Pith. sign in

REVIEW 4 major objections 4 minor 111 references

HETA++: Global Structure-from-Motion with Hybrid Explicit Translation Averaging

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read HETA++ claims that hybrid explicit translation averaging—refining relative translations with global rotations, filtering them by a spectral confidence score, and optimizing camera positions jointly with feature tracks—lets global SfM beat p

desk verdict HETA++ is a solid, incremental engineering upgrade to the authors' own HETA, with credible accuracy and runtime gains on standard benchmarks, but the dataset-relative confidence filter and lack of released code keep the central claims from being fully verified. read the letter →

arxiv 2607.15912 v1 pith:C7ZQ66PC submitted 2026-07-17 cs.CV

Peilin Tao , Hainan Cui , Mengqi Rong , Shuhan Shen This is my paper · ORCID
classification cs.CV
keywords Globalstructure-from-motionTranslationaveragingFeaturetracksRotationBundleadjustment3DreconstructionCameraposeestimationSpectralconfidencefiltering
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

HETA++ is trying to establish that global Structure-from-Motion can match or beat incremental reconstruction in accuracy while keeping its speed advantage, if translation averaging is done right. The paper's recipe is hybrid: use relative translations and raw feature tracks together, first re-estimating and filtering relative translations with the help of global rotations, then initializing camera positions and 3D points with convex objectives, refining them with a non-bilinear angle-based objective, and finally polishing poses with selected tracks before full bundle adjustment. A sympathetic reader should care because global SfM is the scalable route for very large unordered image collections, and its weakest step—translation averaging—has been the reason global methods lag in accuracy. If the claim holds, the practical trade-off that forced users toward incremental pipelines largely disappears.

What carries the argument

The central mechanism is the hybrid explicit objective: both camera-to-camera relative translations and camera-to-point feature rays are treated as unit bearing constraints on unknown positions. Three pieces carry the argument. A spectral confidence score, defined from the ratio e2/e1 of the two smallest eigenvalues of the epipolar-plane normals matrix, is mapped through the dataset's 75th percentile to a score in (0,1) that weights each relative translation and drives a 10-degree global-consistency filter. A non-bilinear angle-based residual, Eq. (7), computes the orthogonal distance between observed and predicted directions and saturates at one for angles of 90 degrees or more, giving the

What would settle it

Run a controlled experiment on a scene with ground-truth camera poses: compute refined relative translations while varying the number of feature matches and the parallax angle, then plot the eigenvalue ratio Ra against the actual angular error of each translation. If a sizable share of edges with Ra above the dataset's 75th percentile have larger errors than edges below it—or if the 10-degree global-consistency filter systematically removes accurate edges on near-collinear trajectories—then the confidence heuristic is not selecting the right edges and the reported accuracy gain would not repro

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that a hybrid explicit translation averaging framework yields the most accurate camera positions among global SfM methods on standard unordered benchmarks, and does so at lower runtime. The decisive step is to treat relative translations and feature rays as the same type of normalized directional constraint, then process them in a three-stage sequence: local re-estimation of relative translations with parallax-weighted epipolar-plane normals, global filtering via a spectral confidence score computed from the eigenvalue ratio of the normals matrix, and hybrid initialization by convex L1 objectives followed by non-bilinear angle-based refinement.

Load-bearing premise

The load-bearing premise is that the spectral confidence score—the eigenvalue ratio e2/e1 of epipolar-plane normals, anchored to the dataset's 75th percentile—reliably separates accurate from inaccurate relative translations under the 10-degree consistency filter, and, as the paper's own limitation section records, that the view graph is free enough of structurally ambiguous matches (symmetries, repetitive facades) that such local and global filters can identify outliers; if

Editorial extensions

If this is right

  • Global SfM can approach or match the camera-pose accuracy of incremental pipelines without ordering-dependent drift, since all cameras are optimized together and errors are spread across the scene.
  • Translation averaging no longer needs to choose between relative translations and feature tracks: the hybrid formulation inherits the speed of translation-only methods and the robustness of track-based methods, including on near-collinear trajectories.
  • The spectral eigenvalue-ratio confidence provides a practical, dataset-adaptive way to remove bad relative translations before expensive optimization, which should make global pipelines more robust on Internet photo collections.
  • The joint optimization stage lowers the method's dependence on the accuracy of rotation averaging, which is the other known bottleneck in global SfM.
  • Large-scale reconstructions become cheaper: the paper reports substantial runtime reductions relative to leading global pipelines on the same view-track graph input.

