Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

AttentionGS: Towards Initialization-Free 3D Gaussian Splatting via Structural Attention

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read AttentionGS trains 3D Gaussian Splatting from random point clouds, without SfM initialization, using edge-weighted geometry, opacity-weighted densification, and scheduled appearance losses; on Mip-NeRF 360 it reports a 3.84 PSNR gain over…

desk verdict Plausible and potentially useful, but the SLV-baseline ambiguity and a direct table contradiction leave the headline 3.84 dB gain unestablished. read the letter →

arxiv 2506.23611 v1 pith:JYOTDEVF submitted 2025-06-30 cs.CV

Ziao Liu , Zhenjia Li , Yifeng Shi , Xiangang Li This is my paper
classification cs.CV
keywords 3DGaussianSplattingrandominitializationStructure-from-Motion-freereconstructiongeometricattentionCannyedgeweightingopacity-weighteddensificationappearancenovelviewsynthesis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that 3D Gaussian Splatting can be trained directly from a random point cloud, without the Structure-from-Motion point clouds that normally initialize it, if the loss and densification are reweighted to favor geometry early. On the Mip-NeRF 360 and LLFF benchmarks, the proposed AttentionGS reports large gains over 3DGS, Pixel-GS, and Mip-splatting under random and noisy initialization, and roughly matches 3DGS initialized with COLMAP. If true, this removes a costly preprocessing dependency and makes 3DGS usable in texture-poor or constrained-view settings where SfM fails. The method's claim is that a Canny-edge-weighted geometric loss, a transmittance-weighted densification criterion, and a channel-wise appearance-weighted loss together let random Gaussians find scene structure before they settle into fitting color.

What carries the argument

The load-bearing object is the geometric attention weight $w(x,y)$, computed as the max-normalized absolute difference between Canny-edge-enhanced ground-truth and rendered RGB images; this weight turns the standard L1 loss into an edge-focused L1 loss during early training. Around it sit two supporting mechanisms: the opacity-weighted gradient criterion, which multiplies each Gaussian's NDC-position gradient by its transmittance before comparing against the split threshold, and the appearance attention weight $w_{app}(x,y,c)$, a per-channel normalized RGB error that is blended in through a sigmoid schedule $f(i)$. Together they implement a curriculum: geometry first, appearance second.

What would settle it

Run the same training twice on one Mip-NeRF 360 scene, once with the Canny-based weight $w(x,y)$ detached from autograd and once with gradients forced through it; if the two runs give nearly identical PSNR, the geometric attention is effectively a reweighted L1 loss and the claimed structural mechanism reduces to static edge weighting. A second check: train 3DGS with SLV initialization alone and compare to AttentionGS; if SLV alone closes most of the gap, the attention modules are not the cause of the reported gains.

Watch

Extended reading notes

Core claim

AttentionGS's central claim is that the failure of 3DGS under random initialization is not a fundamental limitation but a training-dynamics problem: randomly placed Gaussians prefer to adjust color instead of position, so the optimizer overfits appearance before geometry forms. The paper proposes to reorder those dynamics by weighting the L1 reconstruction loss by a normalized Canny-edge difference between rendered and ground-truth images, concentrating early optimization on boundaries; replacing the 3DGS densification statistic with a transmittance-weighted gradient so visible surface Gaussians split before occluded ones; and activating a separate per-channel appearance-weighted L1 loss on a sigmoid schedule after structure has formed. With the Sparse-Large-Variance random initialization, the paper reports an average PSNR of 28.73 versus 24.89 for 3DGS on Mip-NeRF 360 under random initialization, with similar gains on LLFF, and performance comparable to 3DGS initialized with COLMAP.

Load-bearing premise

The geometric attention weight comes from an edge-detection step and a max-normalization step, and the paper never states whether that weight is treated as a fixed constant or as part of the differentiated computation, so the reported structural gains rest on an unstated choice about what gradients actually flow.

