Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

VideoLifter: Lifting Videos to 3D with Fast Hierarchical Stereo Alignment

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read VideoLifter reconstructs 3D scenes from unposed monocular video in about 25 minutes, cutting training time by over 82%.

desk verdict Solid fragment-based video-to-3D pipeline, but the 82% speedup claim is confounded by training-view subsampling and needs a matched-frame comparison. read the letter →

arxiv 2501.01949 v3 pith:U5YJGBDC submitted 2025-01-03 cs.CV

Wenyan Cong , Hanqing Zhu , Kevin Wang , Jiahui Lei , Colton Stearns , Yuanhao Cai , Leonidas Guibas , Zhangyang Wang
show 1 more author
Zhiwen Fan
This is my paper · ORCID
classification cs.CV
keywords video-to-3Dreconstruction3DGaussiansplattingcameraposeestimationnovelviewsynthesisstructure-from-motion-freehierarchicalalignmentfragment-basedlearnedstereopriors
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 takes on a practical bottleneck in 3D reconstruction from video: current pose-free methods are either slow (hours of training) or drift over long sequences, and most still rely on structure-from-motion for camera intrinsics. VideoLifter claims a fragment-first path: split the video into short windows, register each window using learned stereo priors while optimizing only a small pose-plus-scale parameter set per frame, then merge the resulting 3D Gaussian clouds in a tree-based hierarchy with key-frame guidance. On Tanks and Temples and CO3D-V2, the authors report cutting training time by more than 82% (roughly 5x) while improving novel-view quality and pose accuracy over state-of-the-art pose-free baselines, with no ground-truth camera parameters required. If this holds, dense scene capture from an unposed monocular video becomes a roughly 25-minute, single-GPU job.

What carries the argument

The load-bearing mechanism is the fragment-and-merge pipeline. A fragment is a disjoint window of $k$ consecutive frames (here $k=4$), anchored by its first frame, the key frame. A learned stereo prior supplies dense 3D point maps and pairwise relative poses; instead of optimizing millions of point coordinates globally, VideoLifter optimizes only a 6-dimensional quaternion pose and a 1-dimensional scale factor per view, and uses PnP-RANSAC and a median ratio for scale. Each fragment is turned into a local 3D Gaussian set, and these are merged through three iterative stages: key-frame-guided inter-fragment alignment, visibility-mask-driven pruning of redundant Gaussians, and joint optimization of Gaussian properties and poses. The visibility mask $M(p)=\operatorname{Conf}(p)>\beta + D(p)>0$ marks pixels the reference fragment already reconstructs, so only complementary Gaussians from the new fragment are carried into the merge.

What would settle it

Run VideoLifter and a strong pose-free baseline on a monocular video with large textureless regions, strong specular highlights, or heavy motion blur, and compare the estimated camera trajectory (ATE) and rendered novel views against ground-truth poses or COLMAP. If the learned stereo prior's point maps are unreliable there and VideoLifter's pose or view-synthesis error collapses relative to baselines that optimize geometry directly, the claim of robust SfM-free reconstruction fails.

Watch

Extended reading notes

Core claim

VideoLifter's central claim is that a globally consistent 3D scene can be reconstructed from a long monocular video without any SfM or precomputed camera parameters, by treating the video as a collection of short fragments rather than as one growing sequence. The paper reports that this fragment-based local-to-global design is both faster and more accurate than the frame-by-frame progressive baselines: more than 82% less training time, superior SSIM/PSNR/LPIPS, and ATE tied with the best pose-free baseline, while also removing the need for supplied intrinsics. The underlying discovery is that dense point maps, relative poses, and scale factors can be extracted from the key frames of each fragment using learned stereo priors, and that the resulting per-fragment 3D Gaussian models can be merged hierarchically without losing local detail or accumulating long-term drift.

Load-bearing premise

The entire geometry—point maps, relative poses, and focal lengths—comes from a pretrained stereo model, so if that model's geometry is wrong for a given video, the fragment registration and all later merging inherit the error and no other component can correct it.

