Pith. sign in

REVIEW 4 major objections 6 minor 62 references

4DGS360 reconstructs a moving object's full 360-degree geometry from a single monocular video, including surfaces the camera never directly sees.

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

2026-08-02 17:40 UTC pith:GTTHIN2K

load-bearing objection A genuinely new initialization idea for 360° dynamic reconstruction, but the load-bearing 'training-free anchor conditioning' is specified only in a notation placeholder and the evaluation lacks error bars and code, so the SOTA claim is not yet verifiable. the 4 major comments →

arxiv 2603.21618 v2 pith:GTTHIN2K submitted 2026-03-23 cs.CV

4DGS360: 360{deg} Gaussian Reconstruction of Dynamic Objects from a Single Video

Jae Won Jang , Yeonjin Chang , Wonsik Shin , Juhwan Cho , Nojun Kwak This is my paper
classification cs.CV
keywords monocular dynamic reconstruction360-degree reconstruction4D reconstructiondynamic Gaussian splatting3D point trackingocclusion handlingnovel view synthesisiPhone360 dataset
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.

Monocular dynamic reconstruction is ill-posed: a single video shows only one side of an object at a time, so the geometry of occluded regions is ambiguous. The paper argues that existing methods fail at extreme novel views because their 2D-native initialization overfits to surfaces visible in each training frame. 4DGS360 instead initializes dynamic Gaussians with AnchorTAP3D, a training-free 3D tracker that uses confident 2D track points as anchors to keep tracking stable for occluded points. This geometry-aware initialization lets the optimization recover coherent 360-degree reconstructions, and the paper reports state-of-the-art results on the new iPhone360 benchmark, the iPhone dataset, and DAVIS.

Core claim

4DGS360 claims that monocular reconstruction fails at extreme viewpoints because 2D-native initialization leaves occluded points without valid depth, overfitting visible surfaces. The fix is AnchorTAP3D, a training-free tracker that turns confident 2D tracks into 3D anchors and conditions a pretrained transformer-based 3D tracker on them, suppressing drift and preserving occluded geometry. Resulting 3D trajectories initialize hierarchical deformable Gaussians, and as-rigid-as-possible regularization yields coherent 360-degree reconstructions. It also introduces iPhone360, a dataset with test cameras up to 135 degrees from training, reporting state-of-the-art results on it, iPhone, and DAVIS.

What carries the argument

AnchorTAP3D is the paper's central mechanism: a unified 2D/3D tracker in which a 2D point tracker (BootsTAP-style) predicts correspondences and confidence, high-confidence points are unprojected to 3D via depth to form a set of anchors, and these anchors condition a pretrained transformer-based 3D tracker (TAPIP3D) inside a sliding temporal window — with no additional training. The anchors act as spatial-temporal constraints that suppress the error drift of naive 3D tracking and resolve the depth ambiguity of occluded regions. The resulting 3D trajectories initialize hierarchical deformable Gaussians, and an As-Rigid-As-Possible (ARAP) loss enforces local rigidity during optimization.

Load-bearing premise

The whole method relies on the untested assumption that a pretrained 3D point tracker can be conditioned on externally supplied 3D anchor points without retraining, and that this conditioning makes it better at tracking occluded points.

What would settle it

A concrete falsifier is to permute or zero out the anchor set (X_A in Eq. 11) and measure whether the 3D tracker's output changes; if predictions are invariant to anchors, the proposed mechanism has no causal effect. A quantitative falsifier is to compare training-free AnchorTAP3D against naive 3D tracking on a point-tracking benchmark with occlusion labels; if anchors do not improve occlusion accuracy, the claimed initialization advantage is not real.

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

If this is right

  • Occluded regions of a dynamic object can be reconstructed from a monocular video, provided those regions are visible in at least one frame.
  • A pretrained 3D point tracker can be steered by 2D-derived anchors without retraining, a capability that could transfer to other tasks that rely on point tracking.
  • The as-rigid-as-possible regularization becomes effective for occluded geometry once initialization is reliable, which is the paper's explanation for the gain over prior methods.
  • The iPhone360 benchmark establishes an evaluation protocol for extreme novel-view synthesis (70-135 degree test cameras) under realistic monocular capture.
  • Combining the geometry-preserving initialization with diffusion-based inpainting gives a better starting point for synthesizing fully unseen background regions.

