Pith. sign in

REVIEW 4 major objections 6 minor 48 references

Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction

T0 review · 4 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 A data augmentation strategy called Puzzles synthesizes unbounded posed video-depth clips from a single image; models trained on only 10% of the original data with it match or beat full-dataset baselines on 7Scenes and DTU.

desk verdict Useful data-centric augmentation, but the headline 10% claim is not yet controlled and needs a matched-budget revision. read the letter →

arxiv 2506.23863 v1 pith:THGGLIFA submitted 2025-06-30 cs.CV

Jiahao Ma , Lei Wang , Miaomiao liu , David Ahmedt-Aristizabal , Chuong Nguyen This is my paper
classification cs.CV
keywords dataaugmentation3Dreconstructionvideo-depthsynthesispointmapregressionkeyframeselectioncameraposeDUSt3Refficiency
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 paper argues that synthetic training data is the bottleneck for end-to-end 3D reconstruction, not network architecture. It introduces Puzzles, a data augmentation framework that converts a single RGB-D image into an unbounded stream of posed video-depth clips by cutting the image into ordered, overlapping patches, assigning each patch a simulated camera trajectory, and rendering the resulting views; a companion step, Clips-to-Clips, extends the idea to video by selecting keyframes that maximize geometric coverage. Integrated into three existing video-based reconstruction pipelines (Spann3R, SLAM3R, Fast3R), Puzzles consistently improves reconstruction accuracy without changing the networks. The strongest claim is that models trained on only ten percent of the original data, augmented with Puzzles, match or exceed the accuracy of models trained on the full dataset. If true, this makes large-scale 3D training feasible from far fewer collected videos and grounds the hope of training from any single image.

What carries the argument

The central object is an image-to-video augmentation engine that turns one RGB-D frame into a posed video-depth clip. Its defining mechanism is the ordered, overlapping patch sequence: each patch shares area with at least one earlier patch, encoding the temporal overlap that makes real video geometrically consistent, and each patch is re-rendered under a simulated camera trajectory. The trajectory is generated by translating the patch (via intrinsics scaling or PnP-estimated extrinsics) and then applying a random rotation about the centroid of the patch's 3D points, constructed with Rodrigues' formula, with invalid views filtered by front-surface and view-frustum coverage tests. Clips-to-Clips generalizes the engine to whole videos via an overlap matrix $O_{ij}$ that measures the fraction of reprojected geometry shared by frames $i$ and $j$, feeding a three-step keyframe selection (validity filtering, longest cover set, redundancy pruning).

What would settle it

Train Spann3R on one-tenth of the blended corpus with Puzzles, but replace the centroid-based camera rotation with plain translation-only patch cropping while keeping every other setting identical; if 7Scenes accuracy no longer matches the full-data baseline, the rotation mechanism is the essential ingredient of the claimed 10% result.

Watch

Extended reading notes

Core claim

At the center of the paper is a new data-centric training recipe for the 3R family of reconstruction models. Puzzles generates synthetic video-depth sequences from a single RGB-D frame: it samples an ordered, overlapping set of patches, calibrates a camera for each patch by rescaling intrinsics or by solving a PnP problem with RANSAC, applies a random rotation about the patch's 3D point-cloud centroid using Rodrigues' formula, and keeps only views that pass front-surface and view-frustum coverage tests, rendering the surviving point clouds with Open3D. On top of this, Clips-to-Clips computes a pairwise geometric overlap matrix for an input video and selects a small set of keyframes that maximally cover the scene with least redundancy, then augments those keyframes. Across 7Scenes, NRGBD, and DTU, adding Puzzles to Spann3R, SLAM3R, and Fast3R improves both accuracy and completion error, and the one-tenth-of-data training regime matches or beats full-data training on 7Scenes and DTU.

Load-bearing premise

The load-bearing premise is that synthetic clips created by cropping a single RGB-D image and re-rendering it under random centroid-based rotations, with occluded areas filled white, are geometrically and photometrically faithful enough to real video that training on them transfers to unseen real scenes.

