Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations

T0 review · 4 major / 5 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 PCR-GS claims that COLMAP-free 3D Gaussian Splatting can handle drastic camera motion by co-regularizing camera poses with DINO feature reprojection and wavelet frequency losses.

desk verdict A sensible, clearly written extension of CF-3DGS with consistent gains on harder splits, but the key ablation conflates the new initialization with the feature loss. read the letter →

arxiv 2507.13891 v2 pith:6WPG3TF3 submitted 2025-07-18 cs.CV

Yu Wei , Jiahui Zhang , Xiaoqin Zhang , Ling Shao , Shijian Lu This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingCOLMAP-freeposeestimationDINOfeaturesfeaturereprojectionwaveletfrequencyregularizationnovelviewsynthesiscameraco-regularization
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

PCR-GS claims that pose-free 3D Gaussian Splatting can handle drastic camera rotations and translations, not just smooth trajectories, by adding two co-regularization losses to the standard photometric training. The first loss reprojects DINO semantic features between adjacent frames and minimizes their discrepancy, which stabilizes relative pose estimation when RGB overlap is small. The second loss decomposes rendered and ground-truth images with a wavelet transform and anneals from low- to high-frequency components, which sharpens the rotation matrix estimate. On 4 fps Tanks&Temples and Free-Dataset splits, the method reports better PSNR/SSIM/LPIPS and lower pose errors than CF-3DGS, Nope-NeRF, BARF, and NeRFmm. If correct, this removes a practical barrier to using 3D-GS on handheld and moving-camera captures without running COLMAP.

What carries the argument

The load-bearing object is the pose co-regularization loss composed of two terms. Feature reprojection regularization samples DINO feature maps at original and reprojected pixel locations and minimizes $\|F_i\langle P_i\rangle - F_{i+1}\langle K P_i T_i \rangle\|_2$, aligning semantic content across wide-baseline adjacent views; a Best-Buddies correspondence search with $N_s=20$ points initializes the relative pose $T_i$. Wavelet-based frequency regularization applies a two-channel filter bank to split images into $LL$, $LH$, $HL$, and $HH$ bands and anneals the loss from low-frequency $LL$ to high-frequency components so that rotation errors, which shift edges and textures, are penalized in the spatial-frequency domain where they are visible.

What would settle it

Run PCR-GS on an image sequence with wide-baseline adjacent frames, such as more than 90 degrees of rotation or large occlusion, and compute relative pose error against COLMAP ground truth; if the DINO-based feature reprojection loss cannot reduce rotation error below the CF-3DGS baseline, the central claim fails. A cheaper test is to replace the DINO feature loss with random features; if rendering quality does not drop, DINO alignment is not the cause of the gain.

Watch

Extended reading notes

Core claim

The paper's central claim is that relative camera poses between adjacent frames can be co-regularized with the 3D-GS scene optimization so that pose-free 3D Gaussian Splatting converges under large inter-frame motion. The key mechanism is feature reprojection: DINO features extracted from frame $i$ are reprojected through the current relative pose into frame $i+1$ and compared with the DINO features there, using an L2 feature loss augmented by a photometric loss and a wavelet frequency loss; this is combined with a Best-Buddies-based initialization of the relative pose. The authors argue that RGB-only regularization is insensitive to small rotational shifts, whereas wavelet high-frequency components localize those shifts spatially, and that DINO features stay semantically aligned where pixel correspondences break down. The reported result is consistent improvement over four COLMAP-free baselines across eight Tanks&Temples scenes sampled at 4 fps and three Free-Dataset scenes, for both novel view synthesis and pose metrics.

Load-bearing premise

The method assumes DINO semantic features from adjacent frames stay aligned under the drastic rotations and translations it targets, and that the 20 Best-Buddies correspondences give a pose initialization good enough to avoid local minima.

Editorial extensions

If this is right

  • COLMAP-free 3D-GS can be applied to casually captured video with aggressive camera motion, where SfM preprocessing would fail or be prohibitively slow.
  • DINO-based semantic alignment provides a pose supervision signal that does not require ground-truth depth or pose, extending to other joint pose-scene optimization settings.
  • The wavelet annealing schedule suggests a general recipe: regularize coarse low-frequency structure first, then high-frequency details, to avoid noise during pose refinement.
  • The Best-Buddies initialization reduces sensitivity to local minima, making the final quality less dependent on the starting identity pose used by prior methods.

