Pith. sign in

REVIEW 3 major objections 6 minor 40 references

C3-GS enhances generalizable Gaussian splatting with three lightweight feature modules—coordinate-guided attention, cross-dimensional attention, and cross-scale fusion—achieving state-of-the-art rendering and generalization from sparse pose

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

C3-GS improves generalizable Gaussian Splatting by adding coordinate-aware attention, cross-dimensional attention, and cross-scale opacity fusion to the MVSGaussian baseline.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Incremental but solid improvement on MVSGaussian for generalizable Gaussian splatting; the modules work in ablations, but the SOTA claim rests on small margins without error bars. the 3 major comments →

arxiv 2508.20754 v1 pith:RTBUR4IV submitted 2025-08-28 cs.CV cs.AI

${C}^{3}$-GS: Learning Context-aware, Cross-dimension, Cross-scale Feature for Generalizable Gaussian Splatting

Yuxi Hu , Jun Zhang , Kuangyi Chen , Zhe Zhang , Friedrich Fraundorfer This is my paper
classification cs.CV cs.AI
keywords generalizable Gaussian splattingnovel view synthesissparse viewsmulti-view stereofeature aggregationcross-scale fusionattention mechanismrendering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper claims that the bottleneck in feed-forward Gaussian splatting is weak feature encoding across spatial, dimensional, and scale axes. It proposes C3-GS, three lightweight additions to the MVSGaussian pipeline: coordinate-guided attention for 2D features, cross-dimensional attention to fuse 2D appearance with 3D cost-volume geometry, and cross-scale fusion to modulate opacity between coarse and fine stages. With only image-level supervision, the framework improves rendering quality and cross-dataset generalization over state-of-the-art generalizable Gaussian and NeRF methods on DTU, Real Forward-facing, NeRF Synthetic, and Tanks and Temples, and also produces sharper intermediate depth maps. A sympathetic reader cares because it suggests that parameter-free feature-aggregation design, not extra supervision or heavier networks, can close the gap between generalizable and per-scene-optimized Gaussian splatting.

Core claim

The paper establishes that enhancing source-view features in three complementary ways—coordinate-aware long-range attention, cross-dimensional 2D-3D attention, and cross-scale opacity modulation—makes feed-forward Gaussian parameter prediction more geometry-consistent and visually accurate. The three modules plug into the MVSGaussian coarse-to-fine cost-volume framework, adding negligible overhead, and produce consistent gains on DTU (3-view PSNR 27.03 → 27.87) and on datasets unseen during training, plus better depth maps. The central claim is that explicit multi-scale, multi-dimensional feature constraints substitute for depth or mask supervision.

What carries the argument

CGA (Coordinate-Guided Attention) pools features along height and width, merges them through a 1D convolution and sigmoid, and applies the resulting attention maps multiplicatively to modulate feature responses, capturing long-range spatial dependencies while keeping coordinate cues. CDA (Cross-Dimensional Attention) concatenates voxel features sampled from the 3D cost volume with aggregated 2D image features, then uses this fused representation as a query in cross-view attention over the source-view 2D features, jointly integrating geometric consistency and appearance detail. CSF (Cross-Scale Fusion) concatenates Gaussian features from adjacent coarse and fine stages, predicts a modulation

Load-bearing premise

The cost volume built via differentiable homography from a few posed views must produce trustworthy Gaussian centers; when that geometry estimate fails under wide baselines or extrapolated viewpoints, the added feature modules cannot compensate.

What would settle it

Run C3-GS and its MVSGaussian baseline on a wide-baseline or view-extrapolation benchmark where target views lie outside the input camera arc; if PSNR degrades to baseline levels, the stated limitation dominates. Also perform a one-module-at-a-time ablation on a held-out dataset to verify the reported gains are not specific to DTU.

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

If this is right

  • If the claim holds, generalizable Gaussian splatting can approach per-scene-optimization quality without per-scene training, using only posed images as supervision.
  • The intermediate depth maps improve as a byproduct, strengthening multi-view reconstruction tasks that consume predicted depth.
  • The modules are lightweight and architecture-agnostic enough to be added to other cost-volume-based generalizable splatting pipelines.
  • Cross-scale opacity modulation suggests that scale relationships among Gaussians are a cheap lever for visual quality; refining only opacity is sufficient in practice.
  • The framework works with varying view counts (2, 3, 4) trained consistently, easing deployment where scene capture density is not fixed.