Editorial extensions

If this is right

  • A monocular video alone—no precomputed camera poses, no intrinsics—can be turned into a dense 3D Gaussian scene in about 25 minutes on one GPU, a roughly 5x speedup over the closest pose-free baseline.
  • Long videos no longer require quadratic memory: fragment registration uses a subgraph whose cost scales linearly with frame count, so sequences that make global stereo optimization run out of memory can still be processed.
  • Because merging is hierarchical and key-frame-guided, drift is compressed along the fragment count rather than the frame count, making pose accuracy competitive with methods that receive ground-truth intrinsics.
  • Removing the COLMAP dependency extends reconstruction to videos where SfM is known to fail, such as low-texture or lighting-challenged footage, which is exactly the regime the paper evaluates.

Reading between the lines

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

  • Editorial: The same 'optimize only pose and scale, not point maps' trick could be ported to other dense 3D representations, such as neural radiance fields, where a fragment-registration stage would similarly cut optimization cost.
  • Editorial: Because the pipeline processes disjoint fragments and merges incrementally, it could be adapted to streaming capture: reconstruct a new chunk as it arrives and splice it into the existing Gaussian field, at the cost of some quality.
  • Editorial: The paper's uniform window choice is a simplification; fragments chosen by visual similarity or estimated motion magnitude could reduce alignment failures on videos with abrupt view changes, though the paper does not test this.
  • Editorial: The claim's ceiling is set by the stereo prior; a direct test would replace the prior with a different geometric estimator and measure the quality change, quantifying how much of the gain comes from the prior versus the fragment-merge design.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. VideoLifter proposes a video-to-3D reconstruction pipeline that avoids SfM and known camera parameters by processing the input video in short fragments, registering each fragment with learned stereo priors from MASt3R, and then merging fragment-level 3D Gaussian reconstructions through a tree-based hierarchical alignment with key-frame guidance. The paper claims a training-time reduction of over 82% relative to CF-3DGS while achieving better novel-view synthesis quality and comparable or better pose accuracy on Tanks and Temples and CO3D-V2. The method is evaluated against several pose-free and pose-optimizing baselines, with ablations on fragment size, confidence threshold, and design components.

Significance. If the reported speedup is genuine, the paper addresses an important practical bottleneck: pose-free video-to-3D reconstruction that scales to long sequences without the drift and memory issues of frame-by-frame progressive methods. The hierarchical local-to-global design with efficient key-frame subgraphs is a sensible and useful contribution, and the experimental comparison is broad, covering rendering quality, pose accuracy, training time, and memory usage. The reliance on a strong pretrained stereo model (MASt3R) is clearly stated, and the paper demonstrates flexibility by comparing with alternative priors (LoFTR + Metric3Dv2). However, the headline efficiency claim is confounded by unequal training-view sampling between the proposed method and the baselines, which means the central quantitative contribution is not yet convincingly established.