Where Pith is reading between the lines

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

  • The anchor-conditioning mechanism may generalize beyond this pipeline: any pretrained 3D tracker could in principle be stabilized by 2D-derived anchors, extending occlusion-aware tracking to SLAM, motion capture, or video editing.
  • A natural stress test is to vary the angular gap between train and test cameras continuously and measure reconstruction quality; the benchmark's 70-135 degree range suggests a degradation curve that could guide capture protocols.
  • Because the method's depth comes from sensors or estimators, a testable extension is to swap in monocular depth predictors and measure how much of the 360-degree benefit survives depth error.
  • The paper does not show whether anchors help by providing global geometry or by locally conditioning attention; an ablation that corrupts anchor positions would separate these explanations.

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

4 major / 6 minor

Summary. The paper proposes 4DGS360, a monocular dynamic object reconstruction method based on deformable 3D Gaussians. Its central novelty is AnchorTAP3D, a 'training-free' extension of the pretrained 3D point tracker TAPIP3D that supposedly conditions on high-confidence 2D tracks as anchors to improve tracking in occluded regions, yielding better initialization for subsequent ARAP-regularized optimization. The paper also introduces iPhone360, a six-scene benchmark with test cameras placed 70–135 degrees from the training views, and reports state-of-the-art CLIP/LPIPS metrics against MoSca and HiMoR on iPhone360 and iPhone, plus qualitative DAVIS results.

Significance. If the anchor-conditioning mechanism works as claimed, the paper would address a well-known failure mode in monocular 4D reconstruction: the inability to reconstruct regions that are never directly visible in the training video. The problem is important, and the proposed iPhone360 dataset plausibly fills a real evaluation gap. However, the central mechanism is under-specified to the point of unreproducibility, and the quantitative evidence consists of small-scale, single-run comparisons with no error bars. The significance cannot be properly assessed from the manuscript as written.

major comments (4)
  1. [Sec. 3.2, Eq. (11)] The central claim is that a frozen TAPIP3D model can be conditioned on an externally supplied anchor set X_A without any training. The paper never states how X_A is injected into the transformer (cross-attention, extra tokens, prompting, or modification of attention masks), how the 2D and 3D trackers are synchronized, or why a pretrained model would respond to an input distribution it has not seen. Since AnchorTAP3D is the only source of the claimed 360-degree gain, the method is unreproducible as written. Please provide the exact network modification or an actual fine-tuning/training procedure and update the 'without additional training' statement accordingly.
  2. [Sec. 4, Tabs. 2–3] All quantitative tables report single-run means without error bars or significance tests, over only 5–6 scenes. The reported improvements (e.g., LPIPS 0.224 vs 0.326 on Goat) may be real, but without variance the state-of-the-art claim is not supported. Moreover, AnchorTAP3D itself is never evaluated quantitatively; Figs. 3 and 8 are only qualitative examples. Please add tracking metrics (e.g., 3D endpoint error, occlusion accuracy) and repeated-run statistics.
  3. [Sec. 4.1, iPhone360 evaluation] The iPhone360 protocol crops to a bounding box from the ground-truth dynamic-object mask and evaluates 'foreground-only rendered results.' This omits the background and may hide poor reconstruction of unseen background regions; it also makes the benchmark's 360-degree claim less direct. Please justify this choice, report full-image metrics as well, and clarify whether the mask comes from GT segmentation (which could leak object location). In addition, the dataset is not released with the paper, so the benchmark cannot be used by others.
  4. [Sec. 3.3 and Supplementary] Several implementation details are deferred to a 'supplementary' that is not included: loss weights, sliding-window length L and overlap, confidence threshold tau, number of Gaussians and nodes, and optimization schedule. More importantly, the anchor-conditioning architecture itself is unspecified. The limitation section correctly notes dependence on pretrained models, but that is not a substitute for the missing technical description of the central contribution.
minor comments (6)
  1. [Abstract, Sec. 1] Typo: 'a advanced' should be 'an advanced.' Also 'earns notable tracking results' is awkward; consider 'achieves state-of-the-art tracking performance.'
  2. [Sec. 2.3] The sentence 'Among 3D tracking models [13,14]' cites CoTracker3 and CoTracker, which are 2D trackers. Please correct the citations to 3D tracking works such as [53] and [60].
  3. [Eq. (12)] The ARAP loss has garbled formatting with nested norms and transforms; please clarify the notation, especially the second term involving T_j^{-1}.
  4. [Table 1 vs. Sec. 4.1] The maximum angular difference for iPhone360 is reported as '80–135' in Table 1 but '70–135' in the text. Please make consistent.
  5. [Fig. 5] The caption says 'MoSca w/ DIFIX3D+' but the text calls the method 'DIFIX3D+' inconsistently; clarify whether this is the same diffusion prior used for all baselines.
  6. [Sec. 4.1] 'DA VIS' should be 'DAVIS' in the section heading and text.