Reading between the lines

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

  • Editorial inference: because DINO features are view-invariant but not occlusion-aware, the method likely degrades on scenes with large occluded regions or repeated texture; a testable extension would be to add a visibility mask learned from the Gaussians.
  • Editorial inference: the wavelet loss operates on rendered versus ground-truth images, so it implicitly assumes photometric consistency; extending it to a perceptual or VGG feature space might transfer the rotation correction to scenes with lighting changes.
  • Editorial inference: the same co-regularization could be applied to monocular SLAM front-ends, replacing RGB-only bundle adjustment with DINO and wavelet residuals and potentially improving robustness to fast motions.
  • Editorial inference: since the method uses only 20 random correspondences for initialization, sensitivity to the random seed is unexplored; a variance study across seeds would show how much of the gain depends on the initialization draw.
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 / 5 minor

Summary. The paper introduces PCR-GS, a COLMAP-free 3D Gaussian Splatting method that jointly optimizes camera poses and a 3D Gaussian scene representation under challenging camera trajectories with large inter-frame motion. Two co-regularizations are proposed: (1) feature reprojection regularization that aligns DINO features extracted from adjacent frames, and (2) wavelet-based frequency regularization that penalizes discrepancies between rendered and ground-truth images in the wavelet domain, with an annealing schedule that gradually emphasizes high-frequency components. The method also uses sparse Best-Buddies correspondences to initialize relative camera poses. Experiments on newly constructed 4 fps Tanks&Temples splits and on the Free-Dataset report consistent improvements over CF-3DGS, Nope-NeRF, BARF, and NeRFmm in both rendering quality and pose accuracy.

Significance. If the reported results hold, PCR-GS is a useful contribution to pose-free 3D Gaussian Splatting: it tackles the practical and previously under-addressed regime of drastically varying camera trajectories, evaluates on 11 real-world scenes with both rendering and pose metrics, and compares against several strong baselines under a common protocol. The use of DINO features for cross-view pose regularization and the wavelet-domain rotation-focused loss are plausible and interesting ideas. However, the current evidence does not cleanly separate the contribution of the feature reprojection loss from the contribution of the new sparse-correspondence initialization, which is load-bearing for one of the two named regularizers. The manuscript would be substantially strengthened by an ablation that isolates these components and by pose-error ablations for each component.

major comments (4)
  1. [Sec. 3.2 and Table 5] The ablation labeled 'Base+FRR' in Table 5 conflates the feature reprojection regularization with the newly introduced Best-Buddies sparse-correspondence initialization. Section 3.2 presents both the feature reprojection loss in Eqs. (6)-(7) and the initialization procedure in Eqs. (8)-(9), and Section 3.4 states that 'before the pose co-regularization, we initialize the relative camera pose by establishing correspondences.' Since the full PCR-GS pipeline includes this initialization, the Base+FRR row very likely adds both the initialization and the FRR loss to the Base. The reported +4.82 dB PSNR gain therefore cannot be attributed to the proposed feature reprojection regularization. Please add ablation rows for initialization-only, FRR-loss-only (with identity initialization), and initialization-plus-FRR, and report all rendering metrics for each row.
  2. [Sec. 4.4, Tables 2-5] The paper claims pose co-regularization improves camera pose estimation, but the ablation study in Table 5 reports only novel-view-synthesis metrics and no pose-error metrics. Since pose accuracy is a central claim of the method (Tables 2 and 4 compare the full method against baselines), the component-wise ablation should also report relative pose rotation error, translation error, and ATE for Base, Base+init, Base+FRR, Base+WFR, and Base+FRR+WFR. Without this, it is not possible to tell whether the two regularizers improve pose estimation independently or whether the pose improvements come primarily from the initialization stage.
  3. [Sec. 3.2, Eq. (9)] The sparse-correspondence initialization is described as a pose optimization over only Ns = 20 randomly selected correspondences, but the manuscript does not analyze the sensitivity to this choice or to outlier correspondences. Given that the Best-Buddies algorithm is used on DINO features under large viewpoint changes, a quantitative evaluation of initialization quality (e.g., pose error before and after initialization, and the effect of varying Ns) would strengthen the claim that the initialization is robust and is not itself the main source of the gains attributed to FRR.
  4. [Sec. 4.4, Table 5] The wavelet-based frequency regularization ablation is not fully isolating the proposed mechanism either. The rows 'Base+WRF(w/o high-freq)' and 'Base+WRF(w/ high-freq)' differ in the use of high-frequency components, but the text does not make clear whether 'w/o high-freq' still uses the full wavelet loss without high-frequency terms or uses no frequency loss at all. Please clarify what exactly each row includes, and also report the marginal contribution of WFR when added to Base+init (rather than only to Base and to Base+FRR), so that the three components can be assessed separately.
minor comments (5)
  1. [Sec. 3.1] There is a typo in 'DINO features are much more stable under drastically view changess'; it should read 'drastic view changes.'
  2. [Sec. 3.3, Eq. (11)] Equation (11) has awkward line-breaking in the norm expression; please format it cleanly so the reader can see that the discrepancy is a weighted combination of component-wise norms.
  3. [Tables 2 and 4] The pose metrics RPEr and RPEt are not explicitly defined with units. Please state whether rotation error is in degrees or radians and whether translation error is in meters or normalized units, and define RPEr, RPEt, and ATE in the evaluation-metrics paragraph.
  4. [Sec. 4.1] The paper states that details of the 4 fps Tanks&Temples construction are in the appendix, but the appendix is not included in the provided text. Since the dataset split differs from CF-3DGS and directly affects all reported numbers, the supplementary material must contain the exact frame indices and evaluation split.
  5. [Sec. 4.3] The baseline names are inconsistent in places ('Barf' versus 'BARF'); please use a single consistent capitalization.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the pose-regularization claims are empirical comparisons validated against external baselines, and neither named mechanism reduces by construction to its own inputs.