major comments (3)
  1. [§4.1, §4.2, Tables 3 and 4] The training-time and quality comparison is confounded by the statement in §4.1: 'We uniformly sample 1/2 and 1/4 training views on Tanks and Temples and CO3D-V2, respectively.' The baselines are re-trained with their official code on the full dataset splits, while VideoLifter optimizes over a subset of frames. The reported times (e.g., 26min vs. 2h20min on Tanks and Temples) therefore reflect both algorithmic efficiency and a reduced optimization workload, and the quality comparison is not apples-to-apples: evaluating on held-out frames after training on fewer views is a different task than training on all views. The abstract's claim of 'reducing training time by over 82%' is load-bearing, and it is not yet established as an algorithmic speedup. The authors should either retrain all baselines on the same subsampled views, or provide a controlled experiment that varies the number of training views for both VideoLifter and CF-3DGS, reporting time and quality as a function of view count. Without this, the central contribution of the paper is not rigorously supported.
  2. [§3.2, Eq. (1), §4.1] The entire pipeline depends on MASt3R's learned priors to produce point maps, relative poses, and focal lengths from image pairs and keyframe subgraphs. The paper evaluates only on Tanks and Temples and CO3D-V2, where MASt3R works well, and provides no failure analysis or robustness bounds for low-texture scenes, unusual lighting, or camera configurations outside MASt3R's training distribution. Since the stated motivation includes 'in-the-wild video scenarios' where SfM fails, the robustness of the prior model is load-bearing for the claimed applicability. Please add experiments on challenging sequences (e.g., low-texture indoor scenes, strong illumination changes, or abrupt viewpoint jumps) or a quantitative analysis of where the method fails and how errors propagate through the hierarchical merging. At minimum, the paper should explicitly discuss the failure modes inherited from MASt3R and the practical limits of the method.
  3. [§4.4, Table 5] The ablation study is conducted on a single CO3D-V2 scene (34_1403_4393) and reports no error bars or multiple seeds. The conclusions about the optimal fragment size k and confidence threshold beta are drawn from one instance, and the optimal values may vary across scenes; for example, the quality difference between k=2 and k=4 is small (SSIM 0.8936 vs. 0.8957) and without variance estimates it is unclear whether this is significant. The paper should report ablations on at least two or three scenes, or provide variance information, to support the claim that the design choices are generally beneficial rather than scene-specific.
minor comments (5)
  1. [Figure 1 caption] The caption uses 'Nope-NeRF' while the text and tables use 'NoPe-NeRF'; please correct the spelling for consistency.
  2. [§3.2, 'Efficient intra-Fragment Feature Registration'] The phrase 'trys to solve' should be 'tries to solve' (also appears near Table 5 in the discussion of MASt3R initialization).
  3. [§3.2, Eq. (1)] The notation in Eq. (1) is somewhat dense: the variables Pv,e, Ov,e, and the summation indices are not fully clarified. Please define all symbols explicitly, especially the role of the confidence map O and the scale factor sigma_e, to make the optimization objective self-contained.
  4. [Table 5] The row 'Local: Use MASt3R MVS Init.' should clarify whether this variant uses the same fragment-based pipeline with MASt3R replacing only the intra-fragment registration, or a different overall setup; this would help readers understand the comparison.
  5. [§4.2, Tables 3 and 4] The ATE metrics are computed against COLMAP-generated poses used as ground truth. Since MASt3R is trained on similar internet-scale imagery, there may be systematic bias; please discuss or provide an additional evaluation (e.g., relative pose error or qualitative trajectory plots) to contextualize the ATE values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's core claims are direct empirical measurements against external baselines, and its only self-citation is used as a comparison point rather than as load-bearing justification.

full rationale

VideoLifter reports measured training times, rendering metrics, and ATE on Tanks and Temples and CO3D-V2, compared with baselines retrained from official code. No output quantity is defined in terms of another, and no fitted parameter is renamed as a prediction; k and beta are ablated as hyperparameters. The pipeline depends on MASt3R, an external pretrained model, and the paper includes an ablation swapping MASt3R for LoFTR+Metric3Dv2, so the prior is not assumed by construction. The self-citation to InstantSplat appears as a baseline and as an ablation variant ('Local: Use MASt3R MVS Init. [11]'), not as authority for the paper's claims. The implementation detail that VideoLifter uniformly samples 1/2 or 1/4 of training views while baselines are retrained on full splits is a fairness/confound concern about the headline speedup, but it is not circularity: training time is measured, not derived from the method's own outputs. The central claims therefore have independent empirical content.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. Its free parameters are standard hyperparameters tuned on a single scene. The key assumptions concern the reliability of the external MASt3R prior and the sufficiency of the sparse keyframe graph.

