Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

PointGS: Point Attention-Aware Sparse View Synthesis with Gaussian Splatting

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that by initializing 3D Gaussians from a stereo foundation model's dense point cloud and enriching each point with variance-aggregated multi-view features refined by local self-attention, sparse-view novel view synthesis…

desk verdict Plausible sparse-view 3DGS pipeline with real-time gains, but the main comparison is confounded by VGGT initialization and the ablation text has a 0.51 vs 0.22 dB inconsistency. read the letter →

arxiv 2506.10335 v1 pith:BBGHOW4Q submitted 2025-06-12 cs.CV

Lintao Xiang , Hongpei Zheng , Yating Huang , Qijun Yang , Hujun Yin This is my paper
classification cs.CV
keywords 3DGaussiansplattingfew-shotnovelviewsynthesissparseviewspoint-wisefeaturesself-attentionstereofoundationmodeldepthregularizationreal-timerendering
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

PointGS is a sparse-view novel view synthesis system built on 3D Gaussian splatting. Instead of relying on structure-from-motion point clouds or spherical-harmonic colors, it initializes Gaussians from dense point clouds and poses produced by a stereo foundation model, assigns each point a feature vector aggregated from multi-scale features of all input views, and refines these features with a local self-attention interaction before decoding Gaussian attributes with two small MLPs. The paper reports that this yields real-time renderings that clearly surpass NeRF-based few-shot methods and match or beat recent few-shot 3DGS methods on the LLFF, Mip-NeRF360, and Deep Blending benchmarks. The intended payoff is that sparse inputs—as few as three views—can produce high-quality interactive rendering without a per-scene structure-from-motion pipeline.

What carries the argument

The load-bearing machinery is the point-wise feature representation combined with a local self-attention interaction. For each 3D point, multi-scale appearance features are sampled from all sparse views and aggregated by variance, $\frac{1}{N}\sum_{i}(f_j^i - \bar{f}_j)^2$, so the descriptor encodes disagreement across views rather than averaged appearance. A point interaction network updates the descriptor: for each point $i$, the $K$ nearest neighbors $\Omega(i)$ are found, a positional encoding $\delta = \theta(x_i - x_j)$ injects relative geometry, and the update $\hat{f}_i = \sum_{f_j \in \Omega(i)} \rho(\gamma(\varphi(f_i) - \psi(f_j) + \delta)) \odot (\alpha(f_j) + \delta)$ mixes neighbor features by learned attention weights. The enriched features are decoded into Gaussian attributes by two lightweight MLPs, and a depth smoothness loss regularizes the rendered depth. This replaces hand-tuned Gaussian geometry optimization with a feature-driven prediction loop.

What would settle it

Run the same 3-view LLFF and 24-view Mip-NeRF360 settings but replace the stereo-model poses with ground-truth poses; if PSNR does not change materially, the reported gains come from the feature and interaction modules rather than the pose initialization. Conversely, on a textureless or repetitive scene where the dense point cloud is known to drift, compare PointGS against a depth-regularized 3DGS baseline: a large quality drop would confirm that the method's accuracy is bounded by its geometric priors.

Watch

Extended reading notes

Core claim

The central discovery is that the color and covariance of each 3D Gaussian can be predicted from learned point-wise appearance features rather than from optimized per-Gaussian spherical-harmonic coefficients, and that this works well when the points come from a dense multi-view stereo prior and when neighboring points exchange information through self-attention. Each Gaussian's feature is built by projecting the point into every input view, sampling feature maps at three scales, and taking the variance across views as an aggregated descriptor; a point interaction network with K-nearest neighbours then updates these descriptors, and two MLPs decode color, opacity, and covariance. The paper reports that every component contributes: removing position encoding, point interaction, or variance fusion degrades PSNR, and increasing K from 0 to 9 improves quality. On LLFF with 3 views it reaches 20.54 dB PSNR against 20.22 dB for the strongest reproduced few-shot 3DGS baseline and well above NeRF-based methods, at 413 FPS; on Mip-NeRF360 with 24 views it reaches 23.48 dB against 22.45 dB for the reproduced baseline; on Deep Blending with 24 views it reaches 24.32 dB.

