Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Test3R: Learning to Reconstruct 3D at Test Time

T0 review · 2 major / 4 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 Test3R claims that test-time visual prompt tuning, driven by a cross-pair consistency objective, makes pretrained pairwise dense 3D reconstruction models more accurate and more consistent across views, and that the same recipe transfers…

desk verdict Test3R's test-time prompt-tuning trick is a real, useful plug-in for DUSt3R-style 3D reconstruction, but the scale-comparability question in the consistency loss leaves the gains one ablation short of fully explained. read the letter →

arxiv 2506.13750 v1 pith:OV4QGJFZ submitted 2025-06-16 cs.CV

Yuheng Yuan , Qiuhong Shen , Shizun Wang , Xingyi Yang , Xinchao Wang This is my paper
classification cs.CV
keywords test-timetrainingvisualprompttuning3Dreconstructionmulti-viewdepthestimationdensematchingpointmapconsistencyDUSt3R
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

Test3R claims that a pretrained pairwise 3D reconstruction model such as DUSt3R can be made more accurate simply by optimizing a small set of visual prompt tokens at test time. From a triplet of images $(I_{ref}, I_{src1}, I_{src2})$, it builds two pointmaps of the reference view, one from each pair, and tunes the prompts so both pointmaps agree. The paper reports that this unlabeled, self-supervised step improves DUSt3R on 3D reconstruction benchmarks and on multi-view depth estimation, reducing the DTU relative depth error from 3.3 to 2.0 and raising the inlier ratio from 69.9 to 84.1 without camera poses or intrinsics. The same recipe also improves MAST3R and MonST3R, which suggests the mechanism transfers to other pairwise dense matchers.

What carries the argument

The mechanism is visual prompt tuning applied to a frozen DUSt3R-style network. A set of learnable prompt tokens is inserted into every encoder layer of the vision transformer; only these tokens are updated at test time (about 0.79M parameters), while all other parameters stay fixed. The training signal is the L2 distance between the two reference-view pointmaps. What makes this work, according to the paper, is that a single scene's triplets provide a strong self-supervised consistency signal that pushes pairwise predictions to agree, effectively replacing the need for a separate global optimization.

What would settle it

Replace the consistency loss with a scale-invariant variant that first estimates the per-pair scale factor, then measure DTU accuracy; if the gains largely vanish, the original loss was exploiting scale freedom rather than true geometric consistency.

Watch

Extended reading notes

Core claim

The central claim is that cross-pair inconsistency is a major source of error in pairwise dense matching, and it can be corrected at test time by enforcing geometric consistency across pairs that share a reference view. The objective is an L2 distance between two reference-view pointmaps, $X_1^{ref,ref}$ and $X_2^{ref,ref}$, which describe the same reference image when paired with two different source views. The authors argue that this single self-supervised objective absorbs the benefits of a global alignment stage: it removes inter-pair inconsistency, corrects imprecise short-baseline geometry by pulling it toward a more global prediction, and adapts the model to the test scene. On DTU, ETH3D, 7Scenes, and NRGBD, the adapted model matches or beats methods that use ground-truth poses and intrinsics or are trained on the same domain.

Load-bearing premise

The load-bearing premise is that the two pointmaps compared in the training objective are already in the same scale; DUSt3R predicts pointmaps with per-pair scale factors, and the paper never justifies comparing them directly with an L2 loss.

Editorial extensions

If this is right

  • On DTU, Test3R lowers absolute relative depth error from 3.3 to 2.0 and raises the 3% inlier ratio from 69.9 to 84.1, with no camera poses, intrinsics, or range data.
  • On 7Scenes and NRGBD, Test3R improves over vanilla DUSt3R on mean accuracy and completion and is competitive with CUT3R, a model trained with a persistent state.
  • Applying the same test-time prompt tuning to MAST3R and MonST3R improves their reconstruction accuracy on 7Scenes.
  • The added cost is small: about 0.79M trainable parameters and roughly 30 seconds of test-time training per scene in the reported setting.
  • The technique is generic to models that output pairwise pointmaps in a common reference frame, so new DUSt3R-style models can be adapted with the same recipe.

