Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Review of Feed-forward 3D Reconstruction: From DUSt3R to VGGT

T0 review · 2 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This survey argues that 3D reconstruction is moving from iterative, fragile SfM+MVS pipelines toward unified feed-forward Transformer networks that output camera poses and dense geometry in a single pass.

desk verdict Useful survey of feed-forward 3D reconstruction with a solid taxonomy, but Table 4's protocol-mixing undercuts the quantitative support for its paradigm-shift claim. read the letter →

arxiv 2507.08448 v1 pith:CVD7KRTJ submitted 2025-07-11 cs.CV cs.AI

Wei Zhang , Yihang Wu , Songhua Li , Wenjie Ma , Xin Ma , Qiang Li , Qi Wang This is my paper
classification cs.CVcs.AI
keywords 3Dreconstructionfeed-forwardmodelDUSt3Rmulti-viewstereocameraposeestimationdensegeometryTransformerVGGT
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

This survey argues that 3D reconstruction is undergoing a paradigm shift: instead of chaining feature matching, RANSAC, bundle adjustment, and multi-view stereo into a fragile iterative pipeline, a new class of feed-forward models, led by DUSt3R, uses a single Transformer-based network to infer camera poses and dense geometry from unconstrained image sets in one forward pass. The review maps the shared architecture, covering dense correspondence learning, joint pose-geometry regression, and strategies for scaling from two views to many views, and contrasts it with classical SfM/MVS and modular deep learning such as MVSNet. The reason to care is that if this direction holds, 3D reconstruction becomes robust in texture-poor and wide-baseline conditions, runs on GPU in seconds rather than minutes or days on CPU, and requires no expert tuning. The survey also names the open problems: accuracy, scalability, dynamic scenes, and uncertainty quantification.

What carries the argument

The central object is the unified feed-forward Transformer network exemplified by DUSt3R. Its key design is that instead of regressing rotation and translation directly, the network predicts per-pixel 3D coordinates in the first image's normalized camera frame, and a differentiable Kabsch-Umeyama alignment between the predicted point cloud and the second view's projected points yields the relative pose while the predicted coordinates themselves form the dense geometry. This pose-from-alignment coupling lets pose and geometry supervise each other. Scaling to many views is carried by confidence-weighted pose-graph or joint point-cloud aggregation methods such as Align3R and Pow3R, with a few emerging true multi-view attention models such as MV-DUSt3R+ that ingest more than two images directly.

What would settle it

Recheck Table 4 against the original DTU benchmark papers for DUSt3R, MASt3R, and VGGT; the missing COLMAP entries and the mismatched scale of DUSt3R's rel value (3.3 versus MVSNet's 1.8) are concrete places where a transcription error would falsify the survey's comparative claim. Alternatively, run a single two-view DUSt3R reconstruction on a texture-less scene and compare runtime and accuracy directly against COLMAP with MVSNet densification.

Watch

Extended reading notes

Core claim

The paper's central claim is that the dominant workflow of 3D reconstruction is moving from 'iterative optimization' to 'end-to-end inference'. It identifies DUSt3R as the seminal work that unified pose and dense geometry prediction, and traces how successors such as MASt3R, VGGT, Align3R, Pow3R, SLAM3R, Driv3R, and Reloc3r share a common technical core: a Transformer-based matcher producing dense probabilistic correspondences, a regression head that decodes 3D coordinates rather than a rotation matrix and translation vector directly, with a differentiable Kabsch-Umeyama alignment step yielding the pose, and aggregation or attention-based strategies to go from pairs to many views. Against COLMAP-style pipelines and MVSNet-style modular learning, the feed-forward family trades some measured accuracy, as reflected in the DTU benchmark table the survey compiles, for greater robustness and operational simplicity.

Load-bearing premise

The synthesis depends on the accuracy of its quantitative benchmark transcription: if the DTU comparisons in Table 4 contain misread or incomparable numbers, the review's evidence that feed-forward models trade accuracy for robustness is unreliable.

Editorial extensions

If this is right

  • If the shift holds, dense reconstruction of a room from a smartphone becomes a seconds-scale GPU operation, with no RANSAC or bundle adjustment.
  • Robustness in texture-less, wide-baseline, and non-Lambertian scenes improves because learned priors substitute for explicit feature detection and geometric verification.
  • The primary bottlenecks move from geometric optimization to GPU memory, training-data scale and diversity, and generalization to out-of-distribution domains.
  • Hybrid pipelines that use a feed-forward network for a robust initial estimate and a lightweight optimization layer for refinement become a natural middle path for high-precision applications.
  • Direct integration with neural rendering, such as NeRF or 3D Gaussian Splatting outputs, could enable real-time novel-view synthesis from sparse image inputs.