Circularity Check

0 steps flagged

No significant circularity: the method is an empirical pipeline evaluated on held-out views, and no prediction reduces to a fitted input by construction.

full rationale

The paper's derivation chain is an optimization-based reconstruction pipeline: 2D tracking produces high-confidence anchors (Eq. 10), which are used to condition the pretrained 3D tracker TAPIP3D (Eq. 11), yielding 3D trajectories that initialize Gaussians; then rendering losses and ARAP regularization are optimized against training views. None of these equations defines a target quantity in terms of itself, and no fitted parameter is renamed as a prediction. The anchor set is built from the same input tracks, but that is an architectural conditioning choice, not a circular reduction: the claimed output is the 3D trajectory, evaluated on test viewpoints not used in optimization. The SOTA claims are supported by comparisons on the external iPhone and DAVIS benchmarks, and although the new iPhone360 dataset is introduced by the authors, evaluation on it compares against independent baseline methods rather than predicting the methods' own fitted values. The only self-citation (Divcon-NeRF, ref. [19]) appears in related work and is not load-bearing. The manuscript's own limitations section acknowledges dependence on pretrained trackers and inability to synthesize invisible backgrounds, which is consistent with an empirical pipeline rather than a circular argument. The under-specification of how anchors are injected into the frozen TAPIP3D transformer is a reproducibility/correctness concern, not circularity, and cannot be scored as circular without a demonstrated reduction. Therefore, no circular step is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. It relies heavily on pretrained 2D/3D trackers and hand-chosen hyperparameters. The main load-bearing assumptions are that anchor conditioning works on a frozen 3D tracker, that depth/calibration inputs are reliable, and that the self-built iPhone360 benchmark validly measures 360° reconstruction.

free parameters (5)
  • 2D confidence threshold τ
    Eq. (9) uses a confidence mask to decide which 2D track points become 3D anchors; τ is not specified and controls the quality and quantity of anchors.
  • Sliding window length L and overlap = L=16, overlap=8
    Section 3.2 sets the temporal window for the 3D tracker; no ablation is shown to justify the choice.
  • Number of motion clusters B (k-means)
    Section 3.2 clusters trajectories into B groups to initialize motion bases; B is not specified or ablated.
  • Number of Gaussians N and nodes n1
    Initialization details specify random sampling of N trajectories and weighted sampling of n1 Gaussians; values are not given.
  • Loss weights λ_rgb, λ_mask, λ_depth, λ_2dtrack, λ_arap
    Eq. (13) defines a weighted sum of losses; the weights are not listed in the main text and are critical to the optimization balance.
axioms (5)
  • domain assumption Pretrained 2D trackers produce accurate correspondences with well-calibrated confidence scores.
    Eqs. (7)–(9) rely on 2D tracker confidence to select anchors; if confidence is miscalibrated, anchors will be wrong and the 3D tracker will be misled.
  • ad hoc to paper A pretrained 3D tracker (TAPIP3D) can be conditioned on externally supplied 3D anchor points without retraining and without performance degradation.
    Eq. (11) introduces the anchor-conditioned model P^Anchor3D and the paper claims 'without additional training,' but no architectural details or experiments validate that this conditioning works as intended.
  • domain assumption Per-frame depth maps and camera parameters are accurate enough for unprojecting 2D tracks to 3D anchors.
    Eq. (8) unprojects 2D points using depth and camera parameters; the paper acknowledges in Section 2.3 that 3D tracking is sensitive to depth and calibration errors, yet relies on these inputs throughout.
  • domain assumption Dynamic objects can be modeled as locally rigid deformations of canonical Gaussians.
    The ARAP loss in Eq. (12) assumes local rigidity; this fails for topology changes or fluid-like motion, which are not present in the tested scenes.
  • ad hoc to paper The iPhone360 benchmark, with mask-cropped perceptual metrics and test cameras 70–135° apart, is a valid measure of 360° reconstruction quality.
    The evaluation protocol described in Section 4.1 is designed by the authors; no external validation shows that the cropping procedure or the perceptual metrics track 360° geometric fidelity.