Load-bearing premise

The whole pipeline inherits the accuracy of the external stereo model's estimated camera poses and dense point clouds, and of the monocular depth prior used for regularization; if those priors are wrong, the point features and decoded Gaussian parameters carry the error, and the paper provides no sensitivity analysis for that dependence.

Editorial extensions

If this is right

  • With as few as three input views, renderings reach quality above NeRF-based few-shot methods while keeping interactive frame rates, making sparse-view capture practical for AR, VR, and robotics.
  • The method removes the need for structure-from-motion in sparse-view settings: camera poses and dense point clouds come from a stereo foundation model and are refined by bundle adjustment.
  • Variance-based multi-view fusion is a cheap way to inject cross-view information into each Gaussian, and the paper shows it outperforms simple averaging of features.
  • Local self-attention among Gaussian points is the main quality driver: metrics improve monotonically as the number of neighbors K grows from 0 to 9, so the neighbourhood interaction is what carries the gain.
  • Depth regularization, consisting of a rendered-depth versus monocular-depth-prior term plus a depth-smoothness term, adds a small but consistent improvement of roughly 0.1 dB per term on LLFF.

Reading between the lines

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

  • If the stereo foundation model's poses or dense geometry are inaccurate in textureless or heavily occluded regions, the point features inherit that error; an ablation that swaps in ground-truth poses would reveal how much of the reported gain comes from better initialization versus the learned features.
  • Because the feature aggregation and point-attention stages are agnostic to how the point cloud was produced, the same front-end could be coupled to other geometry priors, letting future work improve robustness without redesigning the decoder.
  • The reported FPS is already interactive, but the point-interaction overhead could be reduced with sparse or clustered attention, as the paper itself suggests, pushing frame rates higher while preserving quality.
  • The explicit point-wise features give a direct handle on which Gaussian to modify, suggesting the framework could extend to dynamic scenes or semantic editing without structural changes.
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 manuscript proposes PointGS, a 3D Gaussian splatting framework for sparse-view novel view synthesis. It initializes Gaussians from VGGT dense point clouds and estimated poses, extracts multi-scale FPN features for each point, aggregates them with a variance-like operation, applies a local self-attention point interaction network, and decodes Gaussian attributes with two lightweight MLPs. Training combines photometric, depth, and depth-smoothness losses. Experiments on LLFF (3 views), MipNeRF360 (24 views), and Deep Blending (24 views) report improvements over NeRF baselines and competitive or better numbers than existing few-shot 3DGS methods, with real-time rendering speeds.

Significance. If the causal claims are established, the paper would show that sparse-view novel view synthesis can be performed at real-time rates using foundation-model initialization plus point-wise learned features, which is practically valuable. The proposed pipeline is plausible, the writing is clear, and the experiments span three datasets. The ablations show that each internal module contributes positively, although the reported magnitude for the point interaction module is inconsistent. The main weakness is that the central attribution of the gains to the proposed point-attention features is confounded with the VGGT-based initialization, and the paper lacks controlled experiments to separate these factors. The absence of error bars and the mixing of published and reproduced baseline numbers further weaken the quantitative evidence. Overall, the contribution is promising but not yet established as stated.