Editorial extensions

If this is right

  • On 7Scenes, models trained with Puzzles on just 10% of the training data match or exceed the accuracy of models trained on the full dataset; on NRGBD the 10% models do not match full-data, though they still improve over their own 10% baseline.
  • Combining translation and rotation augmentation lowers mean reconstruction error to 0.0330, about 35% better than translation-only augmentation (0.0508), so rotational viewpoint diversity is a major contributor to the gains.
  • Clips-to-Clips gives substantially more stable reconstructions than Image-to-Clips on long scene-level sequences (75+ views), reducing drift at test time.
  • Puzzles works almost as well with predicted monocular depth from an off-the-shelf estimator as with ground-truth depth, meaning it can be applied to large unlabeled image collections.

Reading between the lines

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

  • If the 10%-of-data result holds at larger scale, the practical implication is that the 3R model family's data demands can be decoupled from expensive posed-video capture: any large archive of still RGB-D frames becomes a training source, which may shift the field's bottleneck from data collection to augmentation quality.
  • The white-filled occlusions are a likely weak spot: scenes with sharp depth discontinuities will produce many invalid rotations that are either discarded or filled with white, so the method's effective coverage on such scenes is an empirical question worth measuring directly.
  • The overlap-matrix keyframe selection is a portable idea: it could serve as a general frame-sampling or video-summarization heuristic for any multi-view training pipeline that expects geometric continuity.
  • Since the augmentation is architecture-agnostic, a natural next test is whether the same image-to-clips recipe boosts other DUSt3R-family tasks such as camera pose regression or novel-view synthesis; the paper does not report such experiments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes Puzzles, a data augmentation framework that generates pseudo-video clips with depth and pose from single RGB-D images (Image-to-Clips) and from keyframes selected via overlap-matrix pruning (Clips-to-Clips). The authors integrate Puzzles into three video-based 3R-series pipelines (Spann3R, SLAM3R, Fast3R), retraining each on a ~14M-image corpus with and without the augmentation and evaluating on 7Scenes, NRGBD, and DTU. They report consistent full-data gains and claim that models trained on only 10% of the original data plus Puzzles match or exceed the full-data baselines.

Significance. If the headline 10% claim were properly controlled, the paper would make a useful contribution: a plug-and-play, architecture-agnostic augmentation that reduces reliance on large posed video-depth datasets, with code released and evaluation on benchmarks not used for tuning. Strengths include validation across three 3R architectures and three unseen benchmarks, the predicted-depth-versus-ground-truth robustness check in Figure 7.C, and the coverage-aware keyframe selection. However, the central empirical claim currently rests on an uncontrolled training-budget comparison and single-seed numbers, so the significance is conditional on the revision of those points.