pith-pipeline@v1.3.0-alltime-deepseek · 13806 in / 8714 out tokens · 83090 ms · 2026-08-02T17:40:12.324840+00:00 · methodology

0 comments
read the original abstract

We introduce 4DGS360, a diffusion-free framework for 360$^{\circ}$ dynamic object reconstruction from casual monocular video. Existing methods often fail to reconstruct consistent 360$^{\circ}$ geometry, as their heavy reliance on 2D-native priors causes initial points to overfit to visible surface in each training view. 4DGS360 addresses this challenge through a advanced 3D-native initialization that mitigates the geometric ambiguity of occluded regions. Our proposed 3D tracker, AnchorTAP3D, produces reinforced 3D point trajectories by leveraging confident 2D track points as anchors, suppressing drift and providing reliable initialization that preserves geometry in occluded regions. This initialization, combined with optimization, yields coherent 360$^{\circ}$ 4D reconstructions. We further present iPhone360, a new benchmark where test cameras are placed up to 135$^{\circ}$ apart from training views, enabling 360$^{\circ}$ evaluation that existing datasets cannot provide. Experiments show that 4DGS360 achieves state-of-the-art performance on the iPhone360, iPhone, and DAVIS datasets, both qualitatively and quantitatively.

Figures

Figures reproduced from arXiv: 2603.21618 by Jae Won Jang, Juhwan Cho, Nojun Kwak, Wonsik Shin, Yeonjin Chang.

Figure 1
Figure 1. Figure 1: Our 4DGS360 model reconstructs 360° dynamic object geometry from a monoc￾ular video and produces higher-fidelity renderings than prior methods under both or￾dinary and extreme novel-view conditions. We also introduce the iPhone360 dataset, which captures dynamic objects in a monocular setting and includes test cameras far outside the training trajectory for 360° evaluation. leveraging pretrained 2D point t… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of 4DGS360 (a) illustrates the initialization stage using Anchor￾TAP3D, where reliable geometry can be obtained even when tracked points become occluded. This contrasts with 2D-based models, which cannot estimate depth for oc￾cluded points. (b) presents the overall architecture of AnchorTAP3D, and (c) highlights the difference between tracking with and without anchors, showing that anchors effec￾t… view at source ↗
Figure 3
Figure 3. Figure 3: Tracking Comparison. We compare the naive 3D tracking model TAPIP3D [60] with our anchor-based AnchorTAP3D (Ours). At the target time af￾ter undergoing rotation and occlusion, AnchorTAP3D produces noticeably more stable tracking results. This indicates that anchor-guided tracking is advantageous under chal￾lenging conditions such as occlusion. where Π denotes the camera projection. After projection, each 3… view at source ↗
Figure 4
Figure 4. Figure 4: iPhone360 Qualitative Comparison. In (a), Novel-view GTs are masked and model outputs are shown as renderings of dynamic Gaussians. Blue pyramids denote training cameras, and reds denote test camera locations in the Zoom-out. (b) shows bullet-time rendering results across 360° at a fixed time instant [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results with Diffusion Prior. Our method outperforms baselines when combined with DIFIX3D+ [52] [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: iPhone dataset Qualitative result. Backpack scene shows that our model faithfully reconstructs unseen backside regions, clearly outperforming prior methods. achieves better 360° 4D reconstruction, particularly reconstructing the geometry of occluded regions at extreme viewpoints, compared to ‘w/o 3D init’ and ‘w/o Anchor’. The ‘w/o 3D init’ variant follows prior models [23, 46] in initializing with unproje… view at source ↗
Figure 7
Figure 7. Figure 7: DAVIS dataset Qualitative Comparison. Results show that our model maintains coherent geometry and recovers occluded parts more reliably than previous approaches. GT w/o 3D init w/o Anchor Ours GT w/o 3D init w/o Anchor Ours [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Ablations. Qualitative results for different initialization methods. ‘w/o 3D init’ directly unprojects 2D tracks, ‘w/o Anchor’ leverages naive 3D point tracking model [60], and Ours uses AnchorTAP3D at initialization. The left and right scenes correspond to the walk-around and jelly sequences from iPhone360, respectively. 5 Limitations and Conclusion Limitations. While our method demonstrates improved 360°… 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