major comments (4)
  1. [§2.4(1), §3.3, Tables 1–3] The central quantitative claim is confounded with initialization. PointGS initializes all Gaussians from VGGT dense point clouds and estimated poses, whereas the compared few-shot 3DGS baselines (FSGS, CoherentGS, DNGaussian, DepthGS) rely on COLMAP-based or their original initializations. Since 3DGS is known to be sensitive to the initial point cloud, the reported margins (e.g., 0.32 dB over reproduced FSGS on LLFF and 1.03 dB on MipNeRF360) may be due to VGGT's denser and more globally consistent initialization rather than to the proposed point-attention features. The ablation in Table 4 holds VGGT fixed and only toggles internal modules, so it cannot separate these factors. The authors should include a controlled comparison (e.g., running FSGS/CoherentGS with VGGT initialization, or PointGS with COLMAP initialization) or explicitly restrict the claim to the full pipeline. The Limitations section also does not discuss sensitivity to VGGT and ZoeDepth priors.
  2. [§3.4, Table 4] The ablation text contains an internal inconsistency: it states that removing the point interaction module decreases PSNR by 0.22 dB, but Table 4 shows a drop from 20.54 to 20.03 dB, i.e., 0.51 dB. The text understates the contribution of this module, and the discrepancy should be corrected. Additionally, the 'Color representation' ablation described in §3.4 is not included in Table 4, so the claimed 20.54-to-20.36 drop is not verifiable from the table.
  3. [§3.3, Tables 1–2] The baseline reporting mixes published and reproduced numbers without sufficient detail. Table 2 lists FSGS(paper) at 23.70 dB and FSGS(repro) at 22.45 dB, a 1.25 dB discrepancy, while Table 1 also shows smaller differences (20.31 vs 20.22). The paper says results are 'reported ... or referring to the results reported from the published papers' (§3.1), but does not specify the reproduction settings or explain the discrepancies. A fair comparison requires all baselines to be run under the same protocol, with the same view selection, image resolution, and initialization, or a clear justification for using published numbers.
  4. [§3.3] No statistical significance or variance information is provided. The LLFF experiments use a single random draw of 3 training views, and each table reports one run per method. In the sparse-view setting, results can vary substantially with view selection and random seed; the small margins over reproduced FSGS (0.32 dB on LLFF) may not be significant. The authors should report means and standard deviations over multiple seeds/view selections, or at least specify the exact view split and seed used.
minor comments (5)
  1. [Eq. (5)] The 'variance' in Eq. (5) is not normalized by N; it is the sum of squared deviations. Please either divide by N or describe it as a variance-like aggregation.
  2. [Eq. (10)] The notation ∇² in Eq. (10) is not defined; the Laplacian or Hessian convention should be stated explicitly.
  3. [§3.1] The sentence 'The best quantitative results are reported on these methods following the same experimental settings or referring to the results reported from the published papers' is ambiguous and should be replaced with a precise statement of which numbers were reproduced and which were taken from papers.
  4. [References] Some references are incomplete or duplicated (e.g., [1] and [3] both cite Mip-NeRF360, and [22] lists only two authors); please ensure all references are complete and correctly formatted.
  5. [§3.4] The 'Color representation' ablation is described in the text but missing from Table 4; please add it as a row for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PointGS is an empirical pipeline validated against external benchmarks; no prediction reduces to a fitted parameter or self-citation chain.

full rationale

The paper's central claim is an empirical rendering-quality result evaluated on LLFF, MipNeRF360, and Deep Blending against published NeRF and 3DGS baselines. The method is a composition of learned modules: VGGT initialization, FPN feature extraction, variance-based multi-view fusion (Eq. 5), local self-attention interaction (Eq. 7), and MLP decoding of Gaussian parameters. None of these steps defines a prediction in terms of the quantity it claims to predict; the rendered image is produced by differentiable splatting and compared to held-out ground-truth views. Losses in Section 2.5 are standard photometric, depth, and smoothness regularizers, not fitted parameters disguised as predictions. The ablation study in Table 4 removes components and shows their contributions, which is honest internal evaluation rather than circular reasoning. The absence of a control experiment that gives baselines the same VGGT initialization is a legitimate experimental-design concern about attributing the reported gains, but it is not a circularity of the kind where an equation or fitted parameter reduces to its own input. The paper contains no load-bearing self-citation: the authors do not cite their own prior work, and the external prior models (VGGT, ZoeDepth) are used as off-the-shelf components rather than as justification for the core claim. Therefore no circular step can be exhibited with a specific reduction, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several external pretrained components (VGGT, ZoeDepth, FPN backbone) whose outputs are trusted without sensitivity analysis. The hand-set hyperparameters K=3 and the loss weights are the main free parameters; the K ablation shows the chosen value is not optimal. No new entities are introduced.