Reading between the lines

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

  • This is an editorial extension: the same eigenvalue-ratio confidence heuristic could be lifted out of SfM and applied to any pairwise measurement averaging problem, such as rotation averaging or pose-graph optimization, where per-edge uncertainty can be characterized by the spectrum of a local normal matrix.
  • Also editorial: the non-bilinear projection trick—removing auxiliary scale variables while keeping a bounded residual—is a general pattern for compressing bilinear relaxations when the number of observations is in the millions; other geometric vision pipelines may borrow it for memory savings.
  • Another testable extension: the fixed 10-degree filter threshold and 0.5-degree parallax cutoff could be made scene-adaptive; on weakly constrained graphs a smaller threshold may remove too many edges, while on dense graphs a larger one might admit outliers. Varying these thresholds per dataset would clarify how much of the gain comes from the filter itself.
  • The paper's own limitation on symmetric or repetitive scenes suggests the view-graph disambiguation problem, not translation averaging itself, is the next bottleneck; as the authors note, an external camera-triplet filtering stage turns a failure into a success.
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 / 4 minor

Summary. The paper presents HETA++, a global Structure-from-Motion pipeline built on hybrid explicit translation averaging. The pipeline first refines relative translations using global rotations, assigns confidence scores based on an eigenvalue-ratio spectral heuristic, and filters globally inconsistent relative translations. It then initializes camera positions and 3D points via a convex L1 cross-product objective and alternating triangulation, refines them with a non-bilinear angle-based objective, and finally applies a joint optimization stage with coverage-balanced feature-track selection followed by complete bundle adjustment. Experiments on KITTI, ETH3D, 1DSfM, and LaMAR compare HETA++ against COLMAP, CReTA, LiGT, HETA, GLOMAP, and a feed-forward baseline, with the paper claiming superior accuracy and efficiency.

Significance. If the claims hold, HETA++ would be a practical improvement to global SfM, particularly for near-collinear motion sequences where camera-only translation averaging degenerates. The manuscript has clear strengths: the geometric derivation in Eq. (9) justifying parallax-dependent weighting is internally consistent; the non-bilinear angle objective in Eq. (7) removes per-observation scale variables; and the ablation design is generally careful, e.g., Table 7 isolates the initialization and the joint-optimization modules under a common bundle-adjustment backend. The use of multiple public datasets and a shared view-track graph for all global baselines is also a strength. However, the central 'outperforms state-of-the-art' claim rests on a dataset-relative confidence heuristic whose calibration is not validated, on several hand-set thresholds with no sensitivity analysis, and on single-run benchmark numbers whose margins over GLOMAP on 1DSfM are often below 1 AUC point. The paper's own Section 5 documents failure on structurally ambiguous scenes unless an external disambiguation stage is added. The contribution is therefore solid but incremental, and the evidence as presented is not yet