Reading between the lines

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

  • The DTU numbers in the survey hint that the accuracy gap between feed-forward models and classical MVS is closing quickly, with VGGT's Overall score near those of modular deep-learning baselines, so the robustness-for-accuracy trade-off may be a temporary phase rather than a permanent property of the paradigm.
  • The review's framework implies that evaluation conventions should be rethought: DTU Accuracy and Completeness are designed for pipelines with known poses, while feed-forward models are tested without ground-truth poses, making cross-paradigm comparisons ambiguous.
  • A testable extension would be training a feed-forward model on the blend of datasets the survey recommends, such as MegaDepth, CO3D-v2, ARKitScenes, and Waymo, and measuring zero-shot performance on DTU and Tanks and Temples to quantify the generalization claims.
  • The trajectory the survey sketches toward geometrically-aware large language models suggests that reconstruction outputs may eventually become tokens in a multimodal model rather than standalone point clouds, changing what '3D reconstruction' means as a research target.
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

2 major / 6 minor

Summary. This manuscript is a survey of the emerging family of feed-forward 3D reconstruction models that follow DUSt3R, culminating in systems such as VGGT. It proposes that this line of work constitutes a paradigm shift away from the classical multi-stage SfM/MVS pipeline, and it organizes the field into three technical pillars: learned dense correspondence, joint pose-and-geometry regression, and strategies for scaling from two-view to multi-view inference. The survey also provides a table of representative models, a table of training datasets, a quantitative comparison table on DTU, and a discussion of applications and open challenges. The paper does not introduce new algorithms or experiments; its contribution is the synthesis and the framing of the feed-forward paradigm.

Significance. The survey addresses a timely and rapidly growing area, and its qualitative taxonomy is broadly consistent with the cited literature. The model table and the application survey are useful entry points for researchers. The paper also identifies important open problems such as scalability, dynamic scenes, uncertainty quantification, and integration with neural rendering, and it correctly notes the shift of bottlenecks from geometric optimization to GPU memory, data curation, and out-of-distribution generalization. However, the significance of the central claim depends on the reliability of Table 4, which is currently compromised by protocol mixing and missing baseline entries, and on a precise definition of 'single forward pass' that is not consistently maintained in the text.

major comments (2)
  1. [Section 4, Table 4] Table 4 is the only quantitative evidence supporting the survey's characterization of feed-forward models, but as printed it is not internally consistent. The COLMAP row, which is the canonical baseline used throughout Sections 1 and 3, contains no Accuracy/Completeness/Overall entries, so the decisive classical-pipeline comparison is missing. The inlier-ratio column is headed 'τ ↓', which reverses the direction of the metric: higher inlier ratios are better. Rows are drawn from different sources without stating the DTU split, the depth-fusion procedure, or the metric implementation; for example, DUSt3R's Overall=1.741 sits beside VGGT's Overall=0.382, and MVSNet's rel values are parenthesized while DUSt3R's are not. Because the surrounding text says these standardized protocols facilitate fair comparison, the table as it stands cannot substantiate the claim that feed-forward models trade some accuracy for robustness. The authors should either recompute all rows under one protocol or clearly separate depth and reconstruction evaluations and state that cross-row comparisons are only qualitative.
  2. [Sections 1, 2, and 3] The central 'single forward pass' claim is not consistent with the paper's own technical description. Section 2, in the subsection on scaling from two-view to multi-view scenarios, states that most models first run a two-view network on image pairs and then globally aggregate the pairwise estimates, and Section 3 even contrasts this as a 'concatenation of two inference steps' that avoids iterative loops. Aggregation methods such as Align3R and Pow3R are, by the paper's own description, pose-graph or joint optimization algorithms, so they do involve iterative optimization. The abstract and Section 1 nevertheless present feed-forward models as directly inferring globally consistent poses and dense geometry in a single forward pass. This overstatement should be corrected by distinguishing single-pass architectures such as VGGT and MV-DUSt3R+ from pairwise-plus-global-alignment systems, and by qualifying the claimed break with iterative optimization.
minor comments (6)
  1. [Section 4, Table 4] The header 'τ ↓' should read 'τ ↑' because a higher inlier ratio indicates better performance.
  2. [Table 3] The Habitat row cites reference [29], which is the Habitat platform paper, not the Matterport3D dataset; either rename the dataset or cite the original Matterport3D source.
  3. [Table 3] The DL3DV row lists 2022 and 'Outdoor', but the cited DL3DV-10K paper is from CVPR 2024 and the dataset contains both indoor and outdoor scenes.
  4. [Section 2] The spelling of SPANN3R in the text is inconsistent with the 'Spann3R' entry in Table 1 and reference [19]; please standardize the name.
  5. [Section 4, Table 4] The parentheses around MVSNet's rel and τ values are unexplained; if they indicate a different rescaling, please state this in a footnote or table note.
  6. [Section 5] The claim that attention's quadratic complexity is a fundamental bottleneck should be qualified, since several cited models such as SPANN3R and CUT3R explicitly address this with sparse or persistent-state attention.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: this is a survey whose claims rest on external cited works, not on fitted inputs or self-citations.