free parameters (3)
  • K nearest neighbors = 3
    Number of neighbors in point interaction; chosen by hand. Ablation Table 6 shows PSNR rises with K (K=9 gives 20.71 vs 20.54 for K=3), so the main results use a suboptimal hand-set value.
  • Loss weights lambda_1..lambda_4 = 0.8, 0.2, 0.05, 0.067
    Borrowed from FSGS [40] per Section 2.5; not re-tuned. These weights control color, depth, and smoothness terms and are hand-selected priors.
  • Depth smoothness weights alpha_1, alpha_2 = 0.5
    Set to 0.5 in Eq. 10 with no ablation or justification.
assumptions (5)
  • domain assumption VGGT provides accurate camera poses and dense, globally consistent point clouds from sparse views
    Section 2.4(1) replaces SfM/COLMAP with VGGT for pose and dense initialization; the whole pipeline inherits VGGT's accuracy and the paper provides no sensitivity analysis.
  • domain assumption ZoeDepth produces reliable metric depth priors for depth regularization
    Section 2.5 defines L_depth against ZoeDepth's prediction; if the prior is biased, geometry optimization is biased.
  • domain assumption Multi-scale FPN features sampled at projected 3D points are informative for appearance
    Section 2.4(2) assumes projected 2D features represent the 3D point's appearance; no analysis of feature quality or projection errors is given.
  • domain assumption KNN spatial neighborhoods capture appearance coherence
    Section 2.4(3) assumes nearby points share appearance; ablation supports this but K is a free choice.
  • standard math Standard 3DGS differentiable rasterizer and adaptive densification behave as in [16]
    Method builds on the official 3DGS codebase; the splatting equations in Section 2.3 are taken as given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PointGS: Point Attention-Aware Sparse View Synthesis with Gaussian Splatting." pith.science (2026). https://pith.science/paper/BBGHOW4Q

@misc{pith2026250610335,
  author       = {Pith},
  title        = {Pith review of: PointGS: Point Attention-Aware Sparse View Synthesis with Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BBGHOW4Q}},
  note         = {Machine review of arXiv:2506.10335}
}
read the original abstract

3D Gaussian splatting (3DGS) is an innovative rendering technique that surpasses the neural radiance field (NeRF) in both rendering speed and visual quality by leveraging an explicit 3D scene representation. Existing 3DGS approaches require a large number of calibrated views to generate a consistent and complete scene representation. When input views are limited, 3DGS tends to overfit the training views, leading to noticeable degradation in rendering quality. To address this limitation, we propose a Point-wise Feature-Aware Gaussian Splatting framework that enables real-time, high-quality rendering from sparse training views. Specifically, we first employ the latest stereo foundation model to estimate accurate camera poses and reconstruct a dense point cloud for Gaussian initialization. We then encode the colour attributes of each 3D Gaussian by sampling and aggregating multiscale 2D appearance features from sparse inputs. To enhance point-wise appearance representation, we design a point interaction network based on a self-attention mechanism, allowing each Gaussian point to interact with its nearest neighbors. These enriched features are subsequently decoded into Gaussian parameters through two lightweight multi-layer perceptrons (MLPs) for final rendering. Extensive experiments on diverse benchmarks demonstrate that our method significantly outperforms NeRF-based approaches and achieves competitive performance under few-shot settings compared to the state-of-the-art 3DGS methods.

Figures

Figures reproduced from arXiv: 2506.10335 by the authors.

Figure 1
Figure 1. Overview of the proposed method. Given a few input images, we firstly generate dense point clouds from VGGT [32] and extract multi-scale feature maps via an FPN. Each point is then assigned point-wise appearance features by sampling from these feature maps. To enhance feature representation, a self-attention-based interaction module enables information exchange among neighboring points. Finally, two lightweight MLPs… view at source ↗
Figure 2
Figure 2. Structure of point interaction.The input is a set of point-wise feature vectors f with associated 3D coordinates x. The point attention operators can facilitates information exchange between these localized feature vectors and output the enhanced point features ˆf. To promote smoothness in the rendered depth map, we introduce an additional depth smoothness term, defined as: Lsmooth = e −α1|∇Igt| |∇Dre| + e −α2|∇2 Ig… view at source ↗
Figure 3
Figure 3. Visualized results on LLFF Datasets. Our method generates more realistic renderings with less blur in occluded and thin regions. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualized results on Mip-NeRF360 dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison on Deep Blending with 24 input views. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on the components of PointGS. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in 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. Sparse-View 3D Reconstruction: Recent Advances and Open Challenges