major comments (4)
  1. [§3.3.2–3.3.3, Eq. (11)] The global-consistency filter is load-bearing for the whole pipeline, but its calibration is under-supported. Eq. (11) anchors C_ij=0.5 at the dataset's own 75th percentile Ra_0.75, making the retained edge set relative to the noise level of each scene. Figure 6 is the only calibration evidence; there is no precision/recall analysis, no held-out validation, and no end-to-end ablation isolating the filter. The 10° consistency check in §3.3.3 uses coarse positions obtained from the same noisy translations via Eq. (12), so a miscalibrated Ra can either keep bad edges or discard useful ones. Fig. 11 shows CDFs of relative-translation errors, not final camera-pose AUC with and without the filter. Please add sensitivity studies for Ra_0.75, the 10° threshold, and the 0.5° median-parallax cutoff, and an end-to-end ablation that quantifies the filter's contribution.
  2. [§4.1.1, Table 1] The claim that HETA++ 'achieves the best performance in terms of accuracy' on 1DSfM is supported by an average AUC@3 of 40.6 versus GLOMAP's 39.9 and HETA's 38.4. The margin is under 1 point, and on several scenes (ALM, ELS, GDM, PIC) HETA++ is not the best global method. Since Table 1 reports a single run with no uncertainty, the difference may not be significant. Please provide per-scene win/loss summaries, multiple-run statistics, or a statistical test, and temper the wording to 'competitive accuracy' unless the margins are shown to be reproducible.
  3. [§4.3.1, Algorithm 1] The core robustness claim is conditional on a set of fixed hand-tuned hyperparameters: β=sin(1°)sin(5°), T_α=0.5°, T_β=1°, Nm=20, plus the 10° and 0.5° filtering thresholds, the Ra_0.75 reference, and S1=100, S2=50. The paper states these are used in all experiments but gives no sensitivity analysis or evidence that they transfer across heterogeneous scenes. Given that the spectral confidence proxy and the parallax thresholds directly determine which relative translations survive, the absence of a sensitivity sweep is a gap. Please report how the results vary with T_α/T_β and with the percentile anchor, or provide an adaptive selection rule.
  4. [§5, Fig. 12] The paper candidly admits that HETA++ fails on 'old computer' (ETH3D) and Temple of Heaven unless an external camera-triplet disambiguation stage (Manam and Govindu 2024) is added. These are precisely the structurally ambiguous scenes with repetitive texture that form an important failure class for global SfM. As written, the Abstract claims robustness on 'various real-world datasets,' but the benchmark averages exclude or downweight these collapse cases. The central claim should be scoped to non-ambiguous view graphs, or the disambiguation stage should be integrated into HETA++ and evaluated as part of the pipeline.
minor comments (4)
  1. [Eq. (17)] The notation R_i^T X_ki is ambiguous: if X_ki is meant to be the normalized image ray, define it explicitly near Eq. (3) and use a consistent symbol such as f_ki or X_hat_ki.
  2. [Table 7] The column header 'JO w/ S' should be 'JO w/ Selection' for readability; the text uses 'JO w/o Selection' but the table abbreviates inconsistently. Also, the first column group contains a run-on 'SHETA++' in the header row.
  3. [§4.2.3, Table 6] Runtime units are not stated in the Table 6 caption; please specify seconds for both T and the AUC columns.
  4. [§3.3.1] The 'rotation trick' is attributed to Jiang et al. (2013) but the reference list does not contain a separate entry for that term; please ensure all cited works are resolved.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central accuracy claims are benchmark comparisons against external ground truth; self-citations are incremental comparisons, not load-bearing reductions.

full rationale

HETA++ does not define any target quantity in terms of fitted parameters or self-citations. The final accuracy claims are evaluated on external datasets (1DSfM, KITTI, ETH3D, LaMAR) against ground-truth poses or the provided Bundler reconstruction, so the claim that it "outperforms state-of-the-art methods" is not forced by construction. The closest thing to a self-referential loop is the global-consistency filter in §3.3.3: Eq. (12) computes coarse positions from confidence-weighted relative translations, then discards translations whose direction disagrees with those coarse positions by more than 10°. This is a standard robust consistency filter, not a definitional reduction; an edge's influence on the coarse positions is diluted by the graph, and the downstream metrics are external. The confidence mapping in Eq. (11) is calibrated using the dataset's own Ra0.75, which is a robustness heuristic; even if its calibration were imperfect, it is not a fitted parameter that is later reported as a predicted quantity. The self-citation to HETA (Tao et al. 2024) is used as a baseline and for incremental ablations, not as an unverified uniqueness theorem or ansatz smuggled in by citation. The paper's own §5 admits failure on ambiguous scenes such as "old computer" and Temple of Heaven, and says an external camera-triplet disambiguation module is needed for those cases; this is a stated limitation and a correctness concern, not evidence that the paper's derivation reduces to its inputs. No circular step was found; the appropriate score is 0.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard multi-view geometry plus several domain heuristics; no new physical or conceptual entities are introduced beyond standard 3D points and cameras. The main burden is the set of hand-tuned thresholds and the spectral confidence heuristic in Eq. (11).