Editorial extensions

If this is right

  • Under random point cloud initialization, AttentionGS reports a PSNR gain of 3.84, an SSIM gain of 0.088, and an LPIPS reduction of 0.084 over 3DGS on Mip-NeRF 360.
  • Under noisy initialization, the reported gains over 3DGS are 1.70 PSNR, 0.048 SSIM, and 0.049 LPIPS, showing the method tolerates degraded seeds, not only random ones.
  • On LLFF, which has few frames and narrow baselines, AttentionGS with random initialization reaches 23.18 PSNR versus 10.68 for random-initialized 3DGS and 22.51 for COLMAP-initialized 3DGS.
  • The ablation study attributes the gains to all three components: geometric attention alone adds 1.78 PSNR, opacity-weighted gradients add another 1.41 PSNR, and appearance attention adds 0.65 PSNR.
  • Because the method no longer depends on SfM point clouds, the practical pipeline for novel view synthesis from images could drop the COLMAP preprocessing stage in settings where it is slow or unreliable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implication the paper leaves implicit is that the sigmoid schedule's slope and decay node are global hyperparameters; per-scene scheduling might close the remaining gap to COLMAP initialization.
  • The Canny-edge weighting biases optimization toward intensity boundaries, so the method should be tested on genuinely texture-deficient scenes where SfM fails, rather than only on edge-rich benchmark scenes.
  • The transmittance-weighted densification criterion is independent of the loss weighting and could be transplanted into other Gaussian-splatting variants that already have good initialization, potentially improving surface reconstruction there as well.
  • A direct comparison of 3DGS with SLV initialization alone, without the attention modules, would quantify how much of the random-init gain comes from the initialization strategy rather than the new losses.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. AttentionGS proposes three modifications to 3D Gaussian Splatting to enable training from random point-cloud initialization: a geometric attention loss that reweights L1 photometric error by a Canny-edge-difference map (Eqs. 3-6), an opacity/transmittance-weighted densification criterion (Eq. 7), and a channel-wise appearance attention loss scheduled by a sigmoid function (Eqs. 8-11). The method is evaluated on Mip-NeRF 360 and LLFF under random, noisy, and COLMAP initialization, with ablations in Table 3. The central claim is that AttentionGS improves PSNR by 3.84 dB over 3DGS under random initialization and performs comparably to COLMAP-initialized 3DGS.

Significance. If reproducible, the result is practically valuable because it would decouple 3DGS from SfM point clouds in many scenarios, and the proposed components are simple and interpretable. The paper has clear strengths: it uses standard benchmarks, the ablation in Table 3 is monotone, and the comparison includes relevant baselines. The circularity concern raised in the stress-test does not land: the attention weights are defined from current prediction errors as a training loss reweighting, not as fitted constants used to derive a predicted result. However, the current evidence does not yet establish the claimed gains because of an initialization-protocol ambiguity, an internal table inconsistency, and missing implementation details.