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A comprehensive survey that organizes sparse-view 3D reconstruction methods into geometry-based, NeRF, 3DGS, and diffusion-based categories, with benchmarks and open challenges.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Mip-nerf 360: Un- bounded anti-aliased neural radiance fields

    Jonathan T Barron and Ben Mildenhall. Mip-nerf 360: Un- bounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5

    work page 2022

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5855–5864,

  3. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 1

    work page 2022

  4. [4]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023. 1, 4

    arXiv 2023

  5. [5]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19457–19467, 2024. 2

    2024

  6. [6]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. InProceedings of the IEEE/CVF international conference on computer vision, pages 14124–14133, 2021. 2

    2021

  7. [7]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. InEuropean Conference on Computer Vision, pages 370–386. Springer, 2024. 2, 6

    work page 2024

  8. [8]

    Depth-regularized optimization for 3d gaussian splatting in few-shot images

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 811–820, 2024. 1, 4, 5, 6

    work page 2024

Show all 40 references
  1. [9]

    Depth-supervised nerf: Fewer views and faster train- ing for free

    Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ra- manan. Depth-supervised nerf: Fewer views and faster train- ing for free. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882– 12891, 2022. 2

    2022

  2. [10]

    Lightgaussian: Unbounded 3d gaussian compres- sion with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023

    Zhiwen Fan, Kevin Wang, Kairun Wen, and Zehao Zhu. Lightgaussian: Unbounded 3d gaussian compres- sion with 15x reduction and 200+ fps.arXiv preprint arXiv:2311.17245, 2023. 2

    2023 arXiv

  3. [11]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2

    2022

  4. [12]

    Semantic gaus- sians: Open-vocabulary scene understanding with 3d gaus- sian splatting.arXiv preprint arXiv:2403.15624, 2024

    Jun Guo, Xiaojian Ma, and Yue Fan. Semantic gaus- sians: Open-vocabulary scene understanding with 3d gaus- sian splatting.arXiv preprint arXiv:2403.15624, 2024. 2

    2024 arXiv

  5. [13]

    Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018

    Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering.ACM Transactions on Graphics (ToG), 37(6):1–15, 2018. 5

    2018

  6. [14]

    Hdr-nerf: High dynamic range neural radiance fields

    Xin Huang, Qi Zhang, and Ying Feng. Hdr-nerf: High dynamic range neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18398–18408, 2022. 2

    2022

  7. [15]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885–5894, 2021. 1, 2, 5, 6

    2021

  8. [16]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    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,

  9. [17]

    Compact 3d gaussian repre- sentation for radiance field

    Joo Chan Lee and Daniel Rho. Compact 3d gaussian repre- sentation for radiance field. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21719–21728, 2024. 2

    2024

  10. [18]

    Dngaussian: Optimiz- ing sparse-view 3d gaussian radiance fields with global-local depth normalization

    Jiahe Li, Jiawei Zhang, and Xiao Bai. Dngaussian: Optimiz- ing sparse-view 3d gaussian radiance fields with global-local depth normalization. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 20775–20785, 2024. 2, 4, 6, 7

    2024

  11. [19]

    Vastgaussian: Vast 3d gaussians for large scene reconstruction

    Jiaqi Lin, Zhihao Li, and Xiao Tang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5166–5175, 2024. 2

    2024

  12. [20]

    Compgs: Effi- cient 3d scene representation via compressed gaussian splat- ting.arXiv preprint arXiv:2404.09458, 2024

    Xiangrui Liu, Xinju Wu, and Pingping Zhang. Compgs: Effi- cient 3d scene representation via compressed gaussian splat- ting.arXiv preprint arXiv:2404.09458, 2024. 2

    2024 arXiv

  13. [21]