Reading between the lines

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

  • If the pointmaps are not scale-comparable, the L2 loss may be cutting the wrong objective; a version that first estimates the per-pair scale factor would test whether the reported gains are geometric or partly a convenient scale choice.
  • The objective ignores source-view pointmaps, so a natural extension is to add a symmetric consistency term over the source pointmaps as well, which could further stabilize the reconstruction on wide-baseline views.
  • The same triplet-consistency idea could be applied to other dense geometric outputs, such as depth or normal maps, or to dynamic scenes where per-frame consistency is the natural self-supervision.
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 / 4 minor

Summary. The paper proposes Test3R, a test-time adaptation method for pairwise 3D reconstruction models of the DUSt3R family. For each triplet (I1, I2, I3), the method computes two pointmaps of the reference view I1, one from pair (I1, I2) and one from pair (I1, I3), and optimizes a self-supervised loss (Eq. 7) that penalizes the L2 distance between these two reference-view pointmaps. To keep the adaptation cheap, only learnable visual prompt tokens inserted in the encoder layers are optimized, with the backbone frozen. The authors evaluate Test3R on 7Scenes and NRGBD for 3D reconstruction and on DTU and ETH3D for multi-view depth, reporting consistent improvements over vanilla DUSt3R and competitive or better results than several recent baselines, and also show gains when applying the same prompt-tuning recipe to MAST3R and MonST3R. The main claim is that maximizing cross-pair consistency at test time improves geometric accuracy and generalization to unseen scenes.

Significance. If the claimed causal mechanism is correct, the paper makes a useful and simple contribution: a nearly parameter-free way to specialize a pretrained pairwise reconstruction model to a test scene using only self-supervision, with no camera poses or depth labels, and with negligible parameter overhead. The experiments span multiple datasets and tasks, ablate prompt length and depth, and demonstrate applicability to other models. However, the central objective in Eq. (7) has a scale-comparability problem that is not addressed, so the empirical gains are not yet convincingly attributable to geometric refinement rather than scale alignment. The conclusions are also somewhat stronger than the tables warrant given the lack of error bars, the per-dataset tuning of learning rates, and the marginal ETH3D improvement.

major comments (2)
  1. [Sec. 4.2, Eq. (7); Sec. 3, Eq. (6)] The loss in Eq. (7) directly measures the L2 distance between X1^{ref,ref} and X2^{ref,ref}, but DUSt3R predicts pairwise pointmaps only up to a per-pair scale, as shown by the per-edge scale factors sigma_e in the global alignment objective of Eq. (6). The paper never establishes that the two reference pointmaps are scale-comparable. If the two pairs have different intrinsic scale factors, the objective can be reduced by shrinking or enlarging one pointmap toward the other without improving metric accuracy, and in the limit it can be minimized by collapsing both predictions onto a common arbitrary surface. This makes the causal reading of the reported gains ('better geometry') insecure. Please either justify the scale-comparability premise, modify the objective to be scale-invariant (for example, by estimating and normalizing a per-pair scale before comparison), and add an ablation that separates scale alignment from genuine geometric refinement, or report the results of a scale-normalized version of Eq. (7).
  2. [Sec. 5.1 and 5.2, Tables 1-2; Appendix A] The experimental claims lack error bars and significance tests, and the method has per-dataset hyperparameters that are hand-tuned. Appendix A reports different learning rates for each dataset (0.00001, 0.00008, 0.00004, and 0.00001) without any sensitivity analysis. The ETH3D multi-view depth improvement is marginal (rel 3.3 vs. 3.2, tau 73.0 vs. 74.0), so without run-to-run variance it is difficult to assess whether the gain is real. Please report standard deviations over multiple runs (or across seeds) and a sensitivity study over the learning rate and prompt length.
minor comments (4)
  1. [Table 2] The row for the DUSt3R baseline is labeled 'DUSt3R [1]', but reference [1] is D2-Net; the citation should be [12].
  2. [Throughout] There are several typos and formatting errors: 'recomstruction' in Section 1, 'Prompsts' in Table 5, 'V ariants' in Table 4, 'training time training phase' in Section 4.2, and 'specific images' in the Figure 1 caption.
  3. [Appendix C, Eq. (9)] Equation (9) writes K^{-1}[iD_{i,j}, jD_{i,j}, D_{i,j}] without clarifying the column-vector transposition convention; please make the notation precise.
  4. [Conclusion and abstract] The abstract and conclusion claim state-of-the-art performance, but Table 1 shows MAST3R and CUT3R competitive on several metrics and Table 2 shows a very small ETH3D improvement; the claims should be tempered to match the reported numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the test-time consistency objective is evaluated against external ground-truth metrics, and the only self-citations are non-load-bearing related-work entries.