62 extracted references · 1 canonical work pages

  1. [1]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini- vasan, P.P.: Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5855–5864 (2021)

    2021

  2. [2]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip- nerf 360: Unbounded anti-aliased neural radiance fields. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5470–5479 (2022)

    2022

  3. [3]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Zip-nerf: Anti-aliased grid-based neural radiance fields. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19697–19705 (2023)

    2023

  4. [4]

    In: CVPR (2024)

    Charatan, D., Li, S., Tagliasacchi, A., Sitzmann, V.: pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In: CVPR (2024)

    2024

  5. [5]

    In: Glober- son, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C

    Chen, Y., Zheng, C., Xu, H., Zhuang, B., Vedaldi, A., Cham, T.J., Cai, J.: Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In: Glober- son, A., Mackey, L., Belgrave, D., Fan, A., Paquet, U., Tomczak, J., Zhang, C. (eds.) Advances in Neural Information Processing Systems. vol. 37, pp. 107064–107086. Curran Associates, Inc. (2024).https://do...

    2024

  6. [6]

    Advances in Neural Information Processing Systems37, 96181–96206 (2024)

    Chu, W.H., Ke, L., Fragkiadaki, K.: Dreamscene4d: Dynamic multi-object scene generation from monocular videos. Advances in Neural Information Processing Systems37, 96181–96206 (2024)

    2024

  7. [7]

    In: Proceedings of the Asian Conference on Com- puter Vision (ACCV)

    Doersch, C., Luc, P., Yang, Y., Gokay, D., Koppula, S., Gupta, A., Heyward, J., Rocco, I., Goroshin, R., Carreira, J.a., Zisserman, A.: Bootstap: Bootstrapped training for tracking-any-point. In: Proceedings of the Asian Conference on Com- puter Vision (ACCV). pp. 3257–3274 (December 2024)

    2024

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Doersch, C., Yang, Y., Vecerik, M., Gokay, D., Gupta, A., Aytar, Y., Carreira, J., Zisserman, A.: Tapir: Tracking any point with per-frame initialization and tem- poral refinement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10061–10072 (2023)

    2023

  9. [9]

    ACM Transactions on Graphics (TOG)43(4), 1–13 (2024)

    Duckworth, D., Hedman, P., Reiser, C., Zhizhin, P., Thibert, J.F., Lučić, M., Szeliski, R., Barron, J.T.: Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG)43(4), 1–13 (2024)

    2024

  10. [10]

    In: NeurIPS (2022)

    Gao, H., Li, R., Tulsiani, S., Russell, B., Kanazawa, A.: Monocular dynamic view synthesis: A reality check. In: NeurIPS (2022)

    2022

  11. [11]

    In: European Conference on Computer Vision (2022),https://api.semanticscholar.org/CorpusID:248069518

