment), a framework that introduces 2D motion trajectories as explicit, texture-agnostic supervision for asynchronous dynamic reconstruction. Instead of synchronizing cameras solely through rendered color residuals, ASTRA jointly optimizes temporal offsets and dynamic 3D representations by aligning the projected motion of reconstructed 3D points with observed 2D trajectories, while using dynamic and certainty masking to suppress unreliable trajectory constraints. Extensive experiments on different dynamic Gaussian Splatting backbones show that ASTRA preserves high-frequency spatial details and sustains strong robustness even under severe asynchrony with up to 25-frame offsets, achieving approximately 1.4 dB PSNR improvement, reducing temporal-offset MAE by 54.0\%, and nearly quadrupling the synchronization success rate.

Figures

Figures reproduced from arXiv: 2608.02006 by Hao Zhu, Hongdong Li, Junyu Zhu, Xinzhuo Zhang, Xun Cao, Zhan Ma.

Figure 1
Figure 1. Figure 1: Comprehensive evaluation of the proposed ASTRA framework implemented on 4DGS. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed pipeline for joint temporal synchronization and dynamic reconstruction in unsynchronized multi-camera systems. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Disambiguating textureless regions using motion trajectory su [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Convergence analysis and temporal loss profiles under different guidance strategies. Left: Convergence curves of the synchronization metric [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporal-offset gradient landscapes under the color-guided [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative visual comparison of spatial patches, temporal slices, and motion trajectories at different offsets. The trajectory column displays [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of spatio-temporal flame reconstruction. The top-left illustrates the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Compared to other methods, our approach better preserves structural details and fine textures. Notably, our method retains sharper contours, [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of spatiotemporal image rendering. Our method demonstrates enhanced detail preservation and spatiotemporal [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of reconstruction results under varying temporal offsets. Notably, under severe temporal misalignment (e.g., Offset [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of temporal slice rendering under varying track lengths and temporal offsets. The [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison of the mask ablation study. Visual results [PITH_FULL_IMAGE:figures/full_fig_p014_12.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

42 extracted references · 8 linked inside Pith

  1. [1]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P . P . Srinivasan, M. Tancik, J. T. Barron, R. Ra- mamoorthi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”Communications of the ACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  2. [2]

    3d gaus- sian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaus- sian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  3. [3]

    Neural 3d video synthesis from multi-view video,

    T. Li, M. Slavcheva, M. Zollhoefer, S. Green, C. Lassner, C. Kim, T. Schmidt, S. Lovegrove, M. Goesele, R. Newcombeet al., “Neural 3d video synthesis from multi-view video,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 5521–5531

    2022

  4. [4]

    Nerfies: Deformable neural radiance fields,

    K. Park, U. Sinha, J. T. Barron, S. Bouaziz, D. B. Goldman, S. M. Seitz, and R. Martin-Brualla, “Nerfies: Deformable neural radiance fields,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5865–5874

    2021

  5. [5]

    Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic recon- struction and rendering,

    R. Shao, Z. Zheng, H. Tu, B. Liu, H. Zhang, and Y. Liu, “Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic recon- struction and rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 16 632–16 642

    2023

  6. [6]

    K-planes: Explicit radiance fields in space, time, and appearance,

    S. Fridovich-Keil, G. Meanti, F. R. Warburg, B. Recht, and A. Kanazawa, “K-planes: Explicit radiance fields in space, time, and appearance,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 12 479–12 488

    2023

  7. [7]

    Hexplane: A fast representation for dynamic scenes,

    A. Cao and J. Johnson, “Hexplane: A fast representation for dynamic scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 130–141

    2023

  8. [8]

    D- nerf: Neural radiance fields for dynamic scenes,

    A. Pumarola, E. Corona, G. Pons-Moll, and F. Moreno-Noguer, “D- nerf: Neural radiance fields for dynamic scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 10 318–10 327

    2021

  9. [9]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation,

    Q. Gao, Q. Xu, Z. Cao, B. Mildenhall, W. Ma, L. Chen, D. Tang, and U. Neumann, “Gaussianflow: Splatting gaussian dynamics for 4d content creation,”arXiv preprint arXiv:2403.12365, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    4d gaussian splatting for real-time dynamic scene rendering,

    G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang, “4d gaussian splatting for real-time dynamic scene rendering,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 310–20 320

    2024

  11. [11]

    De- formable 3d gaussians for high-fidelity monocular dynamic scene reconstruction,

    Z. Yang, X. Gao, W. Zhou, S. Jiao, Y. Zhang, and X. Jin, “De- formable 3d gaussians for high-fidelity monocular dynamic scene reconstruction,” inProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, pp. 20 331–20 341

    2024

  12. [12]