full rationale

Test3R's central claim is that optimizing visual prompts with the self-supervised objective of Eq. (7) improves geometric reconstruction. This claim is not circular: the objective compares two DUSt3R pointmaps produced from unlabeled test images, and the reported improvements are measured with external ground-truth benchmarks (Acc/Comp/NC on 7Scenes and NRGBD; rel and τ on DTU and ETH3D) that are never used during prompt optimization. No fitted parameter is relabeled as a prediction, and no uniqueness theorem or prior same-author result is invoked to force the method. The self-citations ([36] GFlow and [37] Pe3R) appear only in the related-work enumeration of dynamic reconstruction and 3D perception and do not justify any engineering or theoretical step. The appendices also provide honest limitations (Appendix D), including dependency on input images and the open question of triplet sampling for many views. A technical caveat, not a circularity, is that Eq. (7) directly subtracts X1^{ref,ref} and X2^{ref,ref} without explicitly normalizing the per-pair scale that DUSt3R's global alignment absorbs through σ_e in Eq. (6); this scale ambiguity is a correctness risk and should be checked, but it does not make the derivation equivalent to its inputs, because success is still adjudicated by external metrics.

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

The method's empirical claims rest on three unproven modeling choices: raw pointmap comparability in Eq. (7), consistency as a proxy for accuracy, and single-epoch prompt tuning preserving pretrained knowledge. Four hyperparameters are hand-set or ablated per benchmark. No invented entities are introduced.

free parameters (4)
  • test-time learning rate (per dataset) = 1e-5 (7Scenes), 8e-5 (NRGBD), 4e-5 (DTU), 1e-5 (ETH3D)
    Hand-selected separately for each benchmark; directly influences convergence and final metrics, with no validation protocol reported.
  • prompt length N_p = 32
    Chosen based on ablation in Table 4; affects capacity and final accuracy.
  • triplet sampling cap = 165
    Randomly sampled when total triplets exceed 165; choice trades compute vs coverage and affects optimization.
  • test-time training epochs = 1
    Fixed at 1 epoch; likely tuned for compute/performance balance.
assumptions (4)
  • ad hoc to paper Pointmaps from different pairs (I1,I2) and (I1,I3) share the same coordinate frame and scale, so Eq. (7) can compare them directly
    DUSt3R pairs carry per-pair scale ambiguity (Eq. 6 introduces sigma_e), yet the loss does not align scale; this is an unstated modeling assumption.
  • domain assumption Minimizing cross-pair consistency is a valid proxy for geometric accuracy; consistency does not collapse to a trivial constant reconstruction
    The self-supervised objective contains no ground truth; the paper assumes the consistent solution is the correct one, supported only by empirical evaluation.
  • domain assumption The frozen pretrained backbone retains reconstruction knowledge while 1 epoch of prompt tuning adapts to the scene without overfitting
    Foundation of the method; no theoretical guarantee, relies on VPT empirical success.
  • domain assumption Evaluation datasets are used only with unlabeled test images and the learned prompts do not incorporate benchmark labels
    Test-time training uses test images, but the method's self-supervised loss does not touch ground-truth geometry; still, per-dataset hyperparameters were selected in the same benchmark loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Test3R: Learning to Reconstruct 3D at Test Time." pith.science (2026). https://pith.science/paper/OV4QGJFZ