major comments (4)
  1. [§4.2, Tables 1 and 2] The numbers for Ours(Random) are inconsistent across the two tables. Table 1 reports Ours(Random) on Mip-NeRF 360 as PSNR 28.73, SSIM 0.866, LPIPS 0.083, while Table 2 reports Ours(Random) as 28.79, 0.867, 0.078 — exactly the values Table 1 lists for Ours under the Noise condition. Since the headline 3.84 dB improvement is computed from Table 1 (28.73 − 24.89), the manuscript must identify which table is correct, report per-seed variance, and regenerate the affected rows. Without this, the main quantitative claim is not reproducible.
  2. [§4.1] The sentence "we employ SLV for random point cloud initialization in all subsequent experiments" does not state whether the 3DGS, Pixel-GS, and Mip-splatting baselines also use SLV with the same random seeds. Because SLV (ref. [9]) is itself an initialization method engineered to make random initialization viable, a comparison in which only AttentionGS uses SLV would confound the effect of the proposed geometric/texture losses with the effect of SLV. Please specify the initialization protocol for every method and, if baselines do use SLV, state it explicitly.
  3. [§3.2, Eqs. (3)-(6)] The paper does not state whether the Canny edge map and the max-normalization in Eq. (5) are detached from the computation graph before L_geo is used for training. If w(x,y) is a constant (stop-gradient applied), L_geo is a reweighted L1 loss; if w is differentiated, the gradient is undefined wherever the max-normalization denominator is zero and non-differentiable through NonMaxSuppression. The same issue affects Eq. (8). Please state the gradient handling, add an epsilon or detach operator, and specify the Canny thresholds and kernel K for reproducibility.
  4. [§4.2, Tables 1 and 3] The paper averages over eight Mip-NeRF 360 scenes but reports per-scene numbers only for Bicycle, Bonsai, and Stump; Garden, Kitchen, Room, Counter, and Flowers are missing, and no error bars or multiple-seed statistics are provided. The LLFF results in Table 2 also lack per-scene listings, and the 3DGS(Random) PSNR of 10.68 is so far below the COLMAP baseline that the evaluation protocol should be verified. Please provide the full per-scene tables and variance information.
minor comments (7)
  1. [References] Reference [20] is cited for the LLFF dataset, but [20] is the iNeRF paper; the LLFF dataset should cite Mildenhall et al. (Local Light Field Fusion). The manuscript also says LLFF comprises 7 scenes, whereas the standard LLFF real-scene set contains 8 scenes (Fern, Flower, Fortress, Horns, Leaves, Orchids, Room, Trex); please clarify.
  2. [Throughout] The acronyms "Mip-GS", "MipGS", and "Mip-splatting" are used interchangeably; choose one consistent name.
  3. [§3.4, Eq. (10)] Equation (10) introduces hyperparameters s and m, but only m is reported (0.25) and s is never specified, nor is a sensitivity analysis provided.
  4. [§3.3, Eq. (7)] In Eq. (7), the denominator sum of transmittances can be zero for occluded Gaussians; the paper should state how this case is handled.
  5. [§4.2, Figure 3] Figure 3 would be easier to interpret if the training and testing PSNR curves were labeled with markers or a legend; the text does not identify which curves correspond to which condition.
  6. [§3.4] The claim that spherical harmonics have "reached a relatively high order" at the appearance stage is not supported by implementation details; report the SH basis schedule.
  7. [§4.4, Table 3] The ablation is monotone, which is good, but the row labels "3DGS+Geo+Opacity-grad" and "Ours" do not make explicit that Ours adds the Appearance Attention module; label the last row as "+Appearance" to match the text.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: claimed gains are external-benchmark results of a loss/densification weighting scheme; the only self-citation (URS-NeRF, ref [19]) is non-load-bearing related work, and SLV (ref [9]) is an external prior method openly credited.

full rationale

AttentionGS is an empirical loss- and densification-weighting scheme, and its central claims are validated on held-out views of external benchmarks (Mip-NeRF 360, LLFF) rather than derived from fitted constants. The three modules—Canny-edge-difference reweighting of the L1 loss (Eqs. 3–6), transmittance-weighted densification (Eq. 7), and per-channel error-reweighted L1 with a sigmoid activation schedule (Eqs. 8–11)—are per-iteration training heuristics whose weights are functions of the current reconstruction; this is the ordinary structure of any reweighted loss, not a quantity defined in terms of the quantity it purports to explain. No predicted constant or closed-form result reduces algebraically to a fitted parameter, and the reported 3.84 dB / 0.088 / 0.084 gains are test-set comparisons. The only self-citation is ref [19] (URS-NeRF, whose authors include the present first author), used in a related-work sentence ("other studies aim to improve rendering quality by refining neural network architectures to enhance image detail and realism [2, 4, 19]"); it supports no central quantity and is not invoked to forbid alternatives. SLV initialization (ref [9]) is an external, openly credited prior method, so adopting it is neither ansatz-smuggling nor self-citation; however, the paper's failure to state whether the Table 1 baselines also receive SLV is a comparison-protocol correctness risk, not circularity, and the Table 1 versus Table 2 discrepancy for Ours(Random) on Mip-NeRF 360 (28.73 vs. 28.79) is a reproducibility concern. The Canny/max-normalization differentiability ambiguity in Eq. 5 affects gradient correctness, not the derivation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The ledger is light: the method introduces no new physical or mathematical entities and only one new internal criterion (transmittance-weighted densification). The main unstated costs are the unspecified hyperparameters (s, Canny thresholds, kernel K) and the unresolved treatment of non-differentiable Canny operations in the loss. All experiments assume known camera poses and the SLV initialization from [9].