free parameters (3)
  • Fragment size k = 4
    Number of frames per fragment; chosen via ablation on a single CO3D-V2 scene. Balancing intra-fragment registration accuracy and inter-fragment alignment complexity.
  • Confidence threshold beta = 0.9
    Threshold in the visibility mask for deciding which Gaussians to merge; ablation over 0.5, 0.9, 0.99 on one scene selects 0.9.
  • Neighbor count in keyframe subgraph = 4
    Number of nearest neighboring frames connected to each key frame in the sparse graph; chosen as a constant with no sensitivity analysis.
assumptions (5)
  • domain assumption All frames share a common intrinsic matrix.
    Stated in Sec 3.1: 'We assume that all frames share a common intrinsic matrix, as they are from a single monocular video.' This can fail with zoom or focus changes.
  • domain assumption MASt3R provides sufficiently accurate pointmaps, poses, and focal lengths for initialization.
    The whole pipeline relies on this pretrained stereo model. The paper demonstrates robustness only on datasets where MASt3R works; no failure analysis is given.
  • domain assumption A subgraph with key frames and their four nearest neighbors preserves alignment accuracy.
    Sec 3.2 states that edges between key frames and distant neighbors can be safely pruned due to co-visibility. This is a modeling choice not independently validated.
  • domain assumption Uniform disjoint windows are sufficient for fragmentation.
    Sec 3.2 uses uniform windows and acknowledges that frame-to-frame similarity could help with abrupt view changes. This assumption could hurt on unconstrained videos.
  • domain assumption Pinhole camera model.
    Conclusion explicitly mentions this as a limitation: 'A key limitation, shared with prior pose-free methods (e.g., CF-3DGS), is the assumption of a pinhole camera model.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of VideoLifter: Lifting Videos to 3D with Fast Hierarchical Stereo Alignment." pith.science (2026). https://pith.science/paper/U5YJGBDC

@misc{pith2026250101949,
  author       = {Pith},
  title        = {Pith review of: VideoLifter: Lifting Videos to 3D with Fast Hierarchical Stereo Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5YJGBDC}},
  note         = {Machine review of arXiv:2501.01949}
}
read the original abstract

Efficiently reconstructing 3D scenes from monocular video remains a core challenge in computer vision, vital for applications in virtual reality, robotics, and scene understanding. Recently, frame-by-frame progressive reconstruction without camera poses is commonly adopted, incurring high computational overhead and compounding errors when scaling to longer videos. To overcome these issues, we introduce VideoLifter, a novel video-to-3D pipeline that leverages a local-to-global strategy on a fragment basis, achieving both extreme efficiency and SOTA quality. Locally, VideoLifter leverages learnable 3D priors to register fragments, extracting essential information for subsequent 3D Gaussian initialization with enforced inter-fragment consistency and optimized efficiency. Globally, it employs a tree-based hierarchical merging method with key frame guidance for inter-fragment alignment, pairwise merging with Gaussian point pruning, and subsequent joint optimization to ensure global consistency while efficiently mitigating cumulative errors. This approach significantly accelerates the reconstruction process, reducing training time by over 82% while holding better visual quality than current SOTA methods.

Figures

Figures reproduced from arXiv: 2501.01949 by the authors.

Figure 1
Figure 1. Novel View Synthesis and Training Time Comparisons. VideoLifter does not require precom￾puted camera parameters (i.e., camera intrinsics K from COLMAP), reduces the training time required by the most relevant baseline CF-3DGS [1] by 82% while improving image quality (SSIM). 1 Introduction Reconstructing 3D scenes from image observations is a longstanding problem in computer vision, with applications spanning AR/VR, … view at source ↗
Figure 2
Figure 2. CF-3DGS’s frame-by-frame pipeline accumulates errors with long-term drift, while our method compresses drift error along # frames. Results are tested on 247_26441_50907 from CO3D-V2. using 3D Gaussians G along with estimated camera intrinsics K and extrinsics T = {Ti ∈ R 3×4} N i=1. We assume that all frames share a common intrinsic matrix, as they are from a single monocular video. Key Challenges While NeRF and 3D-… view at source ↗
Figure 3
Figure 3. Network Architecture. Given uncalibrated images, VideoLifter first employs learned priors for efficient fragment registration. The independently optimized 3D Gaussians from fragment are then hierarchically aligned into a globally coherent 3D representation. In response, we propose VideoLifter, an efficient and high-quality video-to-3D reconstruction pipeline, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p004… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Hierarchical Gaussian Alignment. The process iteratively performs three stages: 1) joint opti￾mization of camera poses and local Gaussians (pink), 2) cross-fragment alignment for new local Gaussian (pur￾ple), and 3) visibility masking and pairwise merging of local Gaus…
Figure 5
Figure 5. Figure 5: Visibility Mask showing rendering G f 1 into next two novel frames in G f 2 . White region de￾notes faithful reconstruction using G f 1 , while black represents pixels unseen from G f 1 . With visibility mask, we select complementary gaussians from G f 2 , merging them…
Figure 6
Figure 6. Figure 6: Visual Comparisons between VideoLifter and other baselines. The insets highlight the details of renderings. VideoLifter achieves faithful 3D reconstruction, preserves better details, and alleviates incremental error in progressive learning. scale inconsistencies, makin…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. E3D-Bench: A Benchmark for End-to-End 3D Geometric Foundation Models