@misc{pith2026250613750,
  author       = {Pith},
  title        = {Pith review of: Test3R: Learning to Reconstruct 3D at Test Time},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OV4QGJFZ}},
  note         = {Machine review of arXiv:2506.13750}
}
abstract

Dense matching methods like DUSt3R regress pairwise pointmaps for 3D reconstruction. However, the reliance on pairwise prediction and the limited generalization capability inherently restrict the global geometric consistency. In this work, we introduce Test3R, a surprisingly simple test-time learning technique that significantly boosts geometric accuracy. Using image triplets ($I_1,I_2,I_3$), Test3R generates reconstructions from pairs ($I_1,I_2$) and ($I_1,I_3$). The core idea is to optimize the network at test time via a self-supervised objective: maximizing the geometric consistency between these two reconstructions relative to the common image $I_1$. This ensures the model produces cross-pair consistent outputs, regardless of the inputs. Extensive experiments demonstrate that our technique significantly outperforms previous state-of-the-art methods on the 3D reconstruction and multi-view depth estimation tasks. Moreover, it is universally applicable and nearly cost-free, making it easily applied to other models and implemented with minimal test-time training overhead and parameter footprint. Code is available at https://github.com/nopQAQ/Test3R.

Figures

Figures reproduced from arXiv: 2506.13750 by the authors.

Figure 1
Figure 1. Given a set of images of a specific images, our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Inconsistency Study. On the left are two image pairs sharing the same reference view I1 but with different source views I2 and I3. On the right are the corresponding point maps, with each color indicating the respective image pair. Despite its huge success, this pair-wise prediction paradigm is inherently problematic. Under such a design, the model considers only two images at a time. Such a constraint leads to seve… view at source ↗
Figure 3
Figure 3. Overview of Test3R. The primary goal of Test3R is to adapt a pretrained reconstruction model fs to the specific distribution of test scenes ft. It achieves this goal by optimizing a set of visual prompts at test time through a self-supervised training objective that maximizes cross-pair consistency between X ref,ref 1 and X ref,ref 2 . Therefore, we show a qualitative analysis of the pointmaps on the DTU [58] and ET… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative Comparison on 3D Reconstruction. [_, Ei ] = Li([Pi−1, Ei−1]) (8) where Pi−1 and Ei−1 are learnable prompt tokens and image patch tokens at i − 1-th Transformer layer. Test-time training. We only fine-tune the parameters of the prompts, while all other param…
Figure 5
Figure 5. Figure 5: Qualitative Comparison on Multi-view Depth. Qualitative Results. The qualitative result is shown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Comparison on cross-pair consistency. The depth map of the same reference view but paired with different source views. Test3R demonstrates superior cross-pair consistency compared to vanilla DUSt3R. C.2 Compared to single forward-based model We compare DUSt3R and Test3…
Figure 7
Figure 7. Figure 7: Qualitative comparisons of DUSt3R and our method. [12] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 206…

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

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A survey of feed-forward 3D reconstruction models that jointly estimate camera poses and dense geometry from images in one network pass.