    Harley, A.W., Fang, Z., Fragkiadaki, K.: Particle video revisited: Tracking through occlusions using point trajectories. In: European Conference on Computer Vision (2022),https://api.semanticscholar.org/CorpusID:248069518

    2022

  12. [12]

    Hu, Y., Liu, Z., Shao, J., Lin, Z., Zhang, J.: Eva-gaussian: 3d gaussian-based real- time human novel view synthesis under diverse camera settings (2024),https: //arxiv.org/abs/2410.01425

    Pith/arXiv arXiv 2024

  13. [13]

    In: Proc

    Karaev, N., Makarov, I., Wang, J., Neverova, N., Vedaldi, A., Rupprecht, C.: Co- tracker3: Simpler and better point tracking by pseudo-labelling real videos. In: Proc. arXiv:2410.11831 (2024)

    Pith/arXiv arXiv 2024

  14. [14]

    In: Proc

    Karaev, N., Rocco, I., Graham, B., Neverova, N., Vedaldi, A., Rupprecht, C.: Cotracker: It is better to track together. In: Proc. ECCV (2024) 16 J.Jang et al

    2024

  15. [15]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4), 139–1 (2023)

    2023

  16. [16]

    Advances in Neural Information Processing Systems37, 129209– 129226 (2024)

    Kim, M., Lim, J., Han, B.: 4d gaussian splatting in the wild with uncertainty-aware regularization. Advances in Neural Information Processing Systems37, 129209– 129226 (2024)

    2024

  17. [17]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4015–4026 (October 2023)

    2023

  18. [18]

    NeurIPS25(2012)

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. NeurIPS25(2012)

    2012

  19. [19]

    arXiv preprint arXiv:2503.12947 (2025)

    Lee, I., Jang, J.W., Seo, S., Kwak, N.: Divcon-nerf: Generating augmented rays with diversity and consistency for few-shot view synthesis. arXiv preprint arXiv:2503.12947 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    arXiv preprint arXiv:2312.02135 (2023)

    Lee, Y.C., Zhang, Z., Blackburn-Matzen, K., Niklaus, S., Zhang, J., Huang, J.B., Liu, F.: Fast view synthesis of casual videos with soup-of-planes. arXiv preprint arXiv:2312.02135 (2023)

    Pith/arXiv arXiv 2023

  21. [21]

    In: Proceedings of the Com- puter Vision and Pattern Recognition Conference

    Lei, J., Weng, Y., Harley, A.W., Guibas, L., Daniilidis, K.: Mosca: Dynamic gaus- sian fusion from casual videos via 4d motion scaffolds. In: Proceedings of the Com- puter Vision and Pattern Recognition Conference. pp. 6165–6177 (2025)

    2025

  22. [22]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural scene flow fields for space-time view synthesis of dynamic scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6498–6508 (2021)

    2021

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Liang, Y., Xu, T., Kikuchi, Y.: Himor: Monocular deformable gaussian reconstruc- tion with hierarchical motion representation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 886–895 (2025)

    2025

  24. [24]

    In: European Conference on Computer Vision

    Liu, T., Wang, G., Hu, S., Shen, L., Ye, X., Zang, Y., Cao, Z., Li, W., Liu, Z.: Mvsgaussian: Fast generalizable gaussian splatting reconstruction from multi-view stereo. In: European Conference on Computer Vision. pp. 37–53. Springer (2025)

    2025

  25. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y., Kopf, J., Huang, J.B.: Robust dynamic radiance fields. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13–23 (2023)

    2023

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lu, T., Yu, M., Xu, L., Xiangli, Y., Wang, L., Lin, D., Dai, B.: Scaffold-gs: Struc- tured 3d gaussians for view-adaptive rendering. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 20654–20664 (2024)

    2024

  27. [27]

    In: 2024 International Conference on 3D Vision (3DV)

    Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In: 2024 International Conference on 3D Vision (3DV). pp. 800–809. IEEE (2024)

    2024

  28. [28]

    In: SIGGRAPH Asia 2024 Conference Papers

    Mallick, S.S., Goel, R., Kerbl, B., Steinberger, M., Carrasco, F.V., De La Torre, F.: Taming 3dgs: High-quality radiance fields with limited resources. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024)