Where Pith is reading between the lines

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

  • One could test whether CDA's query design generalizes to transformer-based MVS features (e.g., MVSFormer) or to monocular depth priors, potentially reducing dependence on the cost volume for Gaussian centers.
  • The ablation pattern—CGA alone helps little while CDA and CSF help more—suggests cross-view consistency and scale modeling matter more than 2D receptive field; ablating under wide-baseline or view-extrapolation conditions could reveal where each module fails first.
  • The opacity-only cross-scale refinement could be stacked or applied iteratively for further multi-scale refinement, a natural extension beyond the paper's single-step design.
  • Because the method is built on MVSGaussian, the gains may be partly inherited from that baseline's cost volume; a ceteris-paribus test on another base (e.g., MVSplat) would show whether the modules transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes C3-GS, a generalizable 3D Gaussian Splatting framework built on top of MVSGaussian. It introduces three lightweight modules: Coordinate-Guided Attention (CGA) for coordinate-aware 2D feature enhancement, Cross-Dimensional Attention (CDA) to fuse 2D appearance features with 3D cost-volume features, and Cross-Scale Fusion (CSF) to modulate Gaussian opacity across coarse-to-fine stages. The method is trained on DTU with only image supervision and evaluated for generalization on Real Forward-facing, NeRF Synthetic, and Tanks and Temples. The paper reports improved PSNR/SSIM/LPIPS and depth accuracy over several baselines, with ablations attributing gains to the proposed modules. Code is released.

Significance. If the reported gains are robust, C3-GS provides a simple and effective extension to MVSGaussian, improving feature learning for generalizable Gaussian Splatting without additional supervision. The three modules are conceptually clear, the experiments cover four standard benchmarks, and the paper includes ablations and resource analysis. The code release is a strength. However, the central 'state-of-the-art' claim is tempered by small and somewhat inconsistent margins, single-run results without uncertainty estimates, and the absence of a capacity-matched baseline. The acknowledged limitations (wide-baseline scenarios, view extrapolation) appropriately bound the scope of the claim.

major comments (3)
  1. [Section 4, Tables 1, 2, 5] The paper reports single-run metrics without error bars or repeated runs. Several decisive margins are small: Real Forward-facing 3-view PSNR improves by only +0.10 dB (23.97 vs 23.87), and the CGA-only ablation in Table 5 gains +0.04 dB. On Real Forward-facing 2-view, the method is worse than the baseline in PSNR (22.32 vs 22.58) and LPIPS (0.201 vs 0.192). Since the MVSGaussian baseline was retrained once (Sec. 4.1), it is unclear whether these differences are beyond training noise. The 'state-of-the-art' claim at 3 and 4 views requires multi-seed or per-scene bootstrap uncertainty estimates.
  2. [Section 4.4 / Table 5] No capacity-matched baseline is included. The proposed modules add parameters and computational cost; without a control that increases capacity equally but without CGA/CDA/CSF, the observed gains cannot be attributed specifically to the proposed mechanisms rather than to added model capacity. A parameter count / FLOP comparison or a capacity-matched ablation is needed to support the claim that the modules themselves are the cause of the improvement.
  3. [Abstract and Section 4.2] The claim of 'state-of-the-art rendering quality and generalization ability' is too broad. While the method is top-ranked for 3 and 4 views on most benchmarks, Table 2 shows a regression on Real Forward-facing 2-view PSNR (22.32 vs 22.58), and Table 1 shows 2-view PSNR below MatchNeRF. The text only explains the latter via MatchNeRF's optical-flow prior. The headline claim should be qualified to the settings where the evidence holds.
minor comments (6)
  1. [Sec. 3.2, Eq. (12)] The decision to update only opacity in CSF is stated as empirically motivated, but no ablation is shown comparing opacity-only against updating all Gaussian attributes. Please add this ablation or soften the claim.
  2. [Sec. 4.1] The text states that MVPGS is 're-trained a fixed number of times' but gives no details; the MVSGaussian baseline training is similarly described without specifying the number of runs or seeds. Please state the exact procedure.
  3. [Table 4] IBRNet's Abs err value of 324 appears anomalously large relative to other methods. Please verify the number or add a note explaining this outlier.
  4. [Fig. 2 caption] The acronyms CGA, CDA, CSF are used in the caption without being spelled out. Define them in the caption or refer to Section 3.
  5. [Supplementary Sec. 6.2] The statement 'we achieve the best PSNR performance in Table 1 under the 2-views setting' is only true when training with 2 views (Table 6). This nuance should be clarified in the main text to avoid confusion.
  6. [Conclusion] The limitations are clearly stated and appreciated. Consider moving them to a dedicated 'Limitations' subsection for greater visibility.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical, externally evaluated improvement; modules are specified by explicit equations and tested on held-out benchmarks.