full rationale

PCR-GS's central claim — that DINO feature reprojection regularization (Eqs. 6-7) and wavelet-based frequency regularization (Eqs. 10-13) improve pose-free 3DGS — is an empirical claim validated against external baselines (CF-3DGS, Nope-NeRF, BARF, NeRFmm) on re-sampled real-world data, not a derivation that reduces to its own inputs. The DINO, Best-Buddies, and DPT priors all come from external, independently trained models ([1], [3], [4], [26]); none of the reported metrics is used to fit a parameter that is then renamed as a prediction. Loss weights (lambda0, lambda1, lambda2 = 0.6, 0.2, 0.2) and annealing thresholds (n0 = 100, n1 = 200) are fixed hyperparameters, and the wavelet frequency objective is a training loss distinct from the held-out evaluation metrics (PSNR, SSIM, LPIPS, RPE, ATE), which are computed against ground-truth images and poses after standard Procrustes alignment. The only self-citation is [40] (VMRF, co-authored by two PCR-GS authors), used descriptively in Related Work; it is not load-bearing, and no uniqueness theorem or ansatz is imported from the authors' prior work. One genuine weakness, flagged per review instructions, is that the Table 5 ablation conflates two changes introduced together in Section 3.2 — the FRR loss (Eqs. 6-7) and the Best-Buddies correspondence initialization (Eqs. 8-9) — in the Base+FRR row, and no pose-error ablation is reported, so the +4.82 dB attributed to FRR may reflect the initialization. This is an experimental attribution gap, not circularity: there is no equation identity and no fitted input reproduced as output, so the paper's derivation chain does not close upon itself.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method rests on four domain assumptions inherited or asserted without direct validation: DINO robustness across drastic viewpoint changes, DPT monocular depth as valid initialization, Best Buddies correspondence quality, and wavelet high-frequency discrepancy as a rotation-error signal. Hyperparameters controlling the loss balance are hand-set and one set of weights is unspecified.