full rationale

This is a literature survey, not a derivation. Its central claim that feed-forward models mark a paradigm shift from iterative optimization to end-to-end inference is supported by citing independent external papers (DUSt3R, MASt3R, VGGT) and by an architectural summary, not by any equation that reduces to its inputs. There are no fitted parameters renamed as predictions, no self-citation chain invoked as an external uniqueness theorem, and no defined quantity that silently contains the target conclusion. The DTU table inconsistencies flagged by the reader (missing COLMAP Acc/Comp/Overall entries, DUSt3R rel=3.3 on a different scale than MVSNet) are genuine concerns about benchmark transcription and protocol consistency; they bear on correctness and evidential support, but they are not circularity because the survey does not use those numbers as inputs to derive its claim. A future protocol-consistent benchmark could change the paper's quantitative characterization, but that would be a factual correction, not a circular dependency. Score 0.

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

The paper is a review and introduces no new parameters, axioms, or entities. It relies on standard geometry and deep learning concepts from the cited literature.

assumptions (1)
  • domain assumption The cited papers are accurately summarized, especially the quantitative results in Table 4.
    The review's synthesis depends on this assumption; any mis-transcription would alter the comparative claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Review of Feed-forward 3D Reconstruction: From DUSt3R to VGGT." pith.science (2026). https://pith.science/paper/CVD7KRTJ

@misc{pith2026250708448,
  author       = {Pith},
  title        = {Pith review of: Review of Feed-forward 3D Reconstruction: From DUSt3R to VGGT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CVD7KRTJ}},
  note         = {Machine review of arXiv:2507.08448}
}
read the original abstract

3D reconstruction, which aims to recover the dense three-dimensional structure of a scene, is a cornerstone technology for numerous applications, including augmented/virtual reality, autonomous driving, and robotics. While traditional pipelines like Structure from Motion (SfM) and Multi-View Stereo (MVS) achieve high precision through iterative optimization, they are limited by complex workflows, high computational cost, and poor robustness in challenging scenarios like texture-less regions. Recently, deep learning has catalyzed a paradigm shift in 3D reconstruction. A new family of models, exemplified by DUSt3R, has pioneered a feed-forward approach. These models employ a unified deep network to jointly infer camera poses and dense geometry directly from an Unconstrained set of images in a single forward pass. This survey provides a systematic review of this emerging domain. We begin by dissecting the technical framework of these feed-forward models, including their Transformer-based correspondence modeling, joint pose and geometry regression mechanisms, and strategies for scaling from two-view to multi-view scenarios. To highlight the disruptive nature of this new paradigm, we contrast it with both traditional pipelines and earlier learning-based methods like MVSNet. Furthermore, we provide an overview of relevant datasets and evaluation metrics. Finally, we discuss the technology's broad application prospects and identify key future challenges and opportunities, such as model accuracy and scalability, and handling dynamic scenes.

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. Event3R: Asynchronous-to-Global 3D Reconstruction from Event Camera via Spatial-Temporal Feature Aggregation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Event3R predicts globally aligned 3D point clouds directly from asynchronous event streams using a DUSt3R-style feed-forward network with temporal attention and masked-bin self-supervision.