full rationale

The paper's derivation chain is a trained neural architecture, not an analytic derivation from fitted constants. The three proposed modules (CGA, CDA, CSF) are defined by explicit equations (Eqs. 1-12) and trained end-to-end using the image-based loss in Eq. 13. There is no step in which an output quantity is defined in terms of the target metric, no fitted parameter is later relabeled as a prediction, and no 'uniqueness theorem' is imported from the authors' prior work to force a particular design. The self-citations [14,39] are used only as inspiration for the CGA module; the module itself is specified and ablated in the present paper, and the state-of-the-art claim rests on held-out benchmarks (DTU for training; Real Forward-facing, NeRF Synthetic, Tanks and Temples for direct generalization) rather than on those citations. The baseline MVSGaussian [23] is retrained from scratch for fair comparison, and the conclusion explicitly acknowledges failure modes in wide-baseline and view-extrapolation settings. The absence of multi-seed error bars is a statistical robustness concern, not evidence of circularity. No circular step can be exhibited from the paper's own equations or citations.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 3 invented entities

The paper introduces no new physical entities, only learned architectural components. The free parameters are standard hyperparameters (loss weights, depth samples, number of views). The main assumptions are inherited from MVSGaussian and typical MVS pipelines: the cost volume from sparse posed views is a reliable geometry prior, and the rasterizer gradients are correct. The only paper-specific assumption is the choice to update opacity alone in CSF, which is justified empirically.

free parameters (3)
  • loss weights (βs, βp, γ(1), γ(2)) = βs=0.1, βp=0.05, γ(1)=0.5, γ(2)=1
    These constants are chosen by hand in Sec 4.1 and affect the magnitude of each loss term in the overall objective (Eq. 13). They are not tuned per dataset and follow prior practice.
  • number of depth hypotheses = 64 (coarse), 8 (fine)
    Set in Sec 4.1. This controls the granularity of the cost volume and depth regression, and is a design choice not independently motivated.
  • number of source views during training = 3
    The model is trained with 3 source views (Sec 4.1) and tested with 2, 3, 4. This choice affects the cost volume and feature fusion pipeline.
axioms (4)
  • standard math The differentiable Gaussian splatting rasterizer provides a correct gradient model for end-to-end training.
    The paper relies on the standard 3D-GS rasterization pipeline (Sec 3.2) without re-deriving it.
  • domain assumption Source images come with accurate camera poses and intrinsic parameters.
    Homography warping for cost volume construction (Sec 3.2) requires known camera geometry.
  • domain assumption A cost volume built from a sparse set of views is a sufficient geometry prior for accurate Gaussian centers.
    The method inherits this from MVSGaussian (Sec 3.2, 'Gaussian Center Prediction from MVS') and does not validate it beyond the tested benchmarks.
  • ad hoc to paper The Cross-Scale Fusion module should update only opacity and not other Gaussian attributes.
    Sec 3.2 states 'we empirically find that updating only the opacity achieves a good balance', which is an empirical assumption without theoretical justification.
invented entities (3)
  • CGA (Coordinate-Guided Attention) independent evidence
    purpose: Adds coordinate-aware long-range feature dependencies to the 2D feature pyramid.
    Validated on held-out benchmarks (Table 5), but mathematically identical to existing coordinate attention, so it is not a fundamentally new entity.
  • CDA (Cross-Dimensional Attention) independent evidence
    purpose: Fuses 2D image features with 3D cost volume features to build spatially aware descriptors.
    Validated on held-out benchmarks through ablations and final performance.
  • CSF (Cross-Scale Fusion) independent evidence
    purpose: Modulates Gaussian opacity across scales to improve detail preservation.
    Validated on held-out benchmarks; a lightweight MLP weight mechanism.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ${C}^{3}$-GS: Learning Context-aware, Cross-dimension, Cross-scale Feature for Generalizable Gaussian Splatting." pith.science (2026). https://pith.science/paper/RTBUR4IV