    2024

  29. [29]

    Commu- nications of the ACM65(1), 99–106 (2021)

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: Nerf: Representing scenes as neural radiance fields for view synthesis. Commu- nications of the ACM65(1), 99–106 (2021)

    2021

  30. [30]

    In: European Conference on Computer Vision

    Paliwal, A., Ye, W., Xiong, J., Kotovenko, D., Ranjan, R., Chandra, V., Kalantari, N.K.: Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In: European Conference on Computer Vision. pp. 19–37. Springer (2024)

    2024

  31. [31]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin- Brualla, R.: Nerfies: Deformable neural radiance fields. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5865–5874 (2021) 4DGS360 17

    2021

  32. [32]

    ACM Trans

    Park,K.,Sinha,U.,Hedman,P.,Barron,J.T.,Bouaziz,S.,Goldman,D.B.,Martin- Brualla, R., Seitz, S.M.: Hypernerf: A higher-dimensional representation for topo- logically varying neural radiance fields. ACM Trans. Graph.40(6) (dec 2021)

    2021

  33. [33]

    Pont-Tuset, J., Perazzi, F., Caelles, S., Arbeláez, P., Sorkine-Hornung, A., Gool, L.V.: The 2017 davis challenge on video object segmentation (2018),https:// arxiv.org/abs/1704.00675

    Pith/arXiv arXiv 2017

  34. [34]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Pumarola, A., Corona, E., Pons-Moll, G., Moreno-Noguer, F.: D-nerf: Neural ra- diance fields for dynamic scenes. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10318–10327 (2021)

    2021

  35. [35]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  36. [36]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

    2022

  37. [37]

    Psychometrika31(1), 1–10 (1966).https://doi.org/10.1007/BF02289451

    Schönemann, P.H.: A generalized solution of the orthogonal procrustes problem. Psychometrika31(1), 1–10 (1966).https://doi.org/10.1007/BF02289451

    doi:10.1007/bf02289451 1966

  38. [38]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Seo, S., Chang, Y., Kwak, N.: Flipnerf: Flipped reflection rays for few-shot novel view synthesis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22883–22893 (2023)

    2023

  39. [39]

    In: European Conference on Computer Vision (ECCV) (2024)

    Shen, Y., Ceylan, D., Guerrero, P., Xu, Z., Mitra, N., Wang, S., Frühstück, A.: Supergaussian: Repurposing video models for 3d super resolution. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  40. [40]

    In: Proceedings of the Fifth Eurographics Symposium on Geometry Processing

    Sorkine, O., Alexa, M.: As-rigid-as-possible surface modeling. In: Proceedings of the Fifth Eurographics Symposium on Geometry Processing. p. 109–116. SGP ’07, Eurographics Association, Goslar, DEU (2007)

    2007

  41. [41]

    In: SIGGRAPH Asia 2024 Conference Papers

    Stearns, C., Harley, A., Uy, M., Dubost, F., Tombari, F., Wetzstein, G., Guibas, L.: Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024)

    2024

  42. [42]

    Team, S.D., Chen, X., Chu, F.J., Gleize, P., Liang, K.J., Sax, A., Tang, H., Wang, W., Guo, M., Hardin, T., Li, X., Lin, A., Liu, J., Ma, Z., Sagar, A., Song, B., Wang, X., Yang, J., Zhang, B., Dollár, P., Gkioxari, G., Feiszli, M., Malik, J.: Sam 3d: 3dfy anything in images (2025),https://arxiv.org/abs/2511.16624

    Pith/arXiv arXiv 2025

  43. [43]

    In: Proceedings of the 35th International Conference on Neural Informa- tion Processing Systems

    Teed, Z., Deng, J.: Droid-slam: deep visual slam for monocular, stereo, and rgb-d cameras. In: Proceedings of the 35th International Conference on Neural Informa- tion Processing Systems. NIPS ’21, Curran Associates Inc., Red Hook, NY, USA (2021)

