Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

Revisiting Depth Representations for Feed-Forward 3D Gaussian Splatting

T0 review · 3 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 This paper claims that the geometric weakness of depth-based feed-forward 3D Gaussian Splatting can be largely fixed by a training loss that distills global pointmaps from a pretrained reconstruction transformer into depth-unprojected…

desk verdict PM-Loss is a simple, plausible regularizer that consistently improves feed-forward 3DGS, but the headline >2 dB gains rely on a fine-tuning comparison that may conflate faster convergence with a better optimum. read the letter →

arxiv 2506.05327 v1 pith:HC2HN5YD submitted 2025-06-05 cs.CV

Duochao Shi , Weijie Wang , Donny Y. Chen , Zeyu Zhang , Jia-Wang Bian , Bohan Zhuang , Chunhua Shen This is my paper · ORCID
classification cs.CV
keywords 3DGaussianSplattingfeed-forward3DGSpointmapdepthunprojectionChamferlossnovelviewsynthesisgeometricregularizationknowledgedistillation
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

Feed-forward 3D Gaussian Splatting models that build Gaussians by unprojecting predicted depth maps inherit a known flaw: depth jumps at object boundaries fragment the point cloud and hurt rendering. This paper argues that the fix is to supervise those depth-derived Gaussian centers with a pointmap, a per-pixel 3D coordinate prediction from a pretrained reconstruction transformer, rather than with more depth supervision. The proposed PM-Loss aligns the pointmap to the predicted Gaussian centers with the closed-form Umeyama transform and then applies a one-directional Chamfer distance in 3D. Applied to MVSplat and DepthSplat on DL3DV and RealEstate10K, the loss raises extrapolation PSNR by roughly 2.05 to 2.66 dB and improves point-cloud accuracy and completeness on DTU. If this holds, depth-based feed-forward 3DGS can gain the geometric smoothness of pointmap models without adopting a pose-free architecture or adding any inference cost.

What carries the argument

The load-bearing object is the pointmap, a pixel-aligned map from image coordinates to world-space 3D points produced by a pretrained transformer, VGGT in the main experiments and Fast3R in an ablation. PM-Loss exploits the one-to-one pixel correspondence between the pointmap and the depth-unprojected Gaussian centers to replace expensive ICP registration with the closed-form Umeyama similarity transform, and then uses a single-directional Chamfer distance in 3D space to supervise the Gaussian centers. The design choice that makes the loss work is the nearest-neighbor computation in world coordinates: using the naive pixel correspondence would collapse the loss into a 2D depth loss, which the paper's ablation shows is worse.

What would settle it

Use a pointmap teacher whose training data is explicitly disjoint from DTU and from the DL3DV and RealEstate10K test splits, and check whether the DTU accuracy-completeness gains and the boundary-focused PSNR gains persist; if the gains disappear, the teacher's prior knowledge of the benchmark geometry, rather than geometric smoothing, is the cause.

Watch

Extended reading notes

Core claim

The paper's central claim is that depth discontinuities, not appearance prediction, are the main bottleneck for geometric quality in feed-forward 3DGS, and that a 3D geometric regularizer distilled from a pretrained pointmap model removes most of the damage. Even though the pointmap's absolute accuracy is lower than a well-tuned depth map in textured regions, its global smoothness and completeness near boundaries make it a better teacher for geometry. The paper encodes this as PM-Loss: unproject predicted depth into Gaussian centers, align the global pointmap to those centers with the Umeyama algorithm, then penalize each center's distance to its nearest neighbor in the aligned pointmap with a Chamfer term. The result is a plug-and-play training loss that improves both the unprojected 3D point clouds and the rendered novel views across different backbones, datasets, and numbers of input views.

Load-bearing premise

The load-bearing premise is that the pretrained pointmap model supplies a geometry prior that is independent of the evaluation benchmarks; if VGGT was trained on the same scenes used to measure DTU accuracy, the pseudo-ground-truth pointmaps encode the very geometry being tested.