free parameters (5)
  • loss weights (lambda0, lambda1, lambda2) = 0.6, 0.2, 0.2
    Set by hand in Section 3.4 with no sensitivity analysis; they determine the balance between RGB, feature, and wavelet losses.
  • wavelet annealing steps (n0, n1) = n0=100, n1=200
    Chosen in Section 3.3 to schedule high-frequency weight; no justification or sensitivity study is provided.
  • number of sparse correspondences Ns = 20
    Randomly selected in Section 3.2 for pose initialization; no ablation on Ns is reported.
  • wavelet component weights w_x = unspecified
    Eq. (11) defines a weighted sum over LL/LH/HL/HH components but the weights are never given, leaving the loss under-specified.
  • DINO feature layer = 9th layer
    Section 4.2 states layer 9 retains good detail; no comparison across layers is reported.
assumptions (4)
  • domain assumption DINO features are stable under drastic viewpoint changes and provide reliable dense correspondences between adjacent frames.
    Used in Section 3.2 to justify feature reprojection and initialization; cited to [1] but not validated on the 4 fps Tanks&Temples splits.
  • domain assumption Monocular depth from DPT yields point clouds accurate enough to initialize per-frame 3D Gaussians.
    Inherited from CF-3DGS [6] in Sections 3.1 and 4.2; a wrong monocular depth scale or distortion would bias all relative pose optimization.
  • domain assumption Discrepancies in wavelet high-frequency bands are a reliable signal for rotation error in camera poses.
    Core motivation of Section 3.3; the paper gives qualitative intuition but no formal or empirical proof that high-frequency discrepancy is monotone in rotation error.
  • domain assumption Best Buddies matching on DINO features produces enough correct correspondences to initialize relative poses.
    Section 3.2 relies on this for the initialization stage; noisy correspondences are only mitigated by random selection of 20, not by robust fitting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations." pith.science (2026). https://pith.science/paper/6WPG3TF3

@misc{pith2026250713891,
  author       = {Pith},
  title        = {Pith review of: PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6WPG3TF3}},
  note         = {Machine review of arXiv:2507.13891}
}
read the original abstract

COLMAP-free 3D Gaussian Splatting (3D-GS) has recently attracted increasing attention due to its remarkable performance in reconstructing high-quality 3D scenes from unposed images or videos. However, it often struggles to handle scenes with complex camera trajectories as featured by drastic rotation and translation across adjacent camera views, leading to degraded estimation of camera poses and further local minima in joint optimization of camera poses and 3D-GS. We propose PCR-GS, an innovative COLMAP-free 3DGS technique that achieves superior 3D scene modeling and camera pose estimation via camera pose co-regularization. PCR-GS achieves regularization from two perspectives. The first is feature reprojection regularization which extracts view-robust DINO features from adjacent camera views and aligns their semantic information for camera pose regularization. The second is wavelet-based frequency regularization which exploits discrepancy in high-frequency details to further optimize the rotation matrix in camera poses. Extensive experiments over multiple real-world scenes show that the proposed PCR-GS achieves superior pose-free 3D-GS scene modeling under dramatic changes of camera trajectories.

Figures

Figures reproduced from arXiv: 2507.13891 by the authors.

Figure 1
Figure 1. The proposed PCR-GS can model scenes with complex camera trajectories without using any camera pose priors. It achieves [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of the proposed PCR-GS. PCR-GS performs feature reprojection regularization and wavelet-based frequency regularization concurrently to optimize relative camera poses. For the feature reprojection regularization, we adopt DINO to extract semantic feature maps Fi and Fi+1 of adjacent frames Ii and Ii+1, and optimize the relative camera pose by minimizing the discrepancies between the reprojected feature … view at source ↗
Figure 3
Figure 3. We therefore align the robust DINO-based semantic features across adjacent views to optimize the camera pose with drastic rotations and translations. Specifically, with DINO features extracted from every frame, we design a feature reprojection regularization technique that optimizes relative camera pose as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparisons of PCR-GS with CF-3DGS [6] and Nope-NeRF [2] in novel view synthesis (on Tanks&Temples). PCR-GS achieves superior image rendering as compared with the two state-of-the-art methods. Pose Estimation. We apply the Procrustes analysis as in [2] to p…
Figure 5
Figure 5. Figure 5: Qualitative comparison on novel view synthesis over the Free-dataset [34]. The proposed PCR-GS generates better details with less artifacts consistently. Scenes CF-3DGS PCR-GS (Ours) PSNR↑ SSIM↑ LPIPS↓ PSNR↑ SSIM↑ LPIPS↓ Free-Dataset Pillar 14.47 0.41 0.61 17.15 0.49 0…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MoGaFace: Momentum-Guided and Texture-Aware Gaussian Avatars for Consistent Facial Geometry

    cs.CV 2025-08 unverdicted novelty 6.0 of 10

    MoGaFace improves 3D head avatar rendering by combining momentum-based expression correction with latent texture attention in Gaussian splatting, boosting novel-view quality under imperfect mesh tracking.