    2021

  44. [44]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Verbin, D., Hedman, P., Mildenhall, B., Zickler, T., Barron, J.T., Srinivasan, P.P.: Ref-nerf: Structured view-dependent appearance for neural radiance fields. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  45. [45]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Wang, C., Chai, M., He, M., Chen, D., Liao, J.: Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 3835–3844 (2022)

    2022

  46. [46]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wang, Q., Ye, V., Gao, H., Zeng, W., Austin, J., Li, Z., Kanazawa, A.: Shape of motion: 4d reconstruction from a single video. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 9660–9672 (2025)

    2025

  47. [47]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B., Revaud, J.: Dust3r: Geometric 3d vision made easy. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20697–20709 (June 2024) 18 J.Jang et al

    2024

  48. [48]

    IEEE Transactions on Image Processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A., Sheikh, H., Simoncelli, E.: Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing 13(4), 600–612 (2004)

    2004

  49. [49]

    Wenbo, C., Ligang, L.: Deblur-gs: 3d gaussian splatting from camera motion blurred images. Proc. ACM Comput. Graph. Interact. Tech. (Proceedings of I3D 2024)7(1) (2024).https://doi.org/10.1145/3651301,http://doi.acm.org/10. 1145/3651301

    work page doi:10.1145/3651301 2024

  50. [50]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wu, D., Liu, F., Hung, Y.H., Qian, Y., Zhan, X., Duan, Y.: 4d-fly: Fast 4d recon- struction from a single monocular video. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16663–16673 (2025)

    2025

  51. [51]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.: 4d gaussian splatting for real-time dynamic scene rendering. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20310– 20320 (2024)

    2024

  52. [52]

    In: Conf

    Wu, J.Z., Zhang, Y., Turki, H., Ren, X., Gao, J., Shou, M.Z., Fidler, S., Gojcic, Z., Ling, H.: Difix3d+: Improving 3d reconstructions with single-step diffusion models. In: Conf. Comput. Vis. Pattern Recog. pp. 26024–26035 (2025)

    2025

  53. [53]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2025),https: //arxiv.org/abs/2507.12462

    Xiao, Y., Wang, J., Xue, N., Karaev, N., Makarov, Y., Kang, B., Zhu, X., Bao, H., Shen, Y., Zhou, X.: Spatialtrackerv2: 3d point tracking made easy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (2025),https: //arxiv.org/abs/2507.12462

    Pith/arXiv arXiv 2025

  54. [54]

    Yang, J., Gao, M., Li, Z., Gao, S., Wang, F., Zheng, F.: Track anything: Segment anything meets videos (2023)

    2023

  55. [55]

    In: CVPR (2024)

    Yang, L., Kang, B., Huang, Z., Xu, X., Feng, J., Zhao, H.: Depth anything: Un- leashing the power of large-scale unlabeled data. In: CVPR (2024)

    2024

  56. [56]

    Advances in Neural Information Processing Systems37, 21875–21911 (2024)

    Yang, L., Kang, B., Huang, Z., Zhao, Z., Xu, X., Feng, J., Zhao, H.: Depth anything v2. Advances in Neural Information Processing Systems37, 21875–21911 (2024)

    2024

  57. [57]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20331– 20341 (2024)

    2024

  58. [58]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Yu, A., Ye, V., Tancik, M., Kanazawa, A.: pixelnerf: Neural radiance fields from one or few images. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4578–4587 (2021)

    2021

  59. [59]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yu, Z., Chen, A., Huang, B., Sattler, T., Geiger, A.: Mip-splatting: Alias-free 3d gaussian splatting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 19447–19456 (June 2024)

    2024

  60. [60]

    arXiv preprint arXiv:2504.14717 (2025)

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

    arXiv 2025

  61. [61]

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectivenessofdeepfeaturesasaperceptualmetric.In:Conf.Comput.Vis.Pattern Recog. pp. 586–595 (2018)

    2018

  62. [62]

    In: European Conference on Computer Vision (ECCV) (2024)

    Zhao, L., Wang, P., Liu, P.: Bad-gaussians: Bundle adjusted deblur gaussian splat- ting. In: European Conference on Computer Vision (ECCV) (2024)

    2024