Editorial extensions

If this is right

  • Any feed-forward 3DGS model that predicts depth and unprojects it can be fine-tuned with PM-Loss without changing its architecture, as demonstrated on MVSplat and DepthSplat.
  • Improved geometry at object boundaries should reduce the black or broken regions that depth-based models produce in extrapolated views.
  • DTU point-cloud metrics improve across 2-view, 4-view, and 6-view inputs, so the benefit is not specific to one training configuration.
  • PM-Loss adds only offline pointmap preprocessing and roughly 60 ms per training step, with no test-time overhead.
  • Because Fast3R also works as the pointmap source, the method is not tied to a single pretrained teacher model.

Reading between the lines

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

  • Editorial inference: because PM-Loss is decoupled from the Gaussian head, the same pointmap-distillation trick should transfer to other depth-unprojected predictors, including monocular depth-conditioned NeRFs or mesh generators, whenever a pixel-aligned global 3D prior exists.
  • Editorial inference: the paper's own limitation note implies the gains will be smallest in low-texture regions such as sky, where pointmap teachers are least reliable; measuring PSNR by image region would test this directly.
  • Editorial inference: a sharper experiment than the paper reports is to measure whether the PSNR gain concentrates in boundary pixels; if it does, the depth-discontinuity story is confirmed, and if it is uniform, the loss may be acting as generic geometry smoothing.
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

3 major / 5 minor

Summary. The paper proposes PM-Loss, a training-time regularization for feed-forward 3D Gaussian Splatting (3DGS) models. The key idea is to distill a geometry prior from a pretrained pointmap regression model (VGGT or Fast3R) into the depth-unprojected Gaussian centers. Because pointmaps are smoother near object boundaries than depth maps, the loss is intended to reduce the fragmented or floating geometry that arises from depth discontinuities. The authors apply PM-Loss to MVSplat and DepthSplat, fine-tuning both baseline and loss-augmented models for 100,000 iterations on DL3DV. They report consistent PSNR/SSIM/LPIPS gains on DL3DV and RealEstate10K under a custom extrapolation protocol, improved accuracy/completeness on DTU geometry evaluation, and ablations showing that a 3D nearest-neighbor Chamfer loss outperforms a 2D one-to-one depth loss and that higher-quality pointmaps help. The paper also documents the runtime and memory overhead of the loss, including an offline preprocessing mode.

Significance. If the central claim holds, PM-Loss is a simple, architecture-agnostic regularizer that improves both rendering and geometry quality of feed-forward 3DGS with modest extra cost. The method is clearly specified, the ablations are informative (Table 3 and Table 4), and the offline pointmap preprocessing makes the approach practical. The paper also includes a candid limitation statement about dependence on pointmap quality. However, the headline evidence is currently subject to two load-bearing concerns: the 100k-iteration fine-tuning comparison may conflate a better optimum with faster convergence, and the DTU geometry evaluation may be partly circular if VGGT was trained on DTU or geometrically equivalent data. The lack of variance or significance testing further weakens the quantitative claims. These issues are addressable with additional experiments and analysis, but they need to be resolved before the paper's main contribution can be fully accepted.