Reference graph

Works this paper leans on

68 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    D2-net: A trainable cnn for joint description and detection of local features

    Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition, pages 8092–8101, 2019

    work page 2019

  2. [2]

    Superpoint: Self-supervised interest point detection and description

    Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224–236, 2018

    work page 2018

  3. [3]

    Distinctive image features from scale-invariant keypoints

    David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91–110, 2004

    2004

  4. [4]

    Neural-guided ransac: Learning where to sample model hypotheses

    Eric Brachmann and Carsten Rother. Neural-guided ransac: Learning where to sample model hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4322–4331, 2019

    work page 2019

  5. [5]

    Lightglue: Local feature matching at light speed

    Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023

    work page 2023

  6. [6]

    Progressive correspondence pruning by consensus learning

    Chen Zhao, Yixiao Ge, Feng Zhu, Rui Zhao, Hongsheng Li, and Mathieu Salzmann. Progressive correspondence pruning by consensus learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6464–6473, 2021

    work page 2021

  7. [7]

    Sfm with mrfs: Discrete- continuous optimization for large-scale structure from motion

    David J Crandall, Andrew Owens, Noah Snavely, and Daniel P Huttenlocher. Sfm with mrfs: Discrete- continuous optimization for large-scale structure from motion. IEEE transactions on pattern analysis and machine intelligence, 35(12):2841–2853, 2012

    work page 2012

  8. [8]

    Pixel-perfect structure- from-motion with featuremetric refinement

    Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure- from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987–5997, 2021

    work page 2021

Show all 68 references
  1. [9]

    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

    2016

  2. [10]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2495–2504, 2020

    2020

  3. [11]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 5...

    2016

  4. [12]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697–20709, 2024

    2024

  5. [13]

    Grounding image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71–91. Springer, 2024

    2024

  6. [14]

    A multiple-baseline stereo

    Masatoshi Okutomi and Takeo Kanade. A multiple-baseline stereo. IEEE Transactions on pattern analysis and machine intelligence, 15(4):353–363, 1993

    1993

  7. [15]

    Visual prompt tuning

    Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser- Nam Lim. Visual prompt tuning. In European conference on computer vision, pages 709–727. Springer, 2022. 13

    2022

  8. [16]

    Monst3r: A simple approach for estimating geometry in the presence of motion

    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

  9. [17]

    The interpretation of structure from motion

    Shimon Ullman. The interpretation of structure from motion. Proceedings of the Royal Society of London. Series B. Biological Sciences, 203(1153):405–426, 1979

    1979

  10. [18]

    Multi-view stereo: A tutorial

    Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and trends® in Computer Graphics and Vision, 9(1-2):1–148, 2015

    2015

  11. [19]

    Massively parallel multiview stereopsis by surface normal diffusion

    Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE international conference on computer vision, pages 873–881, 2015

    2015

  12. [20]

    Adaptive patch deformation for textureless-resilient multi-view stereo

    Yuesong Wang, Zhaojie Zeng, Tao Guan, Wei Yang, Zhuo Chen, Wenkai Liu, Luoyuan Xu, and Yawei Luo. Adaptive patch deformation for textureless-resilient multi-view stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1621–1630, 2023

    2023

  13. [21]

    Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction

    Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403–3416, 2022

    2022

  14. [22]

    Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision

    Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3504–3515, 2020

    2020

  15. [23]

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

    Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5610–5619, 2021

    2021

  16. [24]

    Multiview neural surface reconstruction by disentangling geometry and appearance

    Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492–2502, 2020

    2020

  17. [25]

    Multiview stereo with cascaded epipolar raft

    Zeyu Ma, Zachary Teed, and Jia Deng. Multiview stereo with cascaded epipolar raft. In European Conference on Computer Vision, pages 734–750. Springer, 2022

    2022

  18. [26]

    Rethinking depth estimation for multi-view stereo: A unified representation

    Rui Peng, Rongjie Wang, Zhenyu Wang, Yawen Lai, and Ronggang Wang. Rethinking 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

    2022

  19. [27]

    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 (ECCV), pages 767–783, 2018

    2018

  20. [28]

    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

    2023

  21. [29]