    4k4d: Real-time 4d view synthesis at 4k resolution,

    Z. Xu, S. Peng, H. Lin, G. He, J. Sun, Y. Shen, H. Bao, and X. Zhou, “4k4d: Real-time 4d view synthesis at 4k resolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 029–20 040

    2024

  13. [13]

    Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,

    J. Luiten, G. Kopanas, B. Leibe, and D. Ramanan, “Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis,”arXiv preprint arXiv:2308.09713, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Spacetime gaussian feature splatting for real-time dynamic view synthesis,

    Z. Li, Z. Chen, Z. Li, and Y. Xu, “Spacetime gaussian feature splatting for real-time dynamic view synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 8508–8520

    2024

  15. [15]

    A compact dynamic 3d gaussian representation for real-time dynamic view synthesis,

    K. Katsumata, D. M. Vo, and H. Nakayama, “A compact dynamic 3d gaussian representation for real-time dynamic view synthesis,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 394–412

    2024

  16. [16]

    Fast dynamic radiance fields with time-aware neural voxels,

    J. Fang, T. Yi, X. Wang, L. Xie, X. Zhang, W. Liu, M. Nießner, and Q. Tian, “Fast dynamic radiance fields with time-aware neural voxels,” inSIGGRAPH Asia 2022 Conference Papers, 2022, pp. 1–9

    2022

  17. [17]

    De- formgs: Scene flow in highly deformable scenes for deformable object manipulation,

    B. P . Duisterhof, Z. Mandi, Y. Yao, J.-W. Liu, J. Seidenschwarz, M. Z. Shou, D. Ramanan, S. Song, S. Birchfield, B. Wenet al., “De- formgs: Scene flow in highly deformable scenes for deformable object manipulation,”arXiv preprint arXiv:2312.00583, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Sync- nerf: Generalizing dynamic nerfs to unsynchronized videos,

    S. Kim, J. Bae, Y. Yun, H. Lee, G. Bang, and Y. Uh, “Sync- nerf: Generalizing dynamic nerfs to unsynchronized videos,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 3, 2024, pp. 2777–2785

    2024

  19. [19]

    Hu- mans as a calibration pattern: Dynamic 3d scene reconstruction from unsynchronized and uncalibrated videos,

    C. Choi, J. Kim, G. Cha, M. Kim, D. Wee, and Y. M. Kim, “Hu- mans as a calibration pattern: Dynamic 3d scene reconstruction from unsynchronized and uncalibrated videos,”arXiv preprint arXiv:2412.19089, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    4dslomo: 4d reconstruction for high speed scene with asyn- chronous capture,

    Y. Chen, S. Guo, T. Yang, L. Ding, X. Yu, J. Gu, and T. Xue, “4dslomo: 4d reconstruction for high speed scene with asyn- chronous capture,” inProceedings of the SIGGRAPH Asia 2025 Conference Papers, 2025, pp. 1–11

    2025

  21. [21]

    Dynamic gaussian scene reconstruction from unsynchronized videos,

    Z. Xu, H. Zhou, Y. Liu, W. Xue, H. Pan, W. Wang, and B. Wang, “Dynamic gaussian scene reconstruction from unsynchronized videos,” inProceedings of the AAAI Conference on Artificial Intelli- gence, vol. 40, no. 14, 2026, pp. 11 469–11 477. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 16

    2026

  22. [22]

    Neural scene flow fields for space-time view synthesis of dynamic scenes,

    Z. Li, S. Niklaus, N. Snavely, and O. Wang, “Neural scene flow fields for space-time view synthesis of dynamic scenes,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 6498–6508

    2021

  23. [23]

    Mixed neural voxels for fast multi-view video synthesis,

    F. Wang, S. Tan, X. Li, Z. Tian, Y. Song, and H. Liu, “Mixed neural voxels for fast multi-view video synthesis,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 706–19 716

    2023

  24. [24]

    Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting,

    Z. Yang, H. Yang, Z. Pan, and L. Zhang, “Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting,”arXiv preprint arXiv:2310.10642, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video,

    J. Park, M.-Q. V . Bui, J. L. G. Bello, J. Moon, J. Oh, and M. Kim, “Splinegs: Robust motion-adaptive spline for real-time dynamic 3d gaussians from monocular video,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 866–26 875

    2025

  26. [26]

    Gaustar: Gaussian surface tracking and reconstruction,

    C. Zheng, L. Xue, J. Zarate, and J. Song, “Gaustar: Gaussian surface tracking and reconstruction,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16 543–16 553

    2025

  27. [27]

    Opti- mizing dynamic nerf and 3dgs with no video synchronization,

    S. Kim, J. Bae, Y. Yun, H. Son, H. Lee, G. Bang, and Y. Uh, “Opti- mizing dynamic nerf and 3dgs with no video synchronization,” in ECCV 2024 Workshop on Wild 3D: 3D Modeling, Reconstruction, and Generation in the Wild, 2024

    2024

  28. [28]