major comments (3)
  1. [§4.1, Table 1] The headline claim of >2 dB PSNR gains rests entirely on a 100,000-iteration fine-tuning schedule applied to both the baseline and the PM-Loss model, starting from public pretrained weights. Because PM-Loss injects dense 3D supervision from VGGT, it plausibly accelerates convergence, so the reported gap may reflect training speed rather than a better optimum. Please report learning curves (PSNR versus iterations) and extend the baseline (and ideally the PM variant) to a converged schedule, and state whether the 2.05–2.66 dB gains persist when the baseline is trained until its validation error plateaus.
  2. [§4.2, Table 2] The DTU geometry evaluation is vulnerable to circularity if VGGT was trained on DTU or on common MVS benchmarks that include DTU scenes. Since the pointmap pseudo-ground truth then encodes the same geometry used for evaluation, the improved Accuracy/Completeness/Overall scores may partly reflect prior exposure rather than a generic geometric regularizer. The paper should disclose VGGT's training-data composition or demonstrate the benefit on a geometry benchmark that is provably disjoint from VGGT's training set. I note that the pointmaps are computed on DL3DV training scenes, so this is not strict test-set leakage, but the transfer claim still requires the overlap check.
  3. [§4.1, Appendix A, Table A] The main-table evaluation uses a custom extrapolation protocol, and the gains there (e.g., DepthSplat +2.31 dB on DL3DV) are much larger than on the interpolation setting in Table A (+0.81 dB). The manuscript should justify why extrapolation is the appropriate primary measure rather than a setting that happens to amplify the effect of the regularizer, and it should report variance across seeds or at least paired per-scene significance tests for the headline numbers. Without error bars, the magnitude of the claimed improvement is uncertain.
minor comments (5)
  1. [§3.2, Eq. (8)] The loss is called a 'single-directional Chamfer distance,' which is technically a one-sided Chamfer distance; since symmetric Chamfer is the more common usage, please clarify the terminology in the text after Eq. (8).
  2. [§4.2, Table 7 vs. Appendix B, Table B] The main text reports a +6.09 GB VRAM increase for PM-Loss, while Appendix B reports only +0.96 GB when pointmaps are preprocessed offline. The offline mode is the recommended workflow, so please report that number in the main text and describe the online/offline distinction next to Table 7.
  3. [§4.2] The sentence 'a consistent gain of at least 2 dB in PSNR' should be qualified as applying to Table 1; the DTU results in Table 2 use different metrics and do not support a dB statement.
  4. [Appendix C, Table C] The comparison with NoPoSplat uses MVSplat's PSNR of 26.39, which is much higher than the MVSplat numbers in Table 1 (19.52 on RealEstate10K and 16.79 on DL3DV). The caption should state the evaluation protocol for Table C, otherwise the comparison is difficult to interpret.
  5. [§4.1, Implementation Details] The text says both sets of models were fine-tuned using 'the same batch size,' but the appendix shows DepthSplat uses batch size 1 and MVSplat uses batch size 12. Please clarify that this refers to equal batch size within each backbone, not across backbones.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PM-Loss is an empirical regularizer with an external pointmap prior; all headline results are benchmark comparisons, not constructions.

full rationale

PM-Loss is an explicit training regularizer: Eq. (8) defines a single-directional Chamfer loss between depth-unprojected Gaussian centers X3DGS and an aligned external pointmap X'_PM produced by a pretrained model (VGGT/Fast3R). The paper's central claims are empirical benchmark comparisons (Tables 1-2), not derivations: they fine-tune MVSplat/DepthSplat with and without the loss and measure NVS and point-cloud metrics. The loss is not defined in terms of the evaluation metrics, and the reported gains are not by construction equal to the loss; the Umeyama alignment (Eq. 7) and Chamfer distance (Eq. 8) are standard geometric operations with no hidden fitted parameter that is later reported as a prediction. The self-citations (e.g., MVSplat [8], Zpressor [42], SC-DepthV3 [16]) appear only as baselines or related work, not as load-bearing justifications of the method. The paper explicitly acknowledges in its Limitation section that PM-Loss inherits pointmap errors, which confirms the prior is an external inductive bias rather than an equivalent restatement of the output. The only substantive concern is data hygiene: the paper does not report whether VGGT was trained on the evaluation distributions (DTU in particular), and it criticizes Splatt3R for testing on a dataset used by its pointmap model. That concern, if substantiated, would be test-set leakage, not circularity by construction; there is no equation or fitted parameter in this paper that makes the DTU or NVS results equal to the PM-Loss input.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three domain assumptions: the pointmap prior is accurate and smooth enough to supervise geometry; the Umeyama alignment is robust to boundary outliers; and the pointmap model was not trained on the evaluation benchmarks. The only free parameter introduced by the paper is the loss weight lambda_PM. No new physical or conceptual entities are invented; pointmaps and Chamfer distance are standard tools from prior work.