@misc{pith2026250820754,
  author       = {Pith},
  title        = {Pith review of: $C^3$-GS: Learning Context-aware, Cross-dimension, Cross-scale Feature for Generalizable Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTBUR4IV}},
  note         = {Machine review of arXiv:2508.20754}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Generalizable Gaussian Splatting aims to synthesize novel views for unseen scenes without per-scene optimization. In particular, recent advancements utilize feed-forward networks to predict per-pixel Gaussian parameters, enabling high-quality synthesis from sparse input views. However, existing approaches fall short in encoding discriminative, multi-view consistent features for Gaussian predictions, which struggle to construct accurate geometry with sparse views. To address this, we propose $\mathbf{C}^{3}$-GS, a framework that enhances feature learning by incorporating context-aware, cross-dimension, and cross-scale constraints. Our architecture integrates three lightweight modules into a unified rendering pipeline, improving feature fusion and enabling photorealistic synthesis without requiring additional supervision. Extensive experiments on benchmark datasets validate that $\mathbf{C}^{3}$-GS achieves state-of-the-art rendering quality and generalization ability. Code is available at: https://github.com/YuhsiHu/C3-GS.

Figures

Figures reproduced from arXiv: 2508.20754 by Friedrich Fraundorfer, Jun Zhang, Kuangyi Chen, Yuxi Hu, Zhe Zhang.

Figure 1
Figure 1. Figure 1: Comparison with existing methods. Left: Generalization results on DTU [15] with 3 input views, where our method achieves higher PSNR and SSIM. Right: Synthe￾sized images on “Horse" and “Family" scenes from Tanks and Temples [18], highlighting improved visual fidelity. Recent developments include generalizable methods [4, 7, 23, 29, 33, 40], which are inspired by generalizable NeRF [5, 8, 36], and aim to pr… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture. The proposed C 3 -GS is a coarse-to-fine framework that es￾timates depths and Gaussian representations from low resolution (stage ℓ) to high resolution (stage ℓ+1). It extracts features {Fi} N i=1 from N source images {Ii} N i=1 using a feature pyra￾mid network and CGA. These features are warped into the target camera frustum planes to construct the 3D cost volume C. Regularization vi… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison with state-of-the-art methods [5, 8, 23] using 3 input views on DTU [15], Real Forward-facing [24], NeRF Synthetic [25], and Tanks and Temples [18], arranged top to bottom. 4 Experiments 4.1 Experimental Setup Datasets. As in [5, 23], we train our model using the DTU training set [15] and evaluate on the DTU test set, applying the same dataset split configuration used in [23]. We eva… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of depth maps with MVSGaussian [23]. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: shows rendered images from the ablation in [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: shows more examples of the same experimental setup as [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of rendered images using 3 source views on Real Forward-facing [24], NeRF Synthetic [25], and Tanks and Temples [18] datasets. “Ours” refers to the generalization setting, while “Oursft” represents results obtained after per-scene optimization. Our approach already produces good results in the generalizable set￾ting, and for per-scene optimization, only a few additional iterations ar… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of rendered images using 3 source views on Real Forward-facing [24], NeRF Synthetic [25], and Tanks and Temples [18] datasets af￾ter fine-tuning. “MVSGaussianft" and “Oursft” represent results obtained after per-scene optimization [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Depth maps visualization and corresponding rendered images on DTU [15], Real Forward-facing [24], NeRF Synthetic [25], and Tanks and Temples [18] datasets. We can see that our depth map, as a good intermediate result, provides a solid foundation for the subsequent Gaussian expression. Source views Target view Ours [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Source views visualization under the 3-view setting. We select the three closest views as source views based on the distance of the viewpoints to render the target view. We also show the rendered images [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 38 canonical work pages · 2 internal anchors