    Vggt: Visual geometry grounded transformer

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

    2025 arXiv

  22. [30]

    Slam3r: Real-time dense scene reconstruction from monocular rgb videos

    Yuzheng Liu, Siyan Dong, Shuzhe Wang, Yingda Yin, Yanchao Yang, Qingnan Fan, and Baoquan Chen. Slam3r: Real-time dense scene reconstruction from monocular rgb videos. arXiv preprint arXiv:2412.09401, 2024

    2024 arXiv

  23. [31]

    3d reconstruction with spatial memory

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

    2024 arXiv

  24. [32]

    Pow3r: Em- powering unconstrained 3d reconstruction with camera and scene priors

    Wonbong Jang, Philippe Weinzaepfel, Vincent Leroy, Lourdes Agapito, and Jerome Revaud. Pow3r: Em- powering unconstrained 3d reconstruction with camera and scene priors. arXiv preprint arXiv:2503.17316, 2025

    2025 arXiv

  25. [33]

    Align3r: Aligned monocular depth estimation for dynamic videos

    Jiahao Lu, Tianyu Huang, Peng Li, Zhiyang Dou, Cheng Lin, Zhiming Cui, Zhen Dong, Sai-Kit Yeung, Wenping Wang, and Yuan Liu. Align3r: Aligned monocular depth estimation for dynamic videos. arXiv preprint arXiv:2412.03079, 2024

    2024 arXiv

  26. [34]

    Easi3r: Estimating disentangled motion from dust3r without training

    Xingyu Chen, Yue Chen, Yuliang Xiu, Andreas Geiger, and Anpei Chen. Easi3r: Estimating disentangled motion from dust3r without training. arXiv preprint arXiv:2503.24391, 2025. 14

    2025

  27. [35]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025

    2025 arXiv

  28. [36]

    Gflow: Recovering 4d world from monocular video

    Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 7862–7870, 2025

    2025

  29. [37]

    Pe3r: Perception-efficient 3d reconstruction

    Jie Hu, Shizun Wang, and Xinchao Wang. Pe3r: Perception-efficient 3d reconstruction. arXiv preprint arXiv:2503.07507, 2025

    2025

  30. [38]

    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. arXiv preprint arXiv:2501.13928, 2025

    2025 arXiv

  31. [39]

    Learning by transduction

    Alex Gammerman, V olodya V ovk, and Vladimir Vapnik. Learning by transduction. arXiv preprint arXiv:1301.7375, 2013

    2013 arXiv

  32. [40]

    Estimation of dependences based on empirical data

    Vladimir Vapnik. Estimation of dependences based on empirical data . Springer Science & Business Media, 2006

    2006

  33. [41]

    Large scale transductive svms

    Ronan Collobert, Fabian Sinz, Jason Weston, Léon Bottou, and Thorsten Joachims. Large scale transductive svms. Journal of Machine Learning Research, 7(8), 2006

    2006

  34. [42]

    Learning to classify text using support vector machines, volume 668

    Thorsten Joachims. Learning to classify text using support vector machines, volume 668. Springer Science & Business Media, 2002

    2002

  35. [43]

    Test-time training on nearest neighbors for large language models

    Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. arXiv preprint arXiv:2305.18466, 2023

    2023 arXiv

  36. [44]

    Local learning algorithms

    Léon Bottou and Vladimir Vapnik. Local learning algorithms. Neural computation, 4(6):888–900, 1992

    1992

  37. [45]

    Svm-knn: Discriminative nearest neighbor classification for visual category recognition

    Hao Zhang, Alexander C Berg, Michael Maire, and Jitendra Malik. Svm-knn: Discriminative nearest neighbor classification for visual category recognition. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2, pages 2126–2136. IEEE, 2006

    2006

  38. [46]

    Test-time training with self-supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–9248. PMLR, 2020

    2020

  39. [47]