free parameters (1)
  • lambda_PM = 0.005
    Weighting coefficient for PM-Loss in the total loss (Eq. 9). Set to a single value without a reported sensitivity analysis.
assumptions (3)
  • domain assumption The pretrained pointmap model (VGGT) provides a reliable geometry prior, including at object boundaries, that is more globally smooth and complete than depth-map unprojection.
    Section 3.2 states that pointmaps 'tend to exhibit better geometric smoothness and completeness, especially near object boundaries,' and this is the basis for using them as pseudo-ground-truth. The paper acknowledges in the Limitation section that errors in the pointmap propagate.
  • domain assumption The Umeyama algorithm can align the two point clouds using known pixel correspondences without being destabilized by outliers at depth discontinuities.
    Section 3.2 relies on one-to-one pixel correspondence to estimate a global similarity transform (Eq. 7). If boundary outliers dominate, the alignment could be biased; the paper does not analyze robustness to such outliers.
  • domain assumption The evaluation benchmarks (DTU, RealEstate10K, DL3DV) are not contained in the pointmap model's pretraining data in a way that leaks ground-truth geometry into the pseudo-supervision.
    Used in Section 4.1 when applying VGGT to DTU and then evaluating DTU point cloud quality. The paper does not verify this independence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Revisiting Depth Representations for Feed-Forward 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/HC2HN5YD

@misc{pith2026250605327,
  author       = {Pith},
  title        = {Pith review of: Revisiting Depth Representations for Feed-Forward 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HC2HN5YD}},
  note         = {Machine review of arXiv:2506.05327}
}
read the original abstract

Depth maps are widely used in feed-forward 3D Gaussian Splatting (3DGS) pipelines by unprojecting them into 3D point clouds for novel view synthesis. This approach offers advantages such as efficient training, the use of known camera poses, and accurate geometry estimation. However, depth discontinuities at object boundaries often lead to fragmented or sparse point clouds, degrading rendering quality -- a well-known limitation of depth-based representations. To tackle this issue, we introduce PM-Loss, a novel regularization loss based on a pointmap predicted by a pre-trained transformer. Although the pointmap itself may be less accurate than the depth map, it effectively enforces geometric smoothness, especially around object boundaries. With the improved depth map, our method significantly improves the feed-forward 3DGS across various architectures and scenes, delivering consistently better rendering results. Our project page: https://aim-uofa.github.io/PMLoss

Figures

Figures reproduced from arXiv: 2506.05327 by the authors.

Figure 1
Figure 1. Feed-forward 3DGS models, e.g., DepthSplat [9], rely on unprojected depth to form 3D Gaussians. The inherent discontinuities of depth near object boundaries often propagate into distorted 3D point clouds (top left) and degraded rendering (bottom left). Our PM-Loss addresses this by using the prior from pointmap, achieving higher-quality geometry (top right) and rendering (bottom right). (MVS) [18, 19] process by ref… view at source ↗
Figure 2
Figure 2. Overview of PM-Loss. The process begins by estimating a dense point map of the scene using a pre-trained model. This estimated point map then serves as direct 3D supervision for training a feed-forward 3D Gaussian Splatting model. Crucially, unlike conventional methods relying predominantly on 2D supervision, our approach leverages explicit 3D geometric cues, leading to enhanced 3D shape fidelity. 3 Methodology Our … view at source ↗
Figure 3
Figure 3. Qualitative comparisons on DL3DV(top two rows) and RealEstate10K(bottom two rows) under the 2-view extrapolation setting. Adding PM-Loss leads to significant improvements in rendering object boundaries. The estimated transformation (s ∗ , R∗ , t∗ ) is then applied to each point p ′ k in the original pointmap XPM to obtain the aligned pointmap X′ PM = {s ∗R∗p ′ k + t ∗} Ntotal_pts k=1 . This alignment enables us to c… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison of unprojected 3D Gaussians on DL3DV dataset. Our method effectively constrains the 3D Gaussians, significantly reducing floating artifacts and noise near border. Input Views Depthsplat Depthsplat + PM-Loss [PITH_FULL_IMAGE:figures/full_fig_p008…
Figure 5
Figure 5. Figure 5: Qualitative comparison of unprojected 3D Gaussians on the DTU dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. VolSplat: Rethinking Feed-Forward 3D Gaussian Splatting with Voxel-Aligned Prediction