  1. [1]

    Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth.Transactions on Machine Learning Research, 2023

    Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Multi-view stereo by learning robust image features and temperature-based depth.Transactions on Machine Learning Research, 2023

    work page 2023

  2. [2]

    DiffSSC: Semantic LiDAR Scan Completion using Denoising Diffusion Probabilistic Models

    Helin Cao and Sven Behnke. DiffSSC: Semantic LiDAR scan completion using de- noising diffusion probabilistic models. arXiv preprint arXiv:2409.18092, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    SLCF-Net: Sequential LiDAR-camera fusion for seman- tic scene completion using a 3D recurrent U-Net

    Helin Cao and Sven Behnke. SLCF-Net: Sequential LiDAR-camera fusion for seman- tic scene completion using a 3D recurrent U-Net. In IEEE International Conference on Robotics and Automation, pages 2767–2773, 2024

    work page 2024

  4. [4]

    pixel- splat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixel- splat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 19457–19467, 2024. HU ET AL.: C3-GS 11

    work page 2024

  5. [5]

    Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14124–14133, 2021

    work page 2021

  6. [6]

    Evloc: Event-based visual lo- calization in lidar maps via event-depth registration

    Kuangyi Chen, Jun Zhang, and Friedrich Fraundorfer. Evloc: Event-based visual lo- calization in lidar maps via event-depth registration. In IEEE International Conference on Robotics and Automation , 2025

    work page 2025

  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. In Proceedings of the European Conference on Computer Vision, pages 370–386, 2024

    work page 2024

  8. [8]

    Explicit correspondence matching for generalizable neural radiance fields

    Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  9. [9]

    Transmvsnet: Global context-aware multi-view stereo network with transformers

    Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8585–8594, 2022

    work page 2022

  10. [10]

    Object-centered surface reconstruction: Combining multi-image stereo and shading

    Pascal Fua and Yvan G Leclerc. Object-centered surface reconstruction: Combining multi-image stereo and shading. International Journal of Computer Vision , 16(1):35– 56, 1995

    work page 1995

  11. [11]

    Massively parallel multi- view stereopsis by surface normal diffusion

    Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multi- view stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873–881, 2015

    work page 2015

  12. [12]

    Cas- cade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cas- cade cost volume for high-resolution multi-view stereo and stereo matching. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2495–2504, 2020

    work page 2020

  13. [13]

    3d map recon- struction using a monocular camera for smart cities

    Yuxi Hu, Taimeng Fu, Guanchong Niu, Zixiao Liu, and Man-On Pun. 3d map recon- struction using a monocular camera for smart cities. The Journal of Supercomputing , 78(14):16512–16528, 2022

    work page 2022

  14. [14]

    Icg-mvsnet: Learning intra-view and cross-view rela- tionships for guidance in multi-view stereo

    Yuxi Hu, Jun Zhang, Zhe Zhang, Rafael Weilharter, Yuchen Rao, Kuangyi Chen, Runze Yuan, and Friedrich Fraundorfer. Icg-mvsnet: Learning intra-view and cross-view rela- tionships for guidance in multi-view stereo. In Proceedings of the IEEE International Conference on Multimedia and Expo , 2025

    work page 2025

  15. [15]

    Large scale multi-view stereopsis evaluation

    Rasmus Jensen, Anders Dahl, George V ogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 406–413, 2014

    work page 2014

  16. [16]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graph- ics, 42(4), 2023. 12 HU ET AL.: C3-GS

    work page 2023

  17. [17]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  18. [18]

    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 , 36 (4):1–13, 2017

    work page 2017

  19. [19]

    Srkd: Towards efficient 3d point cloud segmentation via structure-and relation-aware knowledge distillation

    Yuqi Li, Junhao Dong, Zeyu Dong, Chuanguang Yang, Zhulin An, and Yongjun Xu. Srkd: Towards efficient 3d point cloud segmentation via structure-and relation-aware knowledge distillation. arXiv preprint arXiv:2506.17290, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Frequency-Aligned Knowledge Distillation for Lightweight Spatiotemporal Forecasting

