Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Princeton365: A Diverse Dataset with Accurate Camera Pose

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Princeton365 claims millimeter-accurate camera-pose ground truth for 365 diverse videos, collected with calibration boards and a 360-degree camera, plus a scale-aware SLAM metric.

desk verdict A genuinely useful dataset with an externally validated core, but the user-view poses that methods actually consume rest on a relative-pose calibration checked only for self-consistency — and the authors already have the data to close that gap. read the letter →

arxiv 2506.09035 v2 pith:47RTSV3K submitted 2025-06-10 cs.CV

Karhan Kayan , Stamatis Alexandropoulos , Rishabh Jain , Yiming Zuo , Erich Liang , Jia Deng This is my paper
classification cs.CV MSC 68T45
keywords SLAMbenchmarkground-truthcameraposecalibrationboards360-degreeBundlePnPInducedOpticalFlowmetricnovelviewsynthesis6-DoFtrajectory
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

Princeton365 is a dataset of 365 videos — 40 long indoor, 40 long outdoor, and 285 object-scanning sequences — whose ground-truth camera poses the authors claim are millimeter-accurate. The accuracy comes from a collection pipeline built around calibration boards and a 360-degree camera: the camera renders both a user view that sees the scene and a ground-truth view that sees the boards, and an optimization method called Bundle PnP recovers the camera trajectory from detected board markers. External validation against a motion-capture system reports an average trajectory error of 2.88 mm across nine re-enacted sequences, versus 1.14 cm for COLMAP. The paper also introduces the Induced Optical Flow metric, which expresses pose error as the pixel motion it would cause given the scene's depth distribution, allowing SLAM methods to be compared across scenes of very different scale. If the claims hold, researchers gain a benchmark that is simultaneously diverse, full six-degree-of-freedom, and accurate enough to trust as ground truth.

What carries the argument

The load-bearing mechanism is Bundle PnP, a nonlinear least-squares optimization that minimizes the reprojection error of detected calibration markers jointly over the global poses of the boards and the camera, initialized from a pose graph of board positions (built with median-based edge selection, pose-graph optimization, and an optional coplanar snapping step) and multi-board PnP. Its companion, Bundle Rig PnP, extends the same objective to solve for the fixed relative pose between the user view and the ground-truth view — two non-overlapping renderings of the same 360-degree frame — so the millimeter accuracy measured on the board-observing view transfers to the view that SLAM and novel-view-synthesis methods actually consume. The secondary mechanism is the Induced Optical Flow (IOF) metric, which re-projects pixels from the ground-truth pose to the estimated pose, integrates over a depth distribution fitted to each sequence's stereo depth, and reports the area under the flow curve (Flow AUC) as a scale-aware error score.

What would settle it

Film a sequence inside a motion-capture room with reflective markers visible both to the MoCap system and to the rendered user view while calibration boards remain visible to the ground-truth view; if the user-view trajectory produced by the full pipeline, including the Bundle Rig PnP relative pose, disagrees with the MoCap user-view trajectory by substantially more than the reported 2-4 mm ATE, the relative-pose calibration carries a bias that the internal consistency test cannot detect. A cheaper variant attaches an independent laser-tracked target to the rig and compares its user-view trajectory to the dataset's poses.

Watch

Extended reading notes

Core claim

The central claim is that a calibration-board and 360-camera setup can produce large-scale, diverse SLAM ground truth without the accuracy-versus-diversity tradeoff that has forced earlier benchmarks to choose between millimeter-accurate but scene-restricted capture and scalable but coarser methods. The paper argues that rendering two views from one 360-degree camera — a user view that sees the scene and a ground-truth view that sees only calibration boards — and recovering poses from 2D-3D marker correspondences through a pose graph, multi-board PnP, and a Bundle PnP refinement achieves both goals, with a Bundle Rig PnP step that calibrates the fixed relative pose between the two views so the user-view trajectory inherits the board-view accuracy. The reported numbers are an average ATE of 2.88 mm against an external motion-capture system across nine sequences, beating COLMAP on every one, and a dataset of 365 sequences, over two million frames, and an estimated 26 km of travel. A second claim is that the new Induced Optical Flow metric, by weighting pose error against scene depth, makes error scores comparable across trajectories of different scale, which the paper argues trajectory-error metrics such as ATE cannot do.