    cs.CV 2025-09 conditional novelty 6.0 of 10

    VolSplat predicts 3D Gaussians from a shared voxel grid instead of from image pixels, reporting large gains in sparse-view novel view synthesis on RealEstate10K, ScanNet, and ACID.

  2. DriveGen3D: Boosting Feed-Forward Driving Scene Generation with Efficient Video Diffusion

    cs.CV 2025-10 conditional novelty 4.0 of 10

    DriveGen3D makes long driving-video synthesis and 3D scene reconstruction practical by caching only the conditional diffusion branch, quantizing cross-view attention, and fusing temporal context into a feed-forward Ga...

  3. Visual Execution and Validation of Finite-State Machines and Pushdown Automata

    cs.FL 2025-08 unverdicted novelty 4.0 of 10

    Two new visualization tools for the FSM language step through all computations of nondeterministic finite-state machines and pushdown automata and let users check state properties during transitions.

Reference graph

Works this paper leans on

52 extracted references · 28 canonical work pages · cited by 3 Pith papers

  1. [1]

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

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1, 2023

    2023

  2. [2]

    2d gaussian splatting for geometrically accurate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. InACM SIGGRAPH 2024 conference papers, pages 1–11, 2024

    2024

  3. [3]

    Trim 3d gaussian splatting for accurate geometry representation.arXiv preprint arXiv:2406.07499, 2024

    Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation.arXiv preprint arXiv:2406.07499, 2024

    arXiv 2024

  4. [4]

    3dgsr: Implicit surface reconstruction with 3d gaussian splatting.ACM Transactions on Graphics (TOG), 43(6):1–12, 2024

    Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Xiuzhe Wu, Ziyi Yang, Yilun Chen, Jiangmiao Pang, and Xiaojuan Qi. 3dgsr: Implicit surface reconstruction with 3d gaussian splatting.ACM Transactions on Graphics (TOG), 43(6):1–12, 2024

    work page 2024

  5. [5]

    Surface reconstruction from gaussian splatting via novel stereo views.arXiv e-prints, pages arXiv–2404, 2024

    Yaniv Wolf, Amit Bracha, and Ron Kimmel. Surface reconstruction from gaussian splatting via novel stereo views.arXiv e-prints, pages arXiv–2404, 2024

    work page 2024

  6. [6]

    Gsdf: 3dgs meets sdf for improved rendering and reconstruction.arXiv preprint arXiv:2403.16964, 2024

    Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction.arXiv preprint arXiv:2403.16964, 2024

    arXiv 2024

  7. [7]

    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. InCVPR, pages 19457–19467, 2024

    2024

  8. [8]

    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. InECCV, pages 370–386. Springer, 2024

    work page 2024

Show all 52 references
  1. [9]

    Depthsplat: Connecting gaussian splatting and depth

    Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. InCVPR, 2025

    2025

  2. [10]

    Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image.arXiv preprint arXiv:2406.04343, 2024

    Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F Hen- riques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image.arXiv preprint arXiv:2406.04343, 2024

    2024 arXiv

  3. [11]

    Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats

    Chen Ziwen, Hao Tan, Kai Zhang, Sai Bi, Fujun Luan, Yicong Hong, Li Fuxin, and Zexiang Xu. Long-lrm: Long-sequence large reconstruction model for wide-coverage gaussian splats. arXiv preprint arXiv:2410.12781, 2024

    2024 arXiv

  4. [12]

    Gs-lrm: Large reconstruction model for 3d gaussian splatting

    Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. InECCV, pages 1–19. Springer, 2024

    2024

  5. [13]

    Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.NeurIPS, 37:107326–107349, 2024

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat: Generalizable 3d gaussian splatting towards free view synthesis of indoor scenes.NeurIPS, 37:107326–107349, 2024

    2024

  6. [14]

    Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers

    Chuanrui Zhang, Yingshuang Zou, Zhuoling Li, Minmin Yi, and Haoqian Wang. Transplat: Generalizable 3d gaussian splatting from sparse multi-view images with transformers. InAAAI, volume 39, pages 9869–9877, 2025

    2025

  7. [15]

    Predicting sharp and accurate occlusion boundaries in monocular depth estimation using displacement fields

    Michael Ramamonjisoa, Yuming Du, and Vincent Lepetit. Predicting sharp and accurate occlusion boundaries in monocular depth estimation using displacement fields. InCVPR, pages 14648–14657, 2020

    2020

  8. [16]

    Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes.IEEE transactions on pattern analysis and machine intelligence, 46(1):497–508, 2023

    Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes.IEEE transactions on pattern analysis and machine intelligence, 46(1):497–508, 2023

    2023

  9. [17]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InCVPR, pages 20697–20709, 2024

    2024

  10. [18]

    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. InECCV, pages 767–783, 2018

    2018

  11. [19]

    Cascade cost volume for high-resolution multi-view stereo and stereo matching

    Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. InCVPR, pages 2495–2504, 2020. 11

    2020

  12. [20]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass

    Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. InCVPR, 2025

    2025

  13. [21]

    Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds

    Zhenggang Tang, Yuchen Fan, Dilin Wang, Hongyu Xu, Rakesh Ranjan, Alexander Schwing, and Zhicheng Yan. Mv-dust3r+: Single-stage scene reconstruction from sparse views in 2 seconds. InCVPR, 2025

    2025

  14. [22]

    Monst3r: A simple approach for estimating geometry in the presence of motion.arXiv preprint arXiv:2410.03825, 2024

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

    arXiv

  15. [23]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InCVPR, 2025

    2025

  16. [24]

    Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision

    Ruicheng Wang, Sicheng Xu, Cassie Dai, Jianfeng Xiang, Yu Deng, Xin Tong, and Jiaolong Yang. Moge: Unlocking accurate monocular geometry estimation for open-domain images with optimal training supervision. InCVPR, 2025

    2025

  17. [25]

    3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024

    2024 arXiv

  18. [26]

    Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs.arXiv preprint arXiv:2408.13912, 2024

    Brandon Smart, Chuanxia Zheng, Iro Laina, and Victor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs.arXiv preprint arXiv:2408.13912, 2024

    2024 arXiv

  19. [27]

    No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images

    Botao Ye, Sifei Liu, Haofei Xu, Xueting Li, Marc Pollefeys, Ming-Hsuan Yang, and Songyou Peng. No pose, no problem: Surprisingly simple 3d gaussian splats from sparse unposed images. InICLR, 2025

    2025

  20. [28]

    Scannet++: A high-fidelity dataset of 3d indoor scenes

    Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. InICCV, pages 12–22, 2023

    2023

  21. [29]

    Stereo magnifica- tion: learning view synthesis using multiplane images.ACM Transactions on Graphics (TOG), 37(4):1–12, 2018

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnifica- tion: learning view synthesis using multiplane images.ACM Transactions on Graphics (TOG), 37(4):1–12, 2018

    2018

  22. [30]

    Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision

    Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision. InCVPR, pages 22160–22169, 2024