Reference graph

Works this paper leans on

54 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [1]

    SIFT algorithm analysis and optimization,

    Z. Daixian, “SIFT algorithm analysis and optimization,” in 2010 International Conference on Image Analysis and Signal Processing, Apr. 2010, pp. 415–419

    work page 2010

  2. [2]

    Structure-From-Motion Revisited,

    J. L. Schonberger and J.-M. Frahm, “Structure-From-Motion Revisited,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 4104–4113

    work page 2016

  3. [3]

    SuperPoint: Self-Supervised Interest Point Detection and Description,

    D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperPoint: Self-Supervised Interest Point Detection and Description,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 224–236

    work page 2018

  4. [4]

    SuperGlue: Learning Feature Matching With Graph Neural Networks,

    P.-E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich, “SuperGlue: Learning Feature Matching With Graph Neural Networks,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 4938–4947

    work page 2020

  5. [5]

    MVSNet: Depth Inference for Unstructured Multi-view Stereo,

    Y. Yao, Z. Luo, S. Li, T. Fang, and L. Quan, “MVSNet: Depth Inference for Unstructured Multi-view Stereo,” in Computer Vision – ECCV 2018, vol. 11212, V. Ferrari, M. Hebert, C. Sminchisescu, and Y. Weiss, Eds., in Lecture Notes in Computer Science, vol. 11212. , Cham: Springer International Publishing, 2018, pp. 785–801

    work page 2018

  6. [6]

    Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo Matching,

    X. Gu, Z. Fan, S. Zhu, Z. Dai, F. Tan, and P. Tan, “Cascade Cost Volume for High-Resolution Multi-View Stereo and Stereo Matching,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2020, pp. 2492–2501

    work page 2020

  7. [7]

    PatchmatchNet: Learned Multi-View Patchmatch Stereo,

    F. Wang, S. Galliani, C. Vogel, P. Speciale, and M. Pollefeys, “PatchmatchNet: Learned Multi-View Patchmatch Stereo,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14194–14203

    work page 2021

  8. [8]

    DUSt3R: Geometric 3D Vision Made Easy,

    S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud, “DUSt3R: Geometric 3D Vision Made Easy,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2024, pp. 20697–20709

    work page 2024

Show all 54 references
  1. [9]

    Attention is All you Need,

    A. Vaswani et al., “Attention is All you Need,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2017

    2017

  2. [10]

    Grounding Image Matching in 3D with MASt3R,

    V. Leroy, Y. Cabon, and J. Revaud, “Grounding Image Matching in 3D with MASt3R,” in Computer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds., Cham: Springer Nature Switzerland, 2025, pp. 71–91

    2024

  3. [11]

    VGGT: Visual Geometry Grounded Transformer,

    J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “VGGT: Visual Geometry Grounded Transformer,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5294–5306

    2025

  4. [12]

    Align3R: Aligned Monocular Depth Estimation for Dynamic Videos,

    J. Lu et al., “Align3R: Aligned Monocular Depth Estimation for Dynamic Videos,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 22820–22830

    2025

  5. [13]

    Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors,

    W. Jang, P. Weinzaepfel, V. Leroy, L. Agapito, and J. Revaud, “Pow3R: Empowering Unconstrained 3D Reconstruction with Camera and Scene Priors,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1071–1081

    2025

  6. [14]

    SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos,

    Y. Liu et al., “SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16651–16662

    2025

  7. [15]

    Driv3R: Learning Dense 4D Reconstruction for Autonomous Driving,

    X. Fei et al., “Driv3R: Learning Dense 4D Reconstruction for Autonomous Driving,” Dec. 09, 2024, arXiv: arXiv:2412.06777

    2024 arXiv

  8. [16]

    Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization,

    S. Dong et al., “Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localization,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 16739–16752

    2025

  9. [17]

    LoFTR: Detector-Free Local Feature Matching With Transformers,

    J. Sun, Z. Shen, Y. Wang, H. Bao, and X. Zhou, “LoFTR: Detector-Free Local Feature Matching With Transformers,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8922–8931

    2021

  10. [18]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” presented at the International Conference on Learning Representations, Oct. 2020

    2020

  11. [19]

    3D Reconstruction with Spatial Memory,

    H. Wang and L. Agapito, “3D Reconstruction with Spatial Memory,” Aug. 28, 2024, arXiv: arXiv:2408.16061

    2024 arXiv

  12. [20]

    MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion,

    J. Zhang et al., “MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion,” Oct. 04, 2024, arXiv: arXiv:2410.03825

    2024 arXiv

  13. [21]