Load-bearing premise

The accuracy of every user-view pose rests on the assumed fixed spatial relationship between the user view and the board-observing ground-truth view, which the authors calibrate but never verify with an external measurement; a systematic bias in that relationship would shift all user-view poses while passing the paper's internal consistency check.

Editorial extensions

If this is right

  • SLAM and visual-odometry methods can be ranked across indoor, outdoor, and object-scanning trajectories of different scale using the Induced Optical Flow / Flow AUC metric, where trajectory-error numbers such as ATE are not comparable.
  • Accurate six-degree-of-freedom ground truth is no longer confined to one or two rooms: the board-and-360-camera pipeline claims to reach any indoor or outdoor location with roughly five minutes of setup.
  • Novel-view-synthesis evaluation can include fully non-Lambertian scenes and 360-degree camera trajectories without depending on COLMAP poses, which fail on five of the paper's six benchmark scenes while the proposed pipeline succeeds on all.
  • The dataset's scale — 365 sequences, over two million frames, roughly 56 percent with ground-truth pose — gives SLAM researchers a substantially larger and more diverse testbed for locating when and why methods fail.

Reading between the lines

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

  • Because the user view is a re-render of a 360-degree capture, the same recording could be re-rendered with different framings or fields of view after the fact, each inheriting the same ground-truth trajectory — effectively multiplying the number of evaluable sequences beyond 365.
  • The Induced Optical Flow metric's dependence on a fitted depth distribution suggests it could be weighted by task-relevant depth bands (for instance near-field for augmented reality) without any change to the capture pipeline, giving users a tunable notion of error.
  • The pose-graph plus Bundle PnP machinery only needs detectable fiducials, so other groups could port the pipeline to their own markers and scenes and produce comparably accurate ground truth, making mm-accurate benchmarks reproducible beyond this one dataset.
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. The paper introduces Princeton365, a dataset of 365 videos (indoor, outdoor, and object-scanning) with camera poses obtained by a scalable pipeline that combines an Insta360 360-degree camera, calibration boards, and a Bundle PnP optimization. The authors claim millimeter-accuracy poses, validated externally against Vicon MoCap on 9 re-enacted sequences (average ATE 2.88 mm vs COLMAP's 1.14 cm). They also propose a new scale-aware SLAM metric, Induced Optical Flow (IOF), with a Flow AUC variant, and an NVS benchmark including strongly non-Lambertian scenes. The central claims are that the dataset provides accurate ground-truth poses at a scale and diversity not available in existing SLAM benchmarks, and that IOF enables cross-scene comparison. The paper includes extensive ablations of the pose pipeline and an evaluation of several SLAM/VO methods.

Significance. If the accuracy claims hold, Princeton365 would be a valuable community resource: it combines the diversity of large outdoor datasets with the millimeter-level accuracy usually found only in small MoCap-constrained datasets, and it provides 6-DoF motion. The external validation against Vicon Vantage V16 (Table 5) and the thorough ablation study (Tables 7-8) are strong positive features, as is the explicit attempt to validate the pipeline with an independent measurement system. The IOF metric addresses a real limitation of ATE in comparing across scene scales, and the NVS benchmark fills a genuine gap by including non-Lambertian scenes with full 360-degree trajectories. The main reservations concern the accuracy of the user-view poses, which are what SLAM and NVS methods actually consume, and the definitional basis of the IOF metric. These issues are addressable but currently leave the central 'mm-accurate user-view pose' claim insufficiently supported.