free parameters (8)
  • Minimum parallax angle Tα = 0.5°
    Feature matches with parallax below Tα are discarded before relative-translation re-estimation (Algorithm 1); affects which epipolar normals enter Eq. (10).
  • Coplanarity residual threshold Tβ = 1°
    Feature matches with normalized co-planarity residual above sin(Tβ) are rejected (Algorithm 1).
  • Cauchy loss width β = sin(1°)·sin(5°)
    Loss width in Eq. (10) for reweighted relative-translation estimation; value is reported without justification and appears to be a tuned constant.
  • Minimum inlier matches Nm = 20
    Image pairs with fewer than Nm inlier matches after local refinement are discarded from translation graph (Algorithm 1).
  • Rotation/translation consistency angle threshold = 10°
    Edge discarded if original relative rotation differs from globally consistent rotation by >10° (§3.1) or if non-tree relative translation differs from coarse position direction by >10° (§3.3.3).
  • Median parallax cutoff = 0.5°
    Non-tree edges whose median feature parallax is below 0.5° are discarded during global filtering (§3.3.3).
  • Eigenvalue-ratio confidence reference Ra0.75 = 75th percentile of per-dataset Ra
    Mapped to confidence 0.5 in Eq. (11); computed per dataset, making the filter adaptive to each scene and dataset-dependent.
  • Track-selection coverage parameters S1, S2 = S1=100, S2=50
    Control number and spatial balance of selected feature tracks in joint optimization (§3.5.1); chosen without cross-validation.
assumptions (6)
  • domain assumption Two-view relative poses and matches are reliable after COLMAP/PoseLib preprocessing
    Entire pipeline consumes view graph built from these; their errors propagate through all later stages (§3.1).
  • domain assumption Global rotations from Chatterjee-Govindu are sufficiently accurate
    Relative translations are re-estimated and filtered using these rotations; Eq. (8) assumes global rotations correct for normal construction.
  • domain assumption Epipolar-plane normal error formula tanγ=tanθ/sinα holds
    Used to justify parallax weighting in §3.3.1; assumes small feature-ray perturbations and ignores components along epipolar plane.
  • ad hoc to paper Eigenvalue ratio e2/e1 is a valid confidence proxy
    The spectral confidence score Eq. (11) is an ad hoc heuristic justified only by the KITTI-05 plot; no theory or independent validation establishes monotonicity.
  • domain assumption Camera-only L1 initialization Eq. (13) avoids collapse
    The paper acknowledges degenerate/weakly constrained cases but assumes subsequent feature-track constraints recover them (§3.4.1).
  • standard math Optimal λ in bilinear objective Eq. (6) is non-negative and saturates to 0 for θ≥π/2
    Used to derive the non-bilinear Eq. (7); standard Euclidean projection result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HETA++: Global Structure-from-Motion with Hybrid Explicit Translation Averaging." pith.science (2026). https://pith.science/paper/C7ZQ66PC

@misc{pith2026260715912,
  author       = {Pith},
  title        = {Pith review of: HETA++: Global Structure-from-Motion with Hybrid Explicit Translation Averaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7ZQ66PC}},
  note         = {Machine review of arXiv:2607.15912}
}
read the original abstract

Global Structure-from-Motion (SfM) offers advantages over incremental methods in terms of efficiency and error distribution. However, the task of translation averaging remains challenging. Many existing methods rely solely on relative translations or feature tracks, which either degrade under collinear camera motion or are susceptible to outliers. In this paper, we propose a novel hybrid explicit translation averaging framework that incorporates both relative translations and feature tracks. Specifically, we first refine the relative translations using global camera rotations and remove globally inconsistent relative translations. Next, we employ convex distance-based objective functions to estimate the initial camera positions and 3D points, followed by refinement using a non-bilinear angle-based objective function. Furthermore, since camera rotations are fixed during translation averaging, inaccurate camera rotations can severely limit the accuracy of camera positions. To address this issue, we then robustly refine both camera rotations and camera positions with selected feature tracks through bounded angle-based refinement and subsequent reprojection-based bundle adjustment. In this step, feature tracks are selected to maintain a balanced spatial distribution and improve optimization efficiency. Finally, we perform a complete bundle adjustment using all reliable feature tracks to refine the camera parameters and 3D points. Extensive experiments on various sequential and unordered real-world datasets demonstrate the superior accuracy, robustness, and scalability of our approach, outperforming state-of-the-art methods in both accuracy and computational efficiency.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

111 extracted references · 4 linked inside Pith

  1. [1]

    Campbell, S. L. and Gear, C. W. The index of general nonlinear D A E S. Numer. M ath. 1995

    1995

  2. [2]

    Slifka, M. K. and Whitton, J. L. Clinical implications of dysregulated cytokine production. J. M ol. M ed. 2000. doi:10.1007/s001090000086

    doi:10.1007/s001090000086 2000

  3. [3]

    Quasimonotonicity, regularity and duality for nonlinear systems of partial differential equations

    Hamburger, C. Quasimonotonicity, regularity and duality for nonlinear systems of partial differential equations. Ann. Mat. Pura. Appl. 1995

    1995

  4. [4]

    Geddes, K. O. and Czapor, S. R. and Labahn, G. Algorithms for C omputer A lgebra. 1992

    1992

  5. [5]

    Software engineering---from auxiliary to key technologies

    Broy, M. Software engineering---from auxiliary to key technologies. Software Pioneers. 1992