Reference graph

Works this paper leans on

48 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Deep vit features as dense visual descriptors

    Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2(3):4, 2021. 2, 4, 6

    arXiv 2021

  2. [2]

    Nope-nerf: Optimising neu- ral radiance field with no pose prior

    Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope-nerf: Optimising neu- ral radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023. 3, 6, 7

    work page 2023

  3. [3]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

    2021

  4. [4]

    Best-buddies similarity for robust template matching

    Tali Dekel, Shaul Oron, Michael Rubinstein, Shai Avidan, and William T Freeman. Best-buddies similarity for robust template matching. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 2021– 2029, 2015. 5

    work page 2021

  5. [5]

    Compact 3d gaussian splatting for dense visual slam

    Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2

    arXiv 2024

  6. [6]

    Colmap-free 3d gaussian splat- ting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 20796–20805,

    work page

  7. [7]

    Fastnerf: High-fidelity neu- ral rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neu- ral rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14346– 14355, 2021. 2

    work page 2021

  8. [8]

    Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction

    Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 2

    arXiv 2024

Show all 48 references
  1. [9]

    Sherf: Generalizable human nerf from a single image

    Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 9352–9364,

  2. [10]

    Efficientnerf efficient neural radiance fields

    Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf efficient neural radiance fields. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12902–12911, 2022. 2

    2022

  3. [11]

    S3 gaussian: Self-supervised street gaussians for autonomous driving

    Nan Huang, Xiaobao Wei, Wenzhao Zheng, Pengju An, Ming Lu, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. S3 gaussian: Self-supervised street gaussians for autonomous driving. arXiv preprint arXiv:2405.20323, 2024. 2

    2024 arXiv

  4. [12]

    Putting nerf on a diet: Semantically consistent few-shot view synthesis

    Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 2

    2021

  5. [13]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  6. [14]

    Tanks and temples: Benchmarking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG) , 36 (4):1–13, 2017. 5, 6, 7

    2017

  7. [15]

    Gecon- erf: Few-shot neural radiance fields via geometric consis- tency

    Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Gecon- erf: Few-shot neural radiance fields via geometric consis- tency. arXiv preprint arXiv:2301.10941, 2023. 2

    2023 arXiv

  8. [16]

    Dense depth-guided generalizable nerf

    Dongwoo Lee and Kyoung Mu Lee. Dense depth-guided generalizable nerf. IEEE Signal Processing Letters, 30:75– 79, 2023. 2

    2023

  9. [17]

    Sgs-slam: Se- mantic gaussian splatting for neural dense slam

    Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Se- mantic gaussian splatting for neural dense slam. InEuropean Conference on Computer Vision , pages 163–179. Springer,

  10. [18]

    Barf: Bundle-adjusting neural radiance fields

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Si- mon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5741–5751, 2021. 3, 6

    2021

  11. [19]

    Dynvideo-e: Harnessing dynamic nerf for large-scale motion-and view-change human-centric video editing

    Jia-Wei Liu, Yan-Pei Cao, Jay Zhangjie Wu, Weijia Mao, Yuchao Gu, Rui Zhao, Jussi Keppo, Ying Shan, and Mike Zheng Shou. Dynvideo-e: Harnessing dynamic nerf for large-scale motion-and view-change human-centric video editing. In Proceedings of the IEEE/CVF Conference on Compute...

    2024

  12. [20]

    Deblur-nsff: Neural scene flow fields for blurry dynamic scenes

    Achleshwar Luthra, Shiva Souhith Gantha, Xiyun Song, Heather Yu, Zongfang Lin, and Liang Peng. Deblur-nsff: Neural scene flow fields for blurry dynamic scenes. In Pro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 3658–3667, 2024. 2

    2024

  13. [21]

    Gnerf: Gan-based neural radiance field without posed camera

    Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 6351–6361, 2021. 3

    2021

  14. [22]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 2

    2021

  15. [23]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1–15, 2022. 2

    2022

  16. [24]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2

    2021

  17. [25]

    Ct-nerf: Incremental optimiz- ing neural radiance field and poses with complex trajectory