free parameters (4)
  • s (steepness in Eq. 10)
    Controls the schedule of appearance attention activation; no value is given in the paper.
  • Canny edge detector thresholds
    Low and high thresholds for Canny are standard hyperparameters but not specified in Section 3.2.
  • outward convolution kernel K (Eq. 4)
    The kernel that expands edge influence is left undefined.
  • m (decay node in Eq. 10) = 0.25
    Set in the implementation details; determines when appearance attention gains influence.
assumptions (4)
  • domain assumption Canny edge extraction is differentiable or detached without disturbing gradient flow
    The geometric loss L_geo in Eq. 6 depends on w(x,y), which depends on the rendered RGB through Eqs. 3-5. The paper does not explain how the non-differentiable Canny/NonMaxSuppression and max-normalization are handled during backpropagation.
  • domain assumption Camera poses are known and accurate
    The method only removes point-cloud initialization; it relies on poses from the dataset or COLMAP in all experiments.
  • domain assumption SLV initialization provides a suitable start distribution for Gaussian optimization
    All random-init experiments use the SLV strategy from [9], but the paper does not analyze robustness to other random initializations.
  • ad hoc to paper Max-normalization denominators in Eqs. 5 and 8 are non-zero
    A perfect edge or color match gives a zero denominator; no fallback is specified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AttentionGS: Towards Initialization-Free 3D Gaussian Splatting via Structural Attention." pith.science (2026). https://pith.science/paper/JYOTDEVF

@misc{pith2026250623611,
  author       = {Pith},
  title        = {Pith review of: AttentionGS: Towards Initialization-Free 3D Gaussian Splatting via Structural Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JYOTDEVF}},
  note         = {Machine review of arXiv:2506.23611}
}
read the original abstract

3D Gaussian Splatting (3DGS) is a powerful alternative to Neural Radiance Fields (NeRF), excelling in complex scene reconstruction and efficient rendering. However, it relies on high-quality point clouds from Structure-from-Motion (SfM), limiting its applicability. SfM also fails in texture-deficient or constrained-view scenarios, causing severe degradation in 3DGS reconstruction. To address this limitation, we propose AttentionGS, a novel framework that eliminates the dependency on high-quality initial point clouds by leveraging structural attention for direct 3D reconstruction from randomly initialization. In the early training stage, we introduce geometric attention to rapidly recover the global scene structure. As training progresses, we incorporate texture attention to refine fine-grained details and enhance rendering quality. Furthermore, we employ opacity-weighted gradients to guide Gaussian densification, leading to improved surface reconstruction. Extensive experiments on multiple benchmark datasets demonstrate that AttentionGS significantly outperforms state-of-the-art methods, particularly in scenarios where point cloud initialization is unreliable. Our approach paves the way for more robust and flexible 3D Gaussian Splatting in real-world applications.

Figures

Figures reproduced from arXiv: 2506.23611 by the authors.

Figure 1
Figure 1. The comparison between 3DGS and our method in terms of structural recovery is illustrated, with both approaches [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of AttentionGS. Geometric Attention Module, our method can achieve more efficient structural recovery with random point cloud initialization. 3.3 Opacity-weighted Gradients Strategy To further augment the effectiveness of geometric recovery, we have refined the densification strategy of 3DGS. The original 3DGS calculates cumulative gradients using viewpoint-based weighting. While this approach effectively i… view at source ↗
Figure 3
Figure 3. PSNR Evolution in Bonsai Scene Under Random [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison between Attention-GS (Ours) and 3DGS on the Mip-NeRF 360 and LLFF Dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Pano2World: End-to-End 3D Generation via Unified Multi-View Sequences

    cs.CV 2026-07 unverdicted novelty 6.0 of 10

    Pano2World generates an explorable 3D Gaussian scene directly from a single indoor panorama via coarse proxy rendering, view-aware joint denoising, and a latent feature adapter.

  2. You Only Gaussian Once: Controllable 3D Gaussian Splatting for Ultra-Densely Sampled Scenes

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    YOGO reformulates stochastic 3D Gaussian Splatting into a deterministic budget-aware system and supplies an ultra-dense dataset to enforce physical fidelity over viewpoint interpolation.