    1992

  6. [6]

    Conductive P olymers. 1981

    1981

  7. [7]

    Smith, S. E. Neuromuscular blocking drugs in man. Neuromuscular junction. H andbook of experimental pharmacology. 1976

    1976

  8. [8]

    Chung, S. T. and Morris, R. L. Isolation and characterization of plasmid deoxyribonucleic acid from Streptomyces fradiae. 1978

    1978

Show all 111 references
  1. [9]

    and AghaKouchak, A

    Hao, Z. and AghaKouchak, A. and Nakhjiri, N. and Farahmand, A. Global integrated drought monitoring and prediction system (GIDMaPS) data sets. 2014

    2014

  2. [10]

    Babichev, S. A. and Ries, J. and Lvovsky, A. I. Quantum scissors: teleportation of single-mode optical states by means of a nonlocal single photon. 2002

    2002

  3. [11]

    and Buchalla, G

    Beneke, M. and Buchalla, G. and Dunietz, I. Mixing induced CP asymmetries in inclusive B decays. Phys. L ett. 1997. arXiv:0707.3168

    1997 arXiv

  4. [12]

    deep SIP : deep learning of S upernova I a P arameters

    Stahl, B. deep SIP : deep learning of S upernova I a P arameters. 2020. ascl:2006.023

    2020

  5. [13]

    Abbott, T. M. C. and others. Dark Energy Survey Year 1 Results: Constraints on Extended Cosmological Models from Galaxy Clustering and Weak Lensing. Phys. Rev. D. 2019. doi:10.1103/PhysRevD.99.123505. arXiv:1810.02499

    2019 arXiv

  6. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Revisiting Global Translation Estimation with Feature Tracks , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  7. [15]

    European Conference on Computer Vision (ECCV) , pages=

    Pan, Linfei and Barath, Daniel and Pollefeys, Marc and Sch\". European Conference on Computer Vision (ECCV) , pages=. 2025 , organization=

    2025

  8. [16]

    Advances in Neural Information Processing Systems , volume=

    Sensitivity in translation averaging , author=. Advances in Neural Information Processing Systems , volume=

  9. [17]

    Ground Camera Images and UAV 3D Model Registration for Outdoor Augmented Reality , year=

    Liu, Weiquan and Wang, Cheng and Zang, Yu and Lai, Shang-Hong and Weng, Dongdong and Bian, Xuesheng and Lin, Xiuhong and Shen, Xuelun and Li, Jonathan , booktitle=. Ground Camera Images and UAV 3D Model Registration for Outdoor Augmented Reality , year=

  10. [18]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    From coarse to fine: Robust hierarchical localization at large scale , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

  11. [19]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Back to the feature: Learning robust camera localization from pixels to pose , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

  12. [20]

    IEEE International Conference on Computer Vision , pages=

    On the limits of pseudo ground truth in visual camera re-localisation , author=. IEEE International Conference on Computer Vision , pages=

  13. [21]

    Communications of the ACM , volume=

    Nerf: Representing scenes as neural radiance fields for view synthesis , author=. Communications of the ACM , volume=. 2021 , publisher=

    2021

  14. [22]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Block-nerf: Scalable large scene neural view synthesis , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

  15. [23]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Grid-guided Neural Radiance Fields for Large Urban Scenes , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

  16. [24]

    Communications of the ACM , volume=

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography , author=. Communications of the ACM , volume=. 1981 , publisher=

    1981

  17. [25]

    IEEE International Conference on Computer Vision , pages=

    Optimizing the viewing graph for structure-from-motion , author=. IEEE International Conference on Computer Vision , pages=

  18. [26]

    IEEE Conference on Computer Vision and Pattern Recognition , pages=

    Efficient initial pose-graph generation for global sfm , author=. IEEE Conference on Computer Vision and Pattern Recognition , pages=

  19. [27]

    IEEE International Conference on Computer Vision , pages=

    Viewing graph solvability via cycle consistency , author=. IEEE International Conference on Computer Vision , pages=

  20. [28]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

    Lie-algebraic averaging for globally consistent motion estimation , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

  21. [29]

    IEEE International Conference on Computer Vision (ICCV) , pages=

    Efficient and robust large-scale rotation averaging , author=. IEEE International Conference on Computer Vision (ICCV) , pages=

  22. [30]

    International Journal of Computer Vision , volume=

    Incremental rotation averaging , author=. International Journal of Computer Vision , volume=. 2021 , publisher=