    Yuqi Li, Chuangang Yang, Hansheng Zeng, Zeyu Dong, Zhulin An, Yongjun Xu, Yingli Tian, and Hao Wu. Frequency-aligned knowledge distillation for lightweight spatiotemporal forecasting. arXiv:2507.02939, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Feature pyramid networks for object detection

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2117– 2125, 2017

    work page 2017

  22. [22]

    Geometry-aware reconstruction and fusion-refined rendering for generalizable neural radiance fields

    Tianqi Liu, Xinyi Ye, Min Shi, Zihao Huang, Zhiyu Pan, Zhan Peng, and Zhiguo Cao. Geometry-aware reconstruction and fusion-refined rendering for generalizable neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7654–7663, 2024

    work page 2024

  23. [23]

    Mvsgaussian: Fast generalizable gaussian splat- ting reconstruction from multi-view stereo

    Tianqi Liu, Guangcong Wang, Shoukang Hu, Liao Shen, Xinyi Ye, Yuhang Zang, Zhiguo Cao, Wei Li, and Ziwei Liu. Mvsgaussian: Fast generalizable gaussian splat- ting reconstruction from multi-view stereo. InProceedings of the European Conference on Computer Vision, pages 37–53, 2025

    work page 2025

  24. [24]

    Local light field fusion: Practical view synthesis with prescriptive sampling guidelines

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics , 38 (4):1–14, 2019

    work page 2019

  25. [25]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision , 2020

    work page 2020

  26. [26]

    Rethink- ing depth estimation for multi-view stereo a unified representation

    Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethink- ing depth estimation for multi-view stereo a unified representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 8645– 8654, 2022

    work page 2022

  27. [27]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 4104–4113, 2016

    work page 2016

  28. [28]

    Pix- elwise view selection for unstructured multi-view stereo

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pix- elwise view selection for unstructured multi-view stereo. In Proceedings of the Euro- pean Conference on Computer Vision, pages 501–518, 2016. HU ET AL.: C3-GS 13

    work page 2016

  29. [29]

    Splatter image: Ultra-fast single-view 3d reconstruction

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208–10217, 2024

    work page 2024

  30. [30]

    Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser

    Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    work page 2021

  31. [31]

    Image quality assessment: from error visibility to structural similarity

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

    work page 2004

  32. [32]

    Gmflow: Learning optical flow via global matching

    Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , pages 8121–8130, 2022

    work page 2022

  33. [33]

    Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views

    Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In Proceedings of the European Conference on Computer Vision , pages 203–220, 2024

    work page 2024

  34. [34]

    Dense hybrid recurrent multi-view stereo net with dynamic consistency checking

    Jianfeng Yan, Zizhuang Wei, Hongwei Yi, Mingyu Ding, Runze Zhang, Yisong Chen, Guoping Wang, and Yu-Wing Tai. Dense hybrid recurrent multi-view stereo net with dynamic consistency checking. In Proceedings of the European Conference on Com- puter Vision, pages 674–689. Springer, 2020

    work page 2020

  35. [35]

    Mvsnet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision, pages 767–783, 2018

    work page 2018

  36. [36]

    pixelNeRF: Neural radi- ance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radi- ance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    work page 2021

  37. [37]

    Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement

    Zehao Yu and Shenghua Gao. Fast-mvsnet: Sparse-to-dense multi-view stereo with learned propagation and gauss-newton refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1949–1958, 2020

    work page 1949

  38. [38]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 586– 595, 2018

    work page 2018

  39. [39]

    Geomvsnet: Learning multi- view stereo with geometry perception

    Zhe Zhang, Rui Peng, Yuxi Hu, and Ronggang Wang. Geomvsnet: Learning multi- view stereo with geometry perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21508–21518, 2023

    work page 2023

  40. [40]

    Ours” refer to the generalization setting in Table 2, while “MVSGaussian f t

    Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680–19690, 2024. HU ET AL.: C3-GS 1 In the supplementary mater...

    work page 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.