Reference graph

Works this paper leans on

29 extracted references · 15 canonical work pages · cited by 2 Pith papers

  1. [9]

    Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Se- ungryong Kim. 2024. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413 (2024)

    arXiv 2024

  2. [1]

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. 2011. Building rome in a day. Commun. ACM 54, 10 (2011), 105–112

    2011

  3. [2]

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. 2021. Mip-nerf: A multiscale repre- sentation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 5855–5864

    work page 2021

  4. [3]

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. 2022. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5470–5479

    2022

  5. [4]

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. 2023. Zip-NeRF: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706 (2023)

    arXiv 2023

  6. [5]

    Yu Chen and Gim Hee Lee. 2023. DBARF: Deep Bundle-Adjusting Generalizable Neural Radiance Fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 24–34

    work page 2023

  7. [6]

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. 2025. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In European Conference on Computer Vision. Springer, 370–386

    work page 2025

  8. [7]

    Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Ragu- ram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazeb- nik, et al. 2010. Building rome on a cloudless day. In Computer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11 . S...

    work page 2010

Show all 29 references
  1. [8]

    Hongyu Fu, Xin Yu, Lincheng Li, and Li Zhang. 2023. CBARF: Cascaded Bundle- Adjusting Neural Radiance Fields from Imperfect Camera Poses. arXiv preprint arXiv:2310.09776 (2023)

    2023 arXiv

  2. [10]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  3. [11]

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. 2021. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision . 5741–5751

    2021

  4. [12]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM 65, 1 (2021), 99–106

    2021

  5. [13]

    Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. 2022. Instant Neural Graphics Primitives with a Multiresolution Hash Encoding. ACM Trans. Graph. 41, 4, Article 102 (July 2022), 15 pages. doi:10.1145/3528223.3530127

    2022

  6. [14]

    Johannes L Schonberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4104–4113

    2016

  7. [15]

    Liang Song, Guangming Wang, Jiuming Liu, Zhenyang Fu, Yanzi Miao, et al

  8. [16]

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. 2024. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20697–20709

    2024

  9. [17]

    arXiv preprint arXiv:2309.05028 (2023)

    SC-NeRF: Self-Correcting Neural Radiance Field with Sparse Views. arXiv preprint arXiv:2309.05028 (2023)

    2023 arXiv

  10. [18]

    Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. 2024. GaussianCity: Generative Gaussian splatting for unbounded 3D city generation. arXiv preprint arXiv:2406.06526 (2024)

    2024 arXiv

  11. [19]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

    2004

  12. [20]

    Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. 2021. inerf: Inverting neural radiance fields for pose estimation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 1323–1330

    2021

  13. [21]

    Bo Xu, Ziao Liu, Mengqi Guo, Jiancheng Li, and Gim Hee Lee. 2024. URS-NeRF: Unordered Rolling Shutter Bundle Adjustment for Neural Radiance Fields. In European Conference on Computer Vision . Springer, 458–475

    2024

  14. [22]

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 19447–19456. MM’33, October 2025, Dublin, Ireland Ziao Liu, Zh...

    2024

  15. [23]

    Xinyou Yin, JAN Goudriaan, Egbert A Lantinga, JAN Vos, and Huub J Spiertz

  16. [24]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  17. [25]

    Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. 2024. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. In European Conference on Computer Vision . Springer, 326–342

    2024

  18. [26]

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. 2024. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825 (2024)

    2024 arXiv

  19. [2003]

    Annals of botany 91, 3 (2003), 361–371

    A flexible sigmoid function of determinate growth. Annals of botany 91, 3 (2003), 361–371

    2003

  20. [2018]

    The unreasonable effectiveness of deep features as a perceptual metric. AttentionGS: Towards Initialization-Free 3D Gaussian Splatting via Structural Attention MM’33, October 2025, Dublin, Ireland Table 3: Quantitative results on the Mip-NeRF 360 dataset. For each metric, the ...

    2025

  21. [2023]

    ACM Trans

    3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42, 4 (2023), 139–1

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.