    2021

  23. [31]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    End-to-end rotation averaging with multi-source propagation , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  24. [32]

    IEEE International Conference on 3D Vision (3DV) , pages=

    It is all in the weights: Robust rotation averaging revisited , author=. IEEE International Conference on 3D Vision (3DV) , pages=

  25. [33]

    Lee, Seong Hun and Civera, Javier , booktitle=

  26. [34]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Revisiting Rotation Averaging: Uncertainties and Robust Losses , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  27. [35]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

    Combining two-view constraints for motion estimation , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

  28. [36]

    IEEE International Conference on Computer Vision (ICCV) , pages=

    Global fusion of relative motions for robust, accurate and scalable structure from motion , author=. IEEE International Conference on Computer Vision (ICCV) , pages=

  29. [37]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Robust camera location estimation by convex programming , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  30. [38]

    European Conference on Computer Vision (ECCV) , pages=

    Shapefit and shapekick for robust, scalable structure from motion , author=. European Conference on Computer Vision (ECCV) , pages=. 2016 , organization=

    2016

  31. [39]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Baseline desensitizing in translation averaging , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  32. [40]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Very large-scale global sfm by distributed motion averaging , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  33. [41]

    European Conference on Computer Vision (ECCV) , pages=

    Correspondence Reweighted Translation Averaging , author=. European Conference on Computer Vision (ECCV) , pages=. 2022 , organization=

    2022

  34. [42]

    IEEE International Conference on Computer Vision (ICCV) , pages=

    A global linear method for camera pose registration , author=. IEEE International Conference on Computer Vision (ICCV) , pages=

  35. [43]

    Linear Global Translation Estimation with Feature Tracks

    Zhaopeng Cui and Nianjuan Jiang and Chengzhou Tang and Ping Tan , booktitle=. Linear Global Translation Estimation with Feature Tracks. , volume=

  36. [44]

    IEEE International Conference on Computer Vision (ICCV) , pages=

    Global structure-from-motion by similarity averaging , author=. IEEE International Conference on Computer Vision (ICCV) , pages=

  37. [45]

    IEEE International Conference on Pattern Recognition (ICPR) , pages=

    Robust global translation averaging with feature tracks , author=. IEEE International Conference on Pattern Recognition (ICPR) , pages=

  38. [46]

    IEEE Robotics and Automation Letters , volume=

    Robust global structure from motion pipeline with parallax on manifold bundle adjustment and initialization , author=. IEEE Robotics and Automation Letters , volume=. 2019 , publisher=

    2019

  39. [47]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    A pose-only solution to visual reconstruction and navigation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2021 , publisher=

    2021

  40. [48]

    SfM with

    Crandall, David J and Owens, Andrew and Snavely, Noah and Huttenlocher, Daniel P , journal=. SfM with. 2012 , publisher=

    2012

  41. [49]

    European Conference on Computer Vision , pages=

    Robust global translations with 1dsfm , author=. European Conference on Computer Vision , pages=. 2014 , organization=

    2014

  42. [50]

    AAAI Conference on Artificial Intelligence , volume=

    MMA: Multi-Camera Based Global Motion Averaging , author=. AAAI Conference on Artificial Intelligence , volume=

  43. [51]

    Reproducible Research in Pattern Recognition: First International Workshop , pages=

    Openmvg: Open multiple view geometry , author=. Reproducible Research in Pattern Recognition: First International Workshop , pages=. 2017 , organization=

    2017

  44. [52]

    Chris Sweeney , title =

  45. [53]

    2003 , publisher=

    Multiple view geometry in computer vision , author=. 2003 , publisher=

    2003

  46. [54]

    Structure-from-Motion Revisited , booktitle=

    Sch\". Structure-from-Motion Revisited , booktitle=

  47. [55]

    Arandjelovic, Relja and Gronat, Petr and Torii, Akihiko and Pajdla, Tomas and Sivic, Josef , booktitle=

  48. [56]

    The International Journal of Robotics Research , volume=

    Vision meets robotics: The kitti dataset , author=. The International Journal of Robotics Research , volume=. 2013 , publisher=

    2013

  49. [57]

    ACM siggraph papers , pages=

    Photo tourism: exploring photo collections in 3D , author=. ACM siggraph papers , pages=

  50. [58]

    European Conference on Visual Media Production , pages=

    Unordered feature tracking made fast and easy , author=. European Conference on Visual Media Production , pages=