major comments (3)
  1. [Sec. 4.5, Sec. F, Sec. G, Table 10] The user-view trajectory, which is the trajectory used for SLAM/NVS evaluation, is obtained by composing the externally validated ground-truth-view trajectory with the fixed relative pose T^u_gt estimated by Bundle Rig PnP. The only validation reported for T^u_gt is the repeatability experiment in Table 10 (mean rotation difference 0.0705 deg, translation difference 0.891 mm). A common-mode bias in the fisheye intrinsics, the 360-degree stitching center, or the rig model would be reproduced in both trajectories and would pass this consistency check while shifting all user-view poses. The paper's Sec. G states that an external measurement of the relative pose is not possible, yet it describes a MoCap-based procedure that yields a relative pose through naive PnP; that procedure is used only to demonstrate high variance of naive PnP, and the mean MoCap-derived relative pose is never compared with the Bundle Rig PnP estimate. Please report that comparison, or explicitly restrict the mm-accuracy claim to the ground-truth-view trajectory and state the user-view accuracy as unvalidated.
  2. [Sec. H, Table 3] The generalization from the 9 MoCap-validated indoor sequences to all 365 sequences rests on the observation that Bundle PnP reprojection errors are similar indoors (0.7373 px) and outdoors (0.7394 px), with p=0.832. Reprojection error is a necessary but not sufficient proxy for absolute pose accuracy: systematic errors in the board pose graph, distortion calibration, or the relative pose calibration can yield low reprojection error while shifting absolute poses. Furthermore, outdoor sequences have only 18.8% posed frames (Table 3), so the validated conditions (short, board-dense MoCap-room re-enactments) differ substantially from the long outdoor portions of the benchmark. Please provide additional evidence for the outdoor and long-sequence cases, or soften the claim that the MoCap validation extends to all sequences.
  3. [Sec. 4.4, Eq. (6), Sec. C, Sec. I] The IOF expectation samples depth d from the sequence-level marginal distribution p(d) independently of the pixel (u,v). In Eq. (6), every pixel integrates ||flow(t,d,u,v)||^2 against the same p(d), so for a given pixel the metric averages over depths that are generally not the actual depth at that pixel. The result is not the optical flow induced by the pose error for the actual scene, but an average over a synthetic depth distribution. Since p(d) is itself obtained by BIC-selected mixtures of Gaussians/Gammas (Sec. C), IOF values may be sensitive to this modeling choice. Please either define the expectation with per-pixel conditional depth distributions p(d|u,v), or provide a comparison against a Monte Carlo evaluation using the actual ZED depth maps; if the marginal formulation is intentional, it should be justified explicitly as the desired semantics of the metric.
minor comments (5)
  1. [Table 3] The column 'Posed dist. covered' gives values such as 2,585 and 922 that appear to be meters, but the unit is not stated in the table header; please add the unit.
  2. [Sec. 5.1, Table 5] There are several typos: 'outperfoms' should be 'outperforms' in Sec. 5.1, and the column headers 'A TE' and 'A ccuracy' in Tables 5 and 1 should be 'ATE' and 'Accuracy'.
  3. [Eq. (5)] The notation t ∼ U(T) overloads T: T is used both for the number of frames and for the trajectory set. Please use a distinct symbol for the frame index range.
  4. [Sec. 4.3.1] In 'Median Edge Selection,' the text says the pose corresponding to the median distance is selected; please clarify whether the median is taken over the scalar board-to-board distances only, or whether a median in SE(3) is computed, and specify the exact criterion.
  5. [Sec. 4.3.1] The 'Snap to Plane' description ('We adjust the x and y axes to preserve the orthogonality of the coordinate system') is vague; please specify the orthogonalization procedure, e.g., Gram-Schmidt on the adjusted axes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the mm-accuracy claim is anchored to an independent Vicon MoCap system, and the IOF metric is a stated definition rather than a fitted prediction.