major comments (4)
  1. [Section 4.1 / Table 1] The headline claim that 1/10 data plus Puzzles matches or exceeds full-data baselines (abstract, Section 1, Section 4.2, Figure 7.B) is not a controlled comparison. Table 1 reports only the data fraction; the text gives no number of augmented samples generated per seed frame, no total sample count, and no number of gradient updates or epochs for the 1/10 versus full conditions. Since Section 3.1 describes the augmentation as unbounded, a 1/10+Puzzles run trained for the same number of epochs as the full baseline could consume more total training examples than the full-data baseline, in which case the result would demonstrate a larger effective training set rather than data efficiency. Please report a controlled training budget, specifying what is held fixed (iterations/epochs, total samples seen, or compute), and re-report the 1/10 comparison under that control.
  2. [Table 1] The table reports single runs without error bars or multiple seeds, yet several 1/10-versus-full differences are small: for example, Spann3R on 7Scenes gives Acc 0.0389 versus 0.0388, and Fast3R gives Acc 0.0402 versus 0.0412. Without variance estimates, the statement that these conditions 'match or exceed' the baseline is not established. Please provide at least three seeds with mean and standard deviation for the headline comparisons, or a statistical test, so the reader can distinguish a real effect from seed noise.
  3. [Appendix B / Section 4.1] Section B states that RandomAffine and RandomPerspective transformations are applied jointly to images, point clouds, and overlap masks, but that 'we do not currently update the associated camera parameters'. This breaks the projective correspondence between the augmented RGB image and the depth/pointmap labels used in training, producing samples with misaligned supervision. Because this is an admitted omission in the manuscript, it must be weighed in the evaluation: either remove these 2D augmentations from the reported pipeline, update the intrinsics/extrinsics accordingly, or ablate them separately and justify why the misalignment does not harm training.
  4. [Section 3.1] The synthetic clip generation fills occluded regions with white and discards a rotation when 'too many occlusions occur', but the paper reports no statistics on how often white-filled frames appear, no fraction of discarded rotations, and no ablation comparing white-fill against alternatives such as inpainting or explicit invalid-region masking. Since synthetic clips are the sole training signal in the 1/10 condition, the transfer claim depends on this choice; a sensitivity analysis of the occlusion handling is needed.
minor comments (6)
  1. [Abstract / Section 4.2] The abstract and conclusion state the 10% result without the caveat given in Section 4.2 that the match holds on 7Scenes but not on NRGBD; please qualify the claim to match the experimental evidence.
  2. [Section 4.1] The text says 'uniform subsamples of varying size' but does not specify whether subsampling is over frames, clips, or scenes; specifying this is necessary to make the data-fraction study in Figure 7.B reproducible.
  3. [Section 3.1 / Eq. (7)] Equation (7) and the implementation detail θ_valid=100° are hard to reconcile: with that threshold, a point whose normal is up to 100° from the viewing direction passes the 'front-surface' test, which appears to admit back-facing geometry. Please clarify the intended inequality and the meaning of the threshold.
  4. [Table 1] The header 'Value△(%)' is ambiguous; please state explicitly that Δ is the relative improvement over the no-Puzzles full-data row and that lower values are better for all metrics.
  5. [Section 4.3] There are typos: 'Cips-to-Clips' should be 'Clips-to-Clips', 'stron potential' should be 'strong potential', and Appendix C contains 'botton row' instead of 'bottom row'.
  6. [Figure 7.B] The panel is described as 'the middle of Figure 7.B'; the layout, axis labels, and legend need to be clarified so the reader can interpret the data-fraction curves directly.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical augmentation scheme validated on external benchmarks, with no fitted quantity renamed as a prediction and no load-bearing self-citation chain.

full rationale

The paper's central claims are experimental, not derivational. 'Integrating Puzzles into existing video-based 3R-series pipelines consistently boosts performance' and 'models trained on only ten percent of the original data augmented with Puzzles still achieve accuracy comparable to those trained on the full dataset' are supported by retraining three existing pipelines (Spann3R, SLAM3R, Fast3R) on a unified corpus and evaluating on 7Scenes, NRGBD, and DTU, which are described in Section 4.1 as unseen evaluation datasets. None of the method's equations define these evaluation metrics: Eq. (2) derives patch intrinsics from the full-image intrinsics by scaling, Eq. (3) solves a PnP problem for patch extrinsics, Eqs. (4)-(6) apply a centroid-based rotation, and Eqs. (7)-(8) are validity filters for generated views. These constructions produce training clips but do not by construction determine Accuracy or Completion scores on the test benchmarks. The augmentation hyperparameters (η = 0.1, τ = 0.2, ρ = 0.7, θvalid = 100°, rotation range [30°, 90°]) are reported as fixed hand-chosen values in Section 4.1, not fitted to the evaluation data; therefore the fitted-input-called-prediction pattern does not apply. The robustness experiment with predicted depth uses MoGE [15], an external off-the-shelf monocular estimator, so it is not a feedback path from the proposed method to its evaluation. No reference in the bibliography is authored by the present paper's authors, so there is no self-citation loop, let alone a load-bearing one. The '10% of original data' comparison does raise an experimental-control concern: the augmented 1/10 condition's generated-sample count and total training iterations are not reported, so the comparison may not be budget-matched. However, that is a reporting and correctness-risk issue, not circularity, because the claimed result does not reduce to the augmentation by definition or by an equation that equates the output with an input. The paper is self-contained as an empirical augmentation study, and no circular step can be exhibited from its own equations or citations.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

All free parameters are hand-chosen augmentation hyperparameters, not fitted to the evaluation data. The axioms are the modeling choices that the method's effectiveness depends on.