  51. [59]

    2023 , month =

    Agarwal, Sameer and Mierle, Keir and The Ceres Solver Team , title =. 2023 , month =

    2023

  52. [60]

    Foundations and Trends

    Distributed optimization and statistical learning via the alternating direction method of multipliers , author=. Foundations and Trends. 2011 , publisher=

    2011

  53. [61]

    IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Design and Evaluation of an Augmented Reality Head-Mounted Display User Interface for Controlling Legged Manipulators , author=. IEEE International Conference on Robotics and Automation (ICRA) , pages=

  54. [62]

    IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

    EPAR: An Efficient and Privacy-Aware Augmented Reality Framework for Indoor Location-Based Services , author=. IEEE/RSJ International Conference on Intelligent Robots and Systems , pages=

  55. [63]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Hybrid rotation averaging: A fast and robust rotation averaging approach , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  56. [64]

    IEEE Transactions on Visualization and Computer Graphics , volume=

    Instant visual odometry initialization for mobile AR , author=. IEEE Transactions on Visualization and Computer Graphics , volume=

  57. [65]

    IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

    Viewing Graph Solvability in Practice , author=. IEEE/CVF International Conference on Computer Vision (ICCV) , pages=

  58. [66]

    IEEE International Conference on 3D Vision (3DV) , pages=

    Reducing drift in structure from motion using extended features , author=. IEEE International Conference on 3D Vision (3DV) , pages=

  59. [67]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Revisiting the P3P Problem , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  60. [68]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Complete solution classification for the perspective-three-point problem , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  61. [69]

    Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms , pages=

    Bundle adjustment—a modern synthesis , author=. Vision Algorithms: Theory and Practice: International Workshop on Vision Algorithms , pages=. 2000 , organization=

    2000

  62. [70]

    2022 , organization=

    Ren, Jie and Liang, Wenteng and Yan, Ran and Mai, Luo and Liu, Shiwen and Liu, Xiao , booktitle=. 2022 , organization=

    2022

  63. [71]

    Communications in Statistics-theory and Methods , volume=

    Robust regression using iteratively reweighted least-squares , author=. Communications in Statistics-theory and Methods , volume=. 1977 , publisher=

    1977

  64. [72]

    International Journal of Computer Vision , volume=

    Distinctive image features from scale-invariant keypoints , author=. International Journal of Computer Vision , volume=. 2004 , publisher=

    2004

  65. [73]

    , journal=

    Umeyama, S. , journal=. Least-squares estimation of transformation parameters between two point patterns , year=

  66. [74]

    Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003

    Locally optimized RANSAC , author=. Pattern Recognition: 25th DAGM Symposium, Magdeburg, Germany, September 10-12, 2003. Proceedings 25 , pages=. 2003 , organization=

    2003

  67. [75]

    Tyrell Rockafellar, Convex Analysis (Princeton University Press, 1970), xviii+ 451 pp

    R. Tyrell Rockafellar, Convex Analysis (Princeton University Press, 1970), xviii+ 451 pp. , author=. Proceedings of the Edinburgh Mathematical Society , volume=. 1973 , publisher=

    1970

  68. [76]

    2004 , publisher=

    Convex optimization , author=. 2004 , publisher=

    2004

  69. [77]

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=

    Removing outliers using the L_ norm , author=. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , volume=. 2006 , organization=

    2006

  70. [78]

    International journal of computer vision , volume=

    Linear multi view reconstruction and camera recovery using a reference plane , author=. International journal of computer vision , volume=. 2002 , publisher=

    2002

  71. [79]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Quasiconvex optimization for robust geometric reconstruction , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2007 , publisher=

    2007

  72. [80]

    Recovering camera motion using

    Sim, Kristy and Hartley, Richard , booktitle=. Recovering camera motion using

  73. [81]

    2008 , publisher=

    Kahl, Fredrik and Hartley, Richard , journal=. 2008 , publisher=

    2008

  74. [82]

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Outlier removal using duality , author=. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  75. [83]

    Computational Optimization and Applications , volume=

    Non-stationary Douglas--Rachford and alternating direction method of multipliers: adaptive step-sizes and convergence , author=. Computational Optimization and Applications , volume=. 2019 , publisher=

    2019

  76. [84]

    Quarterly of applied mathematics , volume=

    A method for the solution of certain non-linear problems in least squares , author=. Quarterly of applied mathematics , volume=

  77. [85]