    Flow4r: Unifying 4d reconstruction and tracking with scene flow,

    S. Qian, G. Zhang, S. Wu, and D. Cremers, “Flow4r: Unifying 4d reconstruction and tracking with scene flow,”arXiv preprint arXiv:2602.14021, 2026

    Pith/arXiv arXiv 2026

  29. [29]

    Tapip3d: Tracking any point in persistent 3d geometry,

    B. Zhang, L. Ke, A. W. Harley, and K. Fragkiadaki, “Tapip3d: Tracking any point in persistent 3d geometry,”arXiv preprint arXiv:2504.14717, 2025

    arXiv 2025

  30. [30]

    Tracking everything everywhere all at once,

    Q. Wang, Y.-Y. Chang, R. Cai, Z. Li, B. Hariharan, A. Holynski, and N. Snavely, “Tracking everything everywhere all at once,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 795–19 806

    2023

  31. [31]

    Dino-tracker: Taming dino for self-supervised point tracking in a single video,

    N. Tumanyan, A. Singer, S. Bagon, and T. Dekel, “Dino-tracker: Taming dino for self-supervised point tracking in a single video,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 367–385

    2024

  32. [32]

    Emergent temporal correspondences from video diffusion trans- formers,

    J. Nam, S. Son, D. Chung, J. Kim, S. Jin, J. Hur, and S. Kim, “Emergent temporal correspondences from video diffusion trans- formers,”arXiv preprint arXiv:2506.17220, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    Point prompt- ing: Counterfactual tracking with video diffusion models,

    A. Shrivastava, S. Mehta, D. Geng, and A. Owens, “Point prompt- ing: Counterfactual tracking with video diffusion models,”arXiv preprint arXiv:2510.11715, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Spatialtracker: Tracking any 2d pixels in 3d space,

    Y. Xiao, Q. Wang, S. Zhang, N. Xue, S. Peng, Y. Shen, and X. Zhou, “Spatialtracker: Tracking any 2d pixels in 3d space,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 20 406–20 417

    2024

  35. [35]

    Spatialtrackerv2: Advancing 3d point tracking with explicit camera motion,

    Y. Xiao, J. Wang, N. Xue, N. Karaev, Y. Makarov, B. Kang, X. Zhu, H. Bao, Y. Shen, and X. Zhou, “Spatialtrackerv2: Advancing 3d point tracking with explicit camera motion,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 6726–6737

    2025

  36. [36]

    Cotracker3: Simpler and better point tracking by pseudo-labelling real videos,

    N. Karaev, Y. Makarov, J. Wang, N. Neverova, A. Vedaldi, and C. Rupprecht, “Cotracker3: Simpler and better point tracking by pseudo-labelling real videos,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 6013–6022

    2025

  37. [37]

    Camliflow: Bidirectional camera-lidar fusion for joint optical flow and scene flow estimation,

    H. Liu, T. Lu, Y. Xu, J. Liu, W. Li, and L. Chen, “Camliflow: Bidirectional camera-lidar fusion for joint optical flow and scene flow estimation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 5791–5801

    2022

  38. [38]

    Roft: Real-time optical flow-aided 6d object pose and velocity tracking,

    N. A. Piga, Y. Onyshchuk, G. Pasquale, U. Pattacini, and L. Natale, “Roft: Real-time optical flow-aided 6d object pose and velocity tracking,”IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 159–166, 2021

    2021

  39. [39]

    Particlenerf: A particle-based encoding for online neural radiance fields,

    J. Abou-Chakra, F. Dayoub, and N. S ¨underhauf, “Particlenerf: A particle-based encoding for online neural radiance fields,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 5975–5984

    2024

  40. [40]

    Hierarchical bayesian guided spatial-, angular-and temporal-consistent view synthesis,

    J. Zhu, H. Zhu, S. Wang, Z. Ma, and X. Cao, “Hierarchical bayesian guided spatial-, angular-and temporal-consistent view synthesis,” IEEE Transactions on Visualization and Computer Graphics, 2025. Junyu Zhureceived the BS degree from Nanjing University of Posts and Telecommunications, in

    2025

  41. [2010]

    He is a Professor at the School of Electronic Science and Engineering, Nanjing University

    He was a Visiting Scholar with the Uni- versity of Texas at Austin, Austin, TX, USA, from 2010 to 2011. He is a Professor at the School of Electronic Science and Engineering, Nanjing University. His current research interests include computational photog- raphy and image-based modeling and rendering

    2010

  42. [2021]

    His research in- terests include novel view synthesis and implicit neural representation

    He is currently working toward the doctoral degree in the School of Electronic Science and Technology, Nanjing University. His research in- terests include novel view synthesis and implicit neural representation. Hao Zhuis an Assistant Professor in the School of Electronic Science and Engineering, Nanjing University. He received the B.S. and Ph.D. de- gre...

    2014