free parameters (7)
  • Rotation angle range for camera rotation augmentation = 30 to 90 degrees
    Hand-chosen hyperparameter controlling the strength of camera rotation augmentation; affects diversity and validity of synthetic views.
  • Valid-view angle threshold theta_valid = 100 degrees
    Threshold for front-surface coverage test; determines which rotated views are kept.
  • Overlap cutoff eta for keyframe validity filtering = 0.1
    Threshold for discarding frames with low maximum overlap in Clips-to-Clips.
  • Visibility threshold tau for overlap matrix = 0.2
    Threshold for depth consistency in Eq. 9 of the overlap matrix computation.
  • Redundancy threshold rho = 0.7
    Threshold for pruning redundant keyframes in Clips-to-Clips.
  • Affine augmentation ranges = rot +/-45 deg, trans 20%, scale 0.8-1.0
    Hand-chosen ranges for 2D affine augmentation in Image-to-Clips, described in Section 3.2 and supplement B.
  • Perspective distortion scale = 0.1
    Hand-chosen scale for RandomPerspective in the 2D augmentation pipeline.
assumptions (5)
  • standard math The pinhole camera model with known intrinsics is sufficient to represent the training and evaluation data.
    Used throughout Section 3.1 for patch calibration and depth projection (Eq. 1-3).
  • domain assumption The scenes are static and rigid; no object or camera motion beyond the simulated camera trajectory.
    The method re-renders a single depth map; dynamic content is out of scope (stated in Section F, limitations).
  • domain assumption RGB-D training data with camera poses is available for the datasets used.
    Clips-to-Clips requires camera parameters and depth maps to compute overlap (Section 3.2, Eq. 9).
  • ad hoc to paper Overlapping patches of a single RGB-D image mimic the multi-view consistency of real video frames.
    This is the core inductive bias of Image-to-Clips, stated in Section 3.1: 'A key property of video is the spatial overlap between consecutive frames'.
  • ad hoc to paper Filling occluded regions with white, and discarding clips with too many occlusions, yields training signal that is net beneficial.
    Stated in Section 3.1: 'If camera rotation reveals occluded areas, we fill them with white; if too many occlusions occur, we discard that patch's rotation.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction." pith.science (2026). https://pith.science/paper/THGGLIFA

@misc{pith2026250623863,
  author       = {Pith},
  title        = {Pith review of: Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/THGGLIFA}},
  note         = {Machine review of arXiv:2506.23863}
}
read the original abstract

Multi-view 3D reconstruction remains a core challenge in computer vision. Recent methods, such as DUST3R and its successors, directly regress pointmaps from image pairs without relying on known scene geometry or camera parameters. However, the performance of these models is constrained by the diversity and scale of available training data. In this work, we introduce Puzzles, a data augmentation strategy that synthesizes an unbounded volume of high-quality posed video-depth data from a single image or video clip. By simulating diverse camera trajectories and realistic scene geometry through targeted image transformations, Puzzles significantly enhances data variety. Extensive experiments show that integrating Puzzles into existing video-based 3D reconstruction pipelines consistently boosts performance without modifying the underlying network architecture. Notably, models trained on only ten percent of the original data augmented with Puzzles still achieve accuracy comparable to those trained on the full dataset. Code is available at https://jiahao-ma.github.io/puzzles/.

Figures

Figures reproduced from arXiv: 2506.23863 by the authors.