full rationale

The central ground-truth claim is validated externally: Tab. 5 compares the pipeline's GT-view trajectory to a Vicon Vantage V16 MoCap system (0.201 mm RMSE accuracy), and the ablations in Tabs. 7 and 8 test pipeline variants against that same external standard. User-view poses are obtained by composing the GT-view trajectory with the fixed relative pose T^u_gt estimated by Bundle Rig PnP (Sec. 4.5, Eq. 15). The paper explicitly says in Sec. G that no external measurement of this relative pose is available and therefore relies on internal validation; Sec. F/Tab. 10 shows only that two trajectories yield consistent relative poses. This is an evidentiary limitation: a common-mode bias would survive the consistency check, so the absolute user-view mm-accuracy claim is less strongly supported than the GT-view claim. However, this is not circular in the definitional sense: T^u_gt is not defined in terms of the user-view poses it is used to produce, and neither the GT-view trajectory nor the relative pose is fitted to the SLAM or NVS outputs it is used to evaluate. The IOF metric (Sec. 4.4, Eqs. 5-6) is a new evaluation definition with stated modeling choices: mixture components selected by BIC, numerical integration over fitted depth distributions, and a 100-px AUC threshold. These are not derived from the method rankings the metric produces, nor are they fitted to make those rankings come out a particular way. No load-bearing self-citation or imported uniqueness theorem appears; citations to DPVO, DPV-SLAM, and DROID-SLAM are evaluations of external systems, not justification of the ground-truth construction. One minor validation-leakage caveat is that the optimal pipeline configuration is selected on the same MoCap sequences later summarized in Tab. 5, which is a statistical generalizability concern rather than a circular derivation. Overall, the derivation chain is self-contained and anchored to an external measurement system, so there is no significant circularity.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on marker detection fidelity, board rigidity and known geometry, optional coplanarity, per-view intrinsic models of rendered 360 views, and transfer of MoCap-room validation to the field. The IOF metric depends on a per-sequence fitted depth distribution and hand-chosen integration bounds. No invented physical entities are introduced; Bundle PnP, Bundle Rig PnP, and IOF are algorithms or metrics, not entities.

free parameters (4)
  • Flow AUC threshold range = 0 to 100 px
    Hand-chosen upper bound of 100 px for the area-under-curve of induced optical flow; affects the numerical scale of the headline metric but not relative method rankings. Sec. 4.4.
  • Depth mixture distribution (Gaussian/Gamma, 1-8 components) = BIC-selected per sequence
    Per-sequence fitted model of ZED stereo depth used to integrate IOF (Eq. 6). The model family and component range are author choices; selection itself is by BIC. Sec. C.
  • Depth integration bounds = 4 standard deviations around component means
    Truncation of the depth integral in the IOF computation (Sec. I); arbitrary but explicitly stated.
  • Average walking speed = 1.4 m/s
    Used to estimate total distance covered (26 km) for unposed frames; stated as an assumption in Sec. N.3.
assumptions (6)
  • domain assumption AprilTag/ChArUco/GridBoard markers are detected correctly and their local 3D coordinates are exactly known.
    The entire pipeline consumes 2D-3D marker correspondences (Sec. 4.3.1); any detection bias or printing distortion flows directly into the poses.
  • standard math Detected board correspondences provide reliable PnP initialization, and pose graph optimization plus Levenberg-Marquardt converge to the correct optimum.
    The pipeline relies on PnP, g2o, and Ceres convergence (Secs. 4.3.1, 4.3.3); no global optimality certificate is supplied.
  • domain assumption Board coplanarity holds when the snapping prior is applied.
    Snapping is enabled only when boards are coplanar, verified with a bubble level in validation (Sec. A.1); in the field this is an assumption.
  • ad hoc to paper The rendered user view and ground-truth view are accurately modeled by separately calibrated intrinsics with per-view distortion.
    Relative pose calibration between views and reprojection rely on the per-view intrinsics model (Sec. 4.2); an unmodeled rendering artifact would bias all poses.
  • ad hoc to paper The walking-speed estimate of 1.4 m/s approximates unposed motion.
    Used to report 26 km estimated total distance (Sec. N.3); clearly stated, low stakes for the central claims.
  • domain assumption MoCap-room validation transfers to real collections.
    Nine re-enacted sequences validate the pipeline indoors (Sec. 5.1); the outdoor transfer rests on reprojection-error equivalence (Sec. H).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Princeton365: A Diverse Dataset with Accurate Camera Pose." pith.science (2026). https://pith.science/paper/47RTSV3K

@misc{pith2026250609035,
  author       = {Pith},
  title        = {Pith review of: Princeton365: A Diverse Dataset with Accurate Camera Pose},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47RTSV3K}},
  note         = {Machine review of arXiv:2506.09035}
}
read the original abstract