    Citygaussian: Real-time high- quality large-scale scene rendering with gaussians.arXiv preprint arXiv:2404.01133, 2024

    Yang Liu and He Guan. Citygaussian: Real-time high- quality large-scale scene rendering with gaussians.arXiv preprint arXiv:2404.01133, 2024. 2

    2024 arXiv

  14. [22]

    Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ACM Transactions on Graphics (ToG), 38(4):1– 14, 2019

    Ben Mildenhall and Pratul P Srinivasan. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines.ACM Transactions on Graphics (ToG), 38(4):1– 14, 2019. 5

    2019

  15. [23]

    Nerf: Representing scenes as neural radi- ance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, and Jonathan T Barron. Nerf: Representing scenes as neural radi- ance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021. 1, 2

    2021

  16. [24]

    Instant neural graphics primitives with a mul- tiresolution hash encoding.ACM transactions on graphics (TOG), 41(4):1–15, 2022

    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. 1, 2

    2022

  17. [25]

    Regnerf: Regularizing neural radiance fields for view syn- thesis from sparse inputs

    Michael Niemeyer, Jonathan T Barron, and Ben Mildenhall. Regnerf: Regularizing neural radiance fields for view syn- thesis from sparse inputs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480–5490, 2022. 2, 5, 6, 7

    2022

  18. [26]

    Coherentgs: Sparse novel view synthesis with coherent 3d gaussians

    Avinash Paliwal, Wei Ye, and Jinhui Xiong. Coherentgs: Sparse novel view synthesis with coherent 3d gaussians. In European Conference on Computer Vision, pages 19–37. Springer, 2024. 2, 6, 7

    2024

  19. [27]

    Self-positioning point-based 9 transformer for point cloud understanding

    Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based 9 transformer for point cloud understanding. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21814–21823, 2023. 4

    2023

  20. [28]

    Efficient 3d semantic segmentation with superpoint transformer

    Damien Robert and Hugo Raguet. Efficient 3d semantic segmentation with superpoint transformer. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17195–17204, 2023. 4

    2023

  21. [29]

    Dense depth pri- ors for neural radiance fields from sparse input views

    Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth pri- ors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12892–12901, 2022. 1

    2022

  22. [30]

    Structure- from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 4104–4113, 2016. 1

    2016

  23. [31]

    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,

  24. [32]

    Vggt: Visual geometry grounded transformer.arXiv preprint arXiv:2503.11651, 2025

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer.arXiv preprint arXiv:2503.11651, 2025. 1, 3

    2025 arXiv

  25. [33]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004

    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. 5

    2004

  26. [34]

    Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo

    Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 5610–5619, 2021. 1

    2021

  27. [35]

    Sparsegs: Real- time 360{\deg}sparse view synthesis using gaussian splat- ting.arXiv preprint arXiv:2312.00206, 2023

    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, 5

    2023 arXiv

  28. [36]

    Murf: Multi-baseline radiance fields

    Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041–20050, 2024. 2, 6

    2024

  29. [37]

    Freenerf: Im- proving few-shot neural rendering with free frequency reg- ularization

    Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Im- proving few-shot neural rendering with free frequency reg- ularization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254–8263,

  30. [38]

    Pose-free neural radiance fields via implicit pose reg- ularization

    Jiahui Zhang, Fangneng Zhan, Yingchen Yu, and Kunhao Liu. Pose-free neural radiance fields via implicit pose reg- ularization. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3534–3543, 2023. 2

    2023

  31. [39]

    The unreasonable effec- tiveness of deep features as a perceptual metric

    Richard Zhang and Phillip Isola. The unreasonable effec- tiveness of deep features as a perceptual metric. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018. 5

    2018

  32. [40]

    Fsgs: Real- time few-shot view synthesis using gaussian splatting.arXiv preprint arXiv:2312.00451, 2023

    Zehao Zhu, Zhiwen Fan, and Yifan Jiang. Fsgs: Real- time few-shot view synthesis using gaussian splatting.arXiv preprint arXiv:2312.00451, 2023. 1, 2, 5, 6, 7 10

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.