    cs.CV 2025-06 conditional novelty 6.0 of 10

    E3D-Bench compares 16 3D geometric foundation models on depth, reconstruction, pose, and view-synthesis tasks with a unified evaluation toolkit.

  2. Decomposing Densification in Gaussian Splatting for Faster 3D Scene Reconstruction

    cs.CV 2025-07 conditional novelty 5.0 of 10

    A split-then-clone densification schedule with energy-guided multi-resolution training roughly halves 3D Gaussian Splatting training time while keeping reconstruction quality.

Reference graph

Works this paper leans on

38 extracted references · 23 canonical work pages · cited by 2 Pith papers

  1. [1]

    Colmap- free 3d gaussian splatting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap- free 3d gaussian splatting. arXiv preprint arXiv:2312.07504, 2023

    arXiv 2023

  2. [2]

    Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing Scenes As Neural Radiance Fields for View Synthesis. Communications of the ACM, 65(1):99–106, 2021

    work page 2021

  3. [3]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1– 14, 2023

    work page 2023

  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

    2016

  5. [5]

    Progressively optimized local radiance fields for robust view synthesis

    Andreas Meuleman, Yu-Lun Liu, Chen Gao, Jia-Bin Huang, Changil Kim, Min H Kim, and Johannes Kopf. Progressively optimized local radiance fields for robust view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16539–16548, 2023

    work page 2023

  6. [6]

    Global fusion of relative motions for robust, accurate and scalable structure from motion

    Pierre Moulon, Pascal Monasse, and Renaud Marlet. Global fusion of relative motions for robust, accurate and scalable structure from motion. In Proceedings of the IEEE international conference on computer vision, pages 3248–3255, 2013

    work page 2013

  7. [7]

    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 interna- tional conference on computer vision, pages 5987–5997, 2021

    work page 2021

  8. [8]

    Nope- nerf: Optimising neural radiance field with no pose prior

    Wenjing Bian, Zirui Wang, Kejie Li, Jia-Wang Bian, and Victor Adrian Prisacariu. Nope- nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4160–4169, 2023

    work page 2023

Show all 38 references
  1. [9]

    Barf: Bundle-adjusting neural radiance fields

    Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741–5751, 2021

    2021

  2. [10]

    Nerf–: Neural radiance fields without known camera parameters

    Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf–: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021

    2021 arXiv

  3. [11]

    Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds

    Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024

    2024 arXiv

  4. [12]

    Grounding image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024

    2024 arXiv

  5. [13]

    Novel view synthesis in tensor space

    Shai Avidan and Amnon Shashua. Novel view synthesis in tensor space. InProceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1034–1040. IEEE, 1997

    1997

  6. [14]

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

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

    2019

  7. [15]

    Point-nerf: Point-based neural radiance fields

    Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5438–5448, 2022

    2022

  8. [16]

    Towards internet-scale multi-view stereo

    Yasutaka Furukawa, Brian Curless, Steven M Seitz, and Richard Szeliski. Towards internet-scale multi-view stereo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 1434–1441. IEEE, 2010. 10

    2010

  9. [17]

    Dtam: Dense tracking and mapping in real-time

    Richard A Newcombe, Steven J Lovegrove, and Andrew J Davison. Dtam: Dense tracking and mapping in real-time. In 2011 international conference on computer vision, pages 2320–2327. IEEE, 2011