    PreF3R: Pose-Free Feed-Forward 3D Gaussian Splatting from Variable-length Image Sequence,

    Z. Chen, J. Yang, and H. Yang, “PreF3R: Pose-Free Feed-Forward 3D Gaussian Splatting from Variable-length Image Sequence,” Nov. 25, 2024, arXiv: arXiv:2411.16877

    2024 arXiv

  14. [22]

    Regist3R: Incremental Registration with Stereo Foundation Model,

    S. Liu, W. Li, P. Qiao, and Y. Dou, “Regist3R: Incremental Registration with Stereo Foundation Model,” Apr. 16, 2025, arXiv: arXiv:2504.12356

    2025

  15. [23]

    MV-DUSt3R+: Single-Stage Scene Reconstruction from Sparse Views In 2 Seconds,

    Z. Tang et al., “MV-DUSt3R+: Single-Stage Scene Reconstruction from Sparse Views In 2 Seconds,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 5283–5293

    2025

  16. [24]

    Easi3R: Estimating Disentangled Motion from DUSt3R Without Training,

    X. Chen, Y. Chen, Y. Xiu, A. Geiger, and A. Chen, “Easi3R: Estimating Disentangled Motion from DUSt3R Without Training,” Mar. 31, 2025, arXiv: arXiv:2503.24391

    2025

  17. [25]

    Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass,

    J. Yang et al., “Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21924–21935

    2025

  18. [26]

    Continuous 3D Perception Model with Persistent State,

    Q. Wang, Y. Zhang, A. Holynski, A. A. Efros, and A. Kanazawa, “Continuous 3D Perception Model with Persistent State,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 10510–10522

    2025

  19. [27]

    MegaDepth: Learning Single-View Depth Prediction from Internet Photos,

    Z. Li and N. Snavely, “MegaDepth: Learning Single-View Depth Prediction from Internet Photos,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA: IEEE, Jun. 2018, pp. 2041–2050. doi: 10.1109/CVPR.2018.00218

    2018

  20. [28]

    AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis,

    K. Vuong, A. Ghosh, D. Ramanan, S. Narasimhan, and S. Tulsiani, “AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesis,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21674–21684

    2025

  21. [29]

    Habitat: A Platform for Embodied AI Research,

    M. Savva et al., “Habitat: A Platform for Embodied AI Research,” presented at the Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 9339–9347

    2019

  22. [30]

    ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data,

    G. Baruch et al., “ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data,” presented at the Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), Jun. 2021

    2021

  23. [31]

    A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation,

    N. Mayer et al., “A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA: IEEE, Jun. 2016, pp. 4040–4048

    2016

  24. [32]

    BlendedMVS: A Large-Scale Dataset for Generalized Multi-View Stereo Networks,

    Y. Yao et al., “BlendedMVS: A Large-Scale Dataset for Generalized Multi-View Stereo Networks,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 1790–1799

    2020

  25. [33]

    ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes,

    C. Yeshwanth, Y.-C. Liu, M. Nießner, and A. Dai, “ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes,” presented at the Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12–22

    2023

  26. [34]

    Common Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category Reconstruction,

    J. Reizenstein, R. Shapovalov, P. Henzler, L. Sbordone, P. Labatut, and D. Novotny, “Common Objects in 3D: Large-Scale Learning and Evaluation of Real-Life 3D Category Reconstruction,” presented at the Proceedings of the IEEE/CVF International Conference on Computer Vision, 20...

    2021

  27. [35]

    Scalability in Perception for Autonomous Driving: Waymo Open Dataset,

    P. Sun et al., “Scalability in Perception for Autonomous Driving: Waymo Open Dataset,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2446–2454

    2020

  28. [36]

    DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision,

    L. Ling et al., “DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22160–22169

    2024

  29. [37]

    RGBD Objects in the Wild: Scaling Real-World 3D Object Learning from RGB-D Videos,

    H. Xia, Y. Fu, S. Liu, and X. Wang, “RGBD Objects in the Wild: Scaling Real-World 3D Object Learning from RGB-D Videos,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 22378–22389

    2024

  30. [38]

    Hypersim: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding,

    M. Roberts et al., “Hypersim: A Photorealistic Synthetic Dataset for Holistic Indoor Scene Understanding,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Montreal, QC, Canada: IEEE, Oct. 2021, pp. 10892–10902

    2021

  31. [39]