We introduce Princeton365, a large-scale diverse dataset of 365 videos with accurate camera pose. Our dataset bridges the gap between accuracy and data diversity in current SLAM benchmarks by introducing a novel ground truth collection framework that leverages calibration boards and a 360-camera. We collect indoor, outdoor, and object scanning videos with synchronized monocular and stereo RGB video outputs as well as IMU. We further propose a new scene scale-aware evaluation metric for SLAM based on the optical flow induced by the camera pose estimation error. In contrast to the current metrics, our new metric allows for comparison between the performance of SLAM methods across scenes as opposed to existing metrics such as Average Trajectory Error (ATE), allowing researchers to analyze the failure modes of their methods. We also propose a challenging Novel View Synthesis benchmark that covers cases not covered by current NVS benchmarks, such as fully non-Lambertian scenes with 360-degree camera trajectories. Please visit https://princeton365.cs.princeton.edu for the dataset, code, videos, and submission.

Figures

Figures reproduced from arXiv: 2506.09035 by the authors.

Figure 1
Figure 1. Randomly picked frames from Princeton365, consisting of three main categories: object scanning, long indoor videos and long [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Top row: Non-Lambertian and dynamic scenes. Top [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. 6 scenes from our NVS benchmark. Our benchmark [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: An illustration of our camera rig. We render a user [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Our ground-truth method. We place calibration boards in a new environment and film using our 360-degree camera and stereo [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: On the Window-Night scene, where COLMAP success [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Comparison of our ground truth trajectory with [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 9
Figure 9. Figure 9: Two types of fiducial calibration boards used in our [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 8
Figure 8. Figure 8: Comparison of our ground truth trajectory with MoCap. [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 10
Figure 10. Figure 10: Histograms of both distance and angle variation ob [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: An illustration of the parametric model we fit to the depth data. The figure shows fitted distributions, BIC scores for model [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: X and Y comparison of six different trajectories as measured by MoCap and the 360 camera. [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 40 canonical work pages

  1. [1]

    Siddharth Agarwal, Ankit V ora, Gaurav Pandey, Wayne Williams, Helen Kourous, and James R. McBride. Ford multi-av seasonal dataset.CoRR, abs/2003.07969, 2020. 3

    work page Pith review arXiv 2003

  2. [2]

    Ceres Solver, 2023

    Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres Solver, 2023. 7

    work page 2023

  3. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 3

    2022

  4. [4]

    The m ´alaga urban dataset: High-rate stereo and lidar in a realistic urban scenario.International Journal of Robotics Research, 33:207–214, 2014

    Jose Luis Blanco, Francisco Moreno, and Javier Gonz ´alez- Jim´enez. The m ´alaga urban dataset: High-rate stereo and lidar in a realistic urban scenario.International Journal of Robotics Research, 33:207–214, 2014. 1, 3

    work page 2014

  5. [5]

    Understanding slerp, then not using it.Game Developer Magazine, 2004

    Jonathan Blow. Understanding slerp, then not using it.Game Developer Magazine, 2004. 14

    work page 2004

  6. [6]

    On the Limits of Pseudo Ground Truth in Visual Camera Re-localisation

    Eric Brachmann, Martin Humenberger, Carsten Rother, and Torsten Sattler. On the Limits of Pseudo Ground Truth in Visual Camera Re-localisation, 2021. arXiv:2109.00524. 1

    work page Pith review arXiv 2021

  7. [7]

    The euroc micro aerial vehicle datasets.The International Journal of Robotics Research, 35 (10):1157–1163, 2016

    Michael Burri, Janosch Nikolic, Pascal Gohl, Thomas Schneider, Joern Rehder, Sammy Omari, Markus W Achte- lik, and Roland Siegwart. The euroc micro aerial vehicle datasets.The International Journal of Robotics Research, 35 (10):1157–1163, 2016. 3

    work page 2016

  8. [8]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam.IEEE transactions on robotics, 37(6):1874– 1890, 2021

    Carlos Campos, Richard Elvira, Juan J G ´omez Rodr´ıguez, Jos´e MM Montiel, and Juan D Tard ´os. Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam.IEEE transactions on robotics, 37(6):1874– 1