    2011

  10. [18]

    Visualsfm: A visual structure from motion system, 2011.URL http://www

    Changchang Wu et al. Visualsfm: A visual structure from motion system, 2011.URL http://www. cs. washington. edu/homes/ccwu/vsfm, 14:2, 2011

    2011

  11. [19]

    Distinctive image features from scale-invariant keypoints

    David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60(2):91–110, 2004

    2004

  12. [20]

    Xufeng Han, Thomas Leung, Yangqing Jia, Rahul Sukthankar, and Alexander C. Berg. Matchnet: Unifying feature and metric learning for patch-based matching. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3279–3286, 2015

    2015

  13. [21]

    Scene reconstruction and visualization from internet photo collections: A survey

    Noah Snavely. Scene reconstruction and visualization from internet photo collections: A survey. IPSJ Transactions on Computer Vision and Applications, 3:44–66, 2011

    2011

  14. [22]

    Improving the efficiency of hierarchical structure-and-motion

    Riccardo Gherardi, Michela Farenzena, and Andrea Fusiello. Improving the efficiency of hierarchical structure-and-motion. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 1594–1600. IEEE, 2010

    2010

  15. [23]

    Self-calibrating neural radiance fields

    Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5846–5854, 2021

    2021

  16. [24]

    Sparf: Neural radiance fields from sparse and noisy poses

    Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190–4200, 2023

    2023

  17. [25]

    Tracknerf: Bundle adjusting nerf from sparse and noisy views via feature tracks

    Jinjie Mai, Wenxuan Zhu, Sara Rojas, Jesus Zarzar, Abdullah Hamdi, Guocheng Qian, Bing Li, Silvio Giancola, and Bernard Ghanem. Tracknerf: Bundle adjusting nerf from sparse and noisy views via feature tracks. arXiv preprint arXiv:2408.10739, 2024

    2024 arXiv

  18. [26]

    Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs

    Zezhou Cheng, Carlos Esteves, Varun Jampani, Abhishek Kar, Subhransu Maji, and Ameesh Makadia. Lu-nerf: Scene and pose estimation by synchronizing local unposed nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 18312– 18321, 2023

    2023

  19. [27]

    A construct-optimize approach to sparse view synthesis without camera pose

    Kaiwen Jiang, Yang Fu, Yash Belhe, Xiaolong Wang, Hao Su, Ravi Ramamoorthi, et al. A construct-optimize approach to sparse view synthesis without camera pose. arXiv preprint arXiv:2405.03659, 2024

    2024 arXiv

  20. [28]

    Splat-slam: Globally optimized rgb-only slam with 3d gaussians

    Erik Sandström, Keisuke Tateno, Michael Oechsle, Michael Niemeyer, Luc Van Gool, Martin R Oswald, and Federico Tombari. Splat-slam: Globally optimized rgb-only slam with 3d gaussians. arXiv preprint arXiv:2405.16544, 2024

    2024 arXiv

  21. [29]

    Loftr: Detector-free local feature matching with transformers

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922–8931, 2021

    2021

  22. [30]

    Gim: Learning generalizable image matcher from internet videos

    Xuelun Shen, Zhipeng Cai, Wei Yin, Matthias Müller, Zijun Li, Kaixuan Wang, Xiaozhi Chen, and Cheng Wang. Gim: Learning generalizable image matcher from internet videos. arXiv preprint arXiv:2402.11095, 2024

    2024 arXiv

  23. [31]

    Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation

    Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis...

    2024

  24. [32]

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

    Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381–395, 1981. 11

    1981

  25. [33]

    Tanks and temples: Bench- marking large-scale scene reconstruction

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Bench- marking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

    2017

  26. [34]

    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

  27. [35]

    Image quality assessment: from error visibility to structural similarity

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

    2004

  28. [36]

    The unrea- sonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unrea- sonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  29. [37]

    A benchmark for the evaluation of rgb-d slam systems

    Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573–580. IEEE, 2012

    2012

  30. [38]

    Zoedepth: Zero-shot transfer by combining relative and metric depth

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 12

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.