Figure 1
Figure 1. Puzzles: A new data augmentation framework for training feedforward 3D reconstruction deep models. A. Existing video-based 3D reconstruction datasets often contain redundant frames due to fixed camera trajectories after recording (A.1). B. Puzzles increases view diversity by generating novel viewpoints from a single image (B.1) and simulates realistic camera trajectories (B.2) to support 3D reconstruction training. … view at source ↗
Figure 2
Figure 2. Puzzles: Image-to-Clips. (A) Starting from a single RGB-D image, we (B) partition it into ordered, overlapping patches, (C) simulate diverse viewpoints by calibrating virtual camera poses, and (D) generate augmented, posed images with aligned depth maps for use in 3D reconstruction. reconstructing hundreds of unordered views in a single forward pass. Despite architectural advances, these 3R-series works [8, 16, 22, … view at source ↗
Figure 3
Figure 3. Example of Image-to-Clips. Given a single input image (left of each block), the proposed Image-to-Clips method samples ordered, overlapping patches (colored boxes) and assigns simulated 6-DoF camera trajectories (frustums) to generate view-consistent video clips. Examples across human-centric illustrations, indoor environments, and outdoor landscapes demonstrate the method’s ability to synthesize diverse and realist… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Puzzles: Clips-to-Clips. (A) We begin by uniformly sampling frames from a video. (B) A pair-wise overlap matrix is computed to measure frame redundancy, with overlap visualized in purple and overlap ratios annotated in red. (C) Low-redundancy keyframes are then selecte…
Figure 5
Figure 5. Figure 5: Keyframe selection. Colors repre￾sent overlapping groups; numbers indicate pair￾wise overlap scores Oij . We discard low-overlap frames, keep the highest-coverage group, and re￾move redundant ones to form the final keyframes. The final keyframe set P removes isolated f…
Figure 6
Figure 6. Figure 6: Comparison of Clips-to-Clips and Image-to-Clips. A. Object-level reconstruction using 10 input views (#V iews = 10), where both methods perform comparably. B. Scene-level reconstruction with 75 input views (#V iews = 75), where Image-to-Clips exhibits drift and unstabl…
Figure 7
Figure 7. Figure 7: Ablation study. A. Comparison on 7Scenes [44] (scene-level) and DTU [46] (object￾level). B. Model accuracy with Puzzles at varying data fractions. C. Image-to-Clips single-image augmentation: model accuracy comparing predicted vs. ground-truth depth. All experiments us…
Figure 8
Figure 8. Figure 8: Example of Clips-to-Clips. Top: Consecutive frames from the original training clip. Middle: Selected Keyframes with corresponding patch selections. Bottom: Synthetic video clips generated from keyframes using the Image-to-Clips process. A Additional Examples of Clips-t…
Figure 9
Figure 9. Figure 9: From random sampling to coverage-aware key-frame selection. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 11
Figure 11. Figure 11: Reconstruction results without (left) and with (right) the proposed Puzzles augmentation. [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 10
Figure 10. Figure 10: Reconstruction accuracy (lower Acc. is better) for translation-only augmen￾tation (TransAug) versus combined translation and rotation augmentation (Trans+RotAug). As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 19 canonical work pages

  1. [1]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam

    Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam. IEEE transactions on robotics, 37(6):1874–1890, 2021

    2021

  2. [2]

    Lsd-slam: Large-scale direct monocular slam

    Jakob Engel, Thomas Schöps, and Daniel Cremers. Lsd-slam: Large-scale direct monocular slam. In European conference on computer vision, pages 834–849. Springer, 2014

    work page 2014

  3. [3]

    Parallel tracking and mapping on a camera phone

    Georg Klein and David Murray. Parallel tracking and mapping on a camera phone. In 2009 8th IEEE International Symposium on Mixed and Augmented Reality, pages 83–86. IEEE, 2009

    work page 2009

  4. [4]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceed- ings of the IEEE conference on computer vision and pattern recognition , pages 4104–4113, 2016

    work page 2016

  5. [5]

    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

    work page 2015

  6. [6]

    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

    work page 2015

  7. [7]

    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 501–518. Springer, 2016

    2016

  8. [8]

    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

Show all 48 references
  1. [9]

    3d reconstruction with spatial memory

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

    2024 arXiv

  2. [10]

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

    Yuzheng Liu, Siyan Dong, Shuzhe Wang, 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

  3. [11]

    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

  4. [12]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017

    2017

  5. [13]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12–22, 2023

    2023

  6. [14]

    Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction

    Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer visi...

    2021

  7. [15]

    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. arXiv preprint arXiv:2410.19115, 2024

    arXiv

  8. [16]

    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

  9. [17]

    Splatt3r: Zero-shot gaussian splatting from uncalibrated image pairs

    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. 10

    2024 arXiv

  10. [18]

    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. arXiv preprint arXiv:2410.24207, 2024

    2024 arXiv

  11. [19]

    Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization

    Siyan Dong, Shuzhe Wang, Shaohui Liu, Lulu Cai, Qingnan Fan, Juho Kannala, and Yanchao Yang. Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization. arXiv preprint arXiv:2412.08376, 2024

    2024 arXiv

  12. [20]

    Driv3r: Learning dense 4d reconstruction for autonomous driving

    Xin Fei, Wenzhao Zheng, Yueqi Duan, Wei Zhan, Masayoshi Tomizuka, Kurt Keutzer, and Jiwen Lu. Driv3r: Learning dense 4d reconstruction for autonomous driving. arXiv preprint arXiv:2412.06777, 2024