    Yunlong Ran, Yanxu Li, Qi Ye, Yuchi Huo, Zechun Bai, Ji- ahao Sun, and Jiming Chen. Ct-nerf: Incremental optimiz- ing neural radiance field and poses with complex trajectory. arXiv preprint arXiv:2404.13896, 2024. 6

    2024 arXiv

  18. [26]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. In Proceedings of 9 the IEEE/CVF international conference on computer vision, pages 12179–12188, 2021. 4, 6

    2021

  19. [27]

    Free view synthesis

    Gernot Riegler and Vladlen Koltun. Free view synthesis. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIX 16, pages 623–640. Springer, 2020. 2

    2020

  20. [28]

    Stable view synthesis

    Gernot Riegler and Vladlen Koltun. Stable view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12216–12225, 2021. 2

    2021

  21. [29]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 4104–4113, 2016. 2

    2016

  22. [30]

    A benchmark for the evalua- tion of rgb-d slam systems

    J ¨urgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evalua- tion of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573–580. IEEE, 2012. 6

    2012

  23. [31]

    Block-nerf: Scalable large scene neural view synthesis

    Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Prad- han, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Reco...

    2022

  24. [32]

    Single-view view synthe- sis with multiplane images

    Richard Tucker and Noah Snavely. Single-view view synthe- sis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 551–560, 2020. 2

    2020

  25. [33]

    Sparsenerf: Distilling depth ranking for few-shot novel view synthesis

    Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zi- wei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 9065–9076,

  26. [34]

    F2- nerf: Fast neural radiance field training with free camera trajectories

    Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2- nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , page...

    2023

  27. [35]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Si- moncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. 6

    2004

  28. [36]

    Nerf–: Neural radiance fields without known camera parameters

    Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 3, 6

    2021 arXiv

  29. [37]

    Sparsegs: Real- time 360 {\deg} sparse view synthesis using gaussian splat- ting

    Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real- time 360 {\deg} sparse view synthesis using gaussian splat- ting. arXiv preprint arXiv:2312.00206, 2023. 1, 2

    2023 arXiv

  30. [38]

    Point- nerf: Point-based neural radiance fields

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point- nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438–5448, 2022. 2

    2022

  31. [39]

    Featuren- erf: Learning generalizable nerfs by distilling foundation models

    Jianglong Ye, Naiyan Wang, and Xiaolong Wang. Featuren- erf: Learning generalizable nerfs by distilling foundation models. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 8962–8973, 2023. 2

    2023

  32. [40]

    Vmrf: View matching neural radiance fields

    Jiahui Zhang, Fangneng Zhan, Rongliang Wu, Yingchen Yu, Wenqing Zhang, Bai Song, Xiaoqin Zhang, and Shijian Lu. Vmrf: View matching neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia , pages 6579–6587, 2022. 3

    2022

  33. [41]

    Cor-gs: sparse-view 3d gaussian splatting via co-regularization

    Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: sparse-view 3d gaussian splatting via co-regularization. In European Conference on Computer Vision, pages 335–352. Springer, 2025. 2

    2025

  34. [42]

    Differentiable point-based radiance fields for efficient view synthesis

    Qiang Zhang, Seung-Hwan Baek, Szymon Rusinkiewicz, and Felix Heide. Differentiable point-based radiance fields for efficient view synthesis. In SIGGRAPH Asia 2022 Con- ference Papers, pages 1–12, 2022. 2

    2022

  35. [43]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 586–595, 2018. 6

    2018

  36. [44]

    Switch-nerf: Learning scene de- composition with mixture of experts for large-scale neural radiance fields

    MI Zhenxing and Dan Xu. Switch-nerf: Learning scene de- composition with mixture of experts for large-scale neural radiance fields. In The Eleventh International Conference on Learning Representations, 2022. 2

    2022

  37. [45]

    Stereo magnification: Learning view synthesis using multiplane images

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2

    2018 arXiv

  38. [46]

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes

    Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic au- tonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages...

    2024

  39. [47]

    Semgauss-slam: Dense semantic gaussian splatting slam

    Siting Zhu, Renjie Qin, Guangming Wang, Jiuming Liu, and Hesheng Wang. Semgauss-slam: Dense semantic gaussian splatting slam. arXiv preprint arXiv:2403.07494, 2024. 2

    2024 arXiv

  40. [48]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision , pages 145–163. Springer, 2025. 2 10

    2025

Pith tools

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