    Moulon, Pierre and Monasse, Pascal and Perrot, Romuald and Marlet, Renaud , booktitle=. Open. 2016 , organization=

    2016

  78. [86]

    Schönberger and Pablo Speciale and Lukas Gruber and Viktor Larsson and Ondrej Miksik and Marc Pollefeys , title =

    Paul-Edouard Sarlin and Mihai Dusmanu and Johannes L. Schönberger and Pablo Speciale and Lukas Gruber and Viktor Larsson and Ondrej Miksik and Marc Pollefeys , title =. European Conference on Computer Vision (ECCV) , pages=. 2022 , organization=

    2022

  79. [87]

    and Galliani, Silvano and Sattler, Torsten and Schindler, Konrad and Pollefeys, Marc and Geiger, Andreas , booktitle=

    Schöps, Thomas and Schönberger, Johannes L. and Galliani, Silvano and Sattler, Torsten and Schindler, Konrad and Pollefeys, Marc and Geiger, Andreas , booktitle=. A Multi-view Stereo Benchmark with High-Resolution Images and Multi-camera Videos , year=

  80. [88]

    International Conference on Computer Vision (ICCV) , pages=

    Philipp Lindenberger and Paul-Edouard Sarlin and Viktor Larsson and Marc Pollefeys , title =. International Conference on Computer Vision (ICCV) , pages=

  81. [89]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Detector-Free Structure from Motion , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  82. [90]

    International Journal of Computer Vision , volume=

    Rotation averaging , author=. International Journal of Computer Vision , volume=. 2013 , publisher=

    2013

  83. [91]

    2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission , pages=

    Global motion estimation from point matches , author=. 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission , pages=. 2012 , organization=

    2012

  84. [92]

    The International Journal of Robotics Research , volume=

    SE-Sync: A certifiably correct algorithm for synchronization over the special Euclidean group , author=. The International Journal of Robotics Research , volume=. 2019 , publisher=

    2019

  85. [93]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Rotation averaging and strong duality , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  86. [94]

    Shonan Rotation Averaging: Global Optimality by Surfing

    Dellaert, Frank and Rosen, David M and Wu, Jing and Mahony, Robert and Carlone, Luca , booktitle=. Shonan Rotation Averaging: Global Optimality by Surfing. 2020 , organization=

    2020

  87. [95]

    and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P

    Barron, Jonathan T. and Mildenhall, Ben and Verbin, Dor and Srinivasan, Pratul P. and Hedman, Peter , booktitle=. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields , year=

  88. [96]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =. 2023 , url =

    2023

  89. [97]

    2020 , booktitle=

    NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis , author=. 2020 , booktitle=

    2020

  90. [98]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation , author =. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  91. [99]

    Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=

    Sch\". Pixelwise View Selection for Unstructured Multi-View Stereo , booktitle=. 2016 , organization=

    2016

  92. [100]

    Mur-Artal, Raúl and Montiel, J. M. M. and Tardós, Juan D. , journal=. doi:10.1109/TRO.2015.2463671 , year=

    2015

  93. [101]

    and Montiel, José M

    Campos, Carlos and Elvira, Richard and Gómez, Juan J. and Montiel, José M. M. and Tardós, Juan D. , journal=

  94. [102]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Leveraging Camera Triplets for Efficient and Accurate Structure-from-Motion , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  95. [103]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Robust relative rotation averaging , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2017 , publisher=

    2017

  96. [104]

    IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Distinguishing the indistinguishable: Exploring structural ambiguities via geodesic context , author=. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  97. [105]

    Viktor Larsson and contributors , URL =

  98. [106]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Dust3r: Geometric 3d vision made easy , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  99. [107]

    European conference on computer vision , pages=

    Grounding image matching in 3d with mast3r , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  100. [108]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Vggt: Visual geometry grounded transformer , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  101. [109]

    arXiv preprint arXiv:2507.13347 , year =

    ^3 : Permutation-Equivariant Visual Geometry Learning , author =. arXiv preprint arXiv:2507.13347 , year =

    arXiv

  102. [110]

    2026 International Conference on 3D Vision (3DV) , pages=

    Mapanything: Universal feed-forward metric 3d reconstruction , author=. 2026 International Conference on 3D Vision (3DV) , pages=. 2026 , organization=

    2026

  103. [111]

    arXiv preprint arXiv:2511.10647 , year=

    Depth anything 3: Recovering the visual space from any views , author=. arXiv preprint arXiv:2511.10647 , year=

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.