    2024 arXiv

  13. [21]

    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

  14. [22]

    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

  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. arXiv preprint arXiv:2503.11651, 2025

    2025 arXiv

  16. [24]

    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. arXiv preprint arXiv:2502.12138, 2025

    2025

  17. [25]

    Random erasing data augmentation

    Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence , volume 34, pages 13001–13008, 2020

    2020

  18. [26]

    Hide-and- seek: A data augmentation technique for weakly-supervised localization and beyond

    Krishna Kumar Singh, Hao Yu, Aron Sarmasi, Gautam Pradeep, and Yong Jae Lee. Hide-and- seek: A data augmentation technique for weakly-supervised localization and beyond. arXiv preprint arXiv:1811.02545, 2018

    2018 arXiv

  19. [27]

    Gridmask data augmentation

    Pengguang Chen, Shu Liu, Hengshuang Zhao, Xingquan Wang, and Jiaya Jia. Gridmask data augmentation. arXiv preprint arXiv:2001.04086, 2020

    2001 arXiv

  20. [28]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  21. [29]

    Augmix: A simple data processing method to improve robustness and uncertainty

    Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshmi- narayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019

    1912 arXiv

  22. [30]

    Improved regularization of convolutional neural networks with cutout

    Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017

    2017 arXiv

  23. [31]

    Cutmix: Regularization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  24. [32]

    Learning data augmentation strategies for object detection

    Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pages 566–58...

    2020

  25. [33]

    Classmix: Segmentation-based data augmentation for semi-supervised learning

    Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369–1378, 2021

    2021

  26. [34]

    Matching anything by segmenting anything

    Siyuan Li, Lei Ke, Martin Danelljan, Luigi Piccinelli, Mattia Segu, Luc Van Gool, and Fisher Yu. Matching anything by segmenting anything. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963–18973, 2024. 11

    2024

  27. [35]

    Bridging images and videos: A simple learning framework for large vocabulary video object detection

    Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In European Conference on Computer Vision, pages 238–258. Springer, 2022

    2022

  28. [36]

    Matchanything: Universal cross-modality image matching with large-scale pre-training

    Xingyi He, Hao Yu, Sida Peng, Dongli Tan, Zehong Shen, Hujun Bao, and Xiaowei Zhou. Matchanything: Universal cross-modality image matching with large-scale pre-training. In Arxiv, 2025

    2025

  29. [37]

    Random sample consensus: a paradigm for model fitting with applica- tions to image analysis and automated cartography

    MA FISCHLER AND. Random sample consensus: a paradigm for model fitting with applica- tions to image analysis and automated cartography. Commun. ACM, 24(6):381–395, 1981

    1981

  30. [38]

    Multiple view geometry in computer vision

    Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003

    2003

  31. [39]

    Rodrigues’ rotation formula—Wikipedia, the free encyclope- dia

    Wikipedia contributors. Rodrigues’ rotation formula—Wikipedia, the free encyclope- dia. https://en.wikipedia.org/wiki/Rodrigues%27_rotation_formula,

  32. [40]

    Open3D: A modern library for 3D data processing

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018

    2018 arXiv

  33. [41]

    Arkitscenes: A diverse real- world dataset for 3d indoor scene understanding using mobile rgb-d data

    Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real- world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021

    2021 arXiv

  34. [42]

    Habitat: A platform for embodied ai research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, ...

    2019

  35. [43]

    Blendedmvs: A large-scale dataset for generalized multi-view stereo networks

    Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1790–1799, 2020

    2020

  36. [44]

    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 2930–2...

    2013

  37. [45]

    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

  38. [46]

    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

  39. [47]

    original sampling

    Riku Murai, Eric Dexheimer, and Andrew J Davison. Mast3r-slam: Real-time dense slam with 3d reconstruction priors. arXiv preprint arXiv:2412.12392, 2024. 12 Figure 8: Example of Clips-to-Clips. Top: Consecutive frames from the original training clip. Middle: Selected Keyframes...

    2024 arXiv

  40. [2025]

    Accessed: 2025-05-04

    2025

Pith tools

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