    Self-supervised policy adaptation during deployment

    Nicklas Hansen, Rishabh Jangir, Yu Sun, Guillem Alenyà, Pieter Abbeel, Alexei A Efros, Lerrel Pinto, and Xiaolong Wang. Self-supervised policy adaptation during deployment. arXiv preprint arXiv:2007.04309, 2020

    2007 arXiv

  40. [48]

    Online learning of unknown dynamics for model-based controllers in legged locomotion

    Yu Sun, Wyatt L Ubellacker, Wen-Loong Ma, Xiang Zhang, Changhao Wang, Noel V Csomay-Shanklin, Masayoshi Tomizuka, Koushil Sreenath, and Aaron D Ames. Online learning of unknown dynamics for model-based controllers in legged locomotion. IEEE Robotics and Automation Letters, 6(4...

    2021

  41. [49]

    Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808–21820, 2021

    2021

  42. [50]

    Robust test-time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15922– 15932, 2023

    2023

  43. [51]

    Language models are few-shot learners

    Ben Mann, N Ryder, M Subbiah, J Kaplan, P Dhariwal, A Neelakantan, P Shyam, G Sastry, A Askell, S Agarwal, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1:3, 2020

    2005 arXiv

  44. [52]

    How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020

    Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423–438, 2020

    2020

  45. [53]

    Autoprompt: Eliciting knowledge from language models with automatically generated prompts

    Taylor Shin, Yasaman Razeghi, Robert L Logan IV , Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020

    2010 arXiv

  46. [54]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 15

    2021 arXiv

  47. [55]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021

    2021 arXiv

  48. [56]

    P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021

    2021 arXiv

  49. [57]

    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, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  50. [58]

    Large-scale data for multiple-view stereopsis

    Henrik Aanæs, Rasmus Ramsbøl Jensen, George V ogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153–168, 2016

    2016

  51. [59]

    A multi-view stereo benchmark with high-resolution images and multi- camera videos

    Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi- camera videos. In Proceedings of the IEEE conference on computer vision and patt...

    2017

  52. [60]

    Visual prompt tuning for test-time domain adaptation

    Yunhe Gao, Xingjian Shi, Yi Zhu, Hao Wang, Zhiqiang Tang, Xiong Zhou, Mu Li, and Dimitris N Metaxas. Visual prompt tuning for test-time domain adaptation. arXiv preprint arXiv:2210.04831, 2022

    2022 arXiv

  53. [61]

    Demon: Depth and motion network for learning monocular stereo

    Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig, Nikolaus Mayer, Eddy Ilg, Alexey Dosovitskiy, and Thomas Brox. Demon: Depth and motion network for learning monocular stereo. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5038–5047, 2017

    2017

  54. [62]

    Deepv2d: Video to depth with differentiable structure from motion

    Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018

    2018 arXiv

  55. [63]

    A benchmark and a baseline for robust multi-view depth estimation

    Philipp Schröppel, Jan Bechtold, Artemij Amiranashvili, and Thomas Brox. A benchmark and a baseline for robust multi-view depth estimation. In 2022 International Conference on 3D Vision (3DV) , pages 637–645. IEEE, 2022

    2022

  56. [64]

    Vis-mvsnet: Visibility-aware multi-view stereo network

    Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1):199–214, 2023

    2023

  57. [65]

    Mvs2d: Efficient multi-view stereo via attention- driven 2d convolutions

    Zhenpei Yang, Zhile Ren, Qi Shan, and Qixing Huang. Mvs2d: Efficient multi-view stereo via attention- driven 2d convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8574–8584, 2022

    2022

  58. [66]

    Scene coordinate regression forests for camera relocalization in rgb-d images

    Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2937, 2013

    2013

  59. [67]

    Neural rgb-d surface reconstruction

    Dejan Azinovi´c, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6290–6301, 2022

    2022

  60. [68]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 16

    2014 arXiv

Pith tools

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