    2024

  23. [31]

    View interpolation for image synthesis

    Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. InPro- ceedings of the 20th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH ’93, page 279–288, New York, NY , USA, 1993. Association for Computing Machinery

    1993

  24. [32]

    View morphing

    Steven M Seitz and Charles R Dyer. View morphing. InProceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 21–30, 1996

    1996

  25. [33]

    Nerf: Representing scenes as neural radiance fields for view synthesis

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

    2021

  26. [34]

    Depth-regularized optimization for 3d gaussian splatting in few-shot images, 2024

    Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images, 2024

    2024

  27. [35]

    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. InECCV, 2024

    2024

  28. [36]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InCVPR, 2024

    2024

  29. [37]

    Depth anything v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. InNeurIPS, 2024

    2024

  30. [38]

    Freesplat++: Generalizable 3d gaussian splatting for efficient indoor scene reconstruction, 2025

    Yunsong Wang, Tianxin Huang, Hanlin Chen, and Gim Hee Lee. Freesplat++: Generalizable 3d gaussian splatting for efficient indoor scene reconstruction, 2025

    2025

  31. [39]

    Epipolar-free 3d gaussian splatting for generalizable novel view synthesis, 2024

    Zhiyuan Min, Yawei Luo, Jianwen Sun, and Yi Yang. Epipolar-free 3d gaussian splatting for generalizable novel view synthesis, 2024

    2024

  32. [40]

    Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views, 2024

    Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Pixelgaussian: Generalizable 3d gaussian reconstruction from arbitrary views, 2024. 12

    2024

  33. [41]

    Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting

    Gyeongjin Kang, Jisang Yoo, Jihyeon Park, Seungtae Nam, Hyeonsoo Im, Sangheon Shin, Sangpil Kim, and Eunbyung Park. Selfsplat: Pose-free and 3d prior-free generalizable 3d gaussian splatting. InCVPR, 2025

    2025

  34. [42]

    Chen, Zeyu Zhang, Duochao Shi, Akide Liu, and Bohan Zhuang

    Weijie Wang, Donny Y . Chen, Zeyu Zhang, Duochao Shi, Akide Liu, and Bohan Zhuang. Zpressor: Bottleneck-aware compression for scalable feed-forward 3dgs, 2025

    2025

  35. [43]

    Mvsplat360: Feed-forward 360 scene synthesis from sparse views.NeurIPS, 37:107064–107086, 2024

    Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views.NeurIPS, 37:107064–107086, 2024

    2024

  36. [44]

    Grounding image matching in 3d with mast3r, 2024

    Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024

    2024

  37. [45]

    Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views

    Shangzhan Zhang, Jianyuan Wang, Yinghao Xu, Nan Xue, Christian Rupprecht, Xiaowei Zhou, Yujun Shen, and Gordon Wetzstein. Flare: Feed-forward geometry, appearance and camera estimation from uncalibrated sparse views. InCVPR, 2025

    2025

  38. [46]

    Efros, and Angjoo Kanazawa

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InCVPR, 2025

    2025

  39. [47]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  40. [48]

    Method for registration of 3-d shapes

    Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. InSensor fusion IV: control paradigms and data structures, volume 1611, pages 586–606. Spie, 1992

    1992

  41. [49]

    S. Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(4):376–380, 1991

    1991

  42. [50]

    Large scale multi-view stereopsis evaluation

    Rasmus Jensen, Anders Dahl, George V ogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. InCVPR, pages 406–413, 2014

    2014

  43. [51]

    Explicit correspondence matching for generalizable neural radiance fields, 2023

    Yuedong Chen, Haofei Xu, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Explicit correspondence matching for generalizable neural radiance fields, 2023

    2023

  44. [52]

    Gaussian heads

    Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d, 2020. 13 A More Experimental Analysis Evaluation under view interpolation settings.Beyond the main paper’s extrapolat...

    2020

Pith tools

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