    The Replica Dataset: A Digital Replica of Indoor Spaces,

    J. Straub et al., “The Replica Dataset: A Digital Replica of Indoor Spaces,” Jun. 13, 2019, arXiv: arXiv:1906.05797. doi: 10.48550/arXiv.1906.05797

    2019 doi arXiv

  32. [40]

    DeepMVS: Learning Multi-View Stereopsis,

    P.-H. Huang, K. Matzen, J. Kopf, N. Ahuja, and J.-B. Huang, “DeepMVS: Learning Multi-View Stereopsis,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 2821–2830

    2018

  33. [41]

    Virtual KITTI 2,

    Y. Cabon, N. Murray, and M. Humenberger, “Virtual KITTI 2,” Jan. 29, 2020, arXiv: arXiv:2001.10773

    2020 arXiv

  34. [42]

    Aria Digital Twin: A New Benchmark Dataset for Egocentric 3D Machine Perception,

    X. Pan et al., “Aria Digital Twin: A New Benchmark Dataset for Egocentric 3D Machine Perception,” presented at the Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 20133–20143

    2023

  35. [43]

    Large Scale Multi-view Stereopsis Evaluation,

    R. Jensen, A. Dahl, G. Vogiatzis, E. Tola, and H. Aanaes, “Large Scale Multi-view Stereopsis Evaluation,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 406–413

    2014

  36. [44]

    Tanks and temples: benchmarking large-scale scene reconstruction,

    A. Knapitsch, J. Park, Q.-Y. Zhou, and V. Koltun, “Tanks and temples: benchmarking large-scale scene reconstruction,” ACM Trans. Graph., vol. 36, no. 4, pp. 1–13, Aug. 2017

    2017

  37. [45]

    A Multi-view Stereo Benchmark with High-Resolution Images and Multi-camera Videos,

    T. Schops et al., “A Multi-view Stereo Benchmark with High-Resolution Images and Multi-camera Videos,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI: IEEE, Jul. 2017, pp. 2538–2547

    2017

  38. [46]

    MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-view Images,

    Y. Chen et al., “MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-view Images,” in Computer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds., Cham: Springer Nature Switzerland, 2025, pp. 370–386

    2024

  39. [47]

    PE3R: Perception-Efficient 3D Reconstruction,

    J. Hu, S. Wang, and X. Wang, “PE3R: Perception-Efficient 3D Reconstruction,” Mar. 10, 2025, arXiv: arXiv:2503.07507

    2025

  40. [48]

    ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Niessner, “ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,” presented at the Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 5828–5839

    2017

  41. [49]

    Massively Parallel Multiview Stereopsis by Surface Normal Diffusion,

    S. Galliani, K. Lasinger, and K. Schindler, “Massively Parallel Multiview Stereopsis by Surface Normal Diffusion,” presented at the Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 873–881

    2015

  42. [50]

    Pixelwise View Selection for Unstructured Multi-View Stereo,

    J. L. Schönberger, E. Zheng, J.-M. Frahm, and M. Pollefeys, “Pixelwise View Selection for Unstructured Multi-View Stereo,” in Computer Vision – ECCV 2016, B. Leibe, J. Matas, N. Sebe, and M. Welling, Eds., Cham: Springer International Publishing, 2016, pp. 501–518

    2016

  43. [51]

    Learning Inverse Depth Regression for Multi-View Stereo with Correlation Cost Volume,

    Q. Xu and W. Tao, “Learning Inverse Depth Regression for Multi-View Stereo with Correlation Cost Volume,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 07, Art. no. 07, Apr. 2020

    2020

  44. [52]

    GeoMVSNet: Learning Multi-View Stereo With Geometry Perception,

    Z. Zhang, R. Peng, Y. Hu, and R. Wang, “GeoMVSNet: Learning Multi-View Stereo With Geometry Perception,” presented at the Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 21508–21518

    2023

  45. [53]

    Test3R: Learning to Reconstruct 3D at Test Time,

    Y. Yuan, Q. Shen, S. Wang, X. Yang, and X. Wang, “Test3R: Learning to Reconstruct 3D at Test Time,” Jun. 16, 2025, arXiv: arXiv:2506.13750

    2025 arXiv

  46. [54]

    MUSt3R: Multi-view Network for Stereo 3D Reconstruction,

    Y. Cabon et al., “MUSt3R: Multi-view Network for Stereo 3D Reconstruction,” presented at the Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 1050–1060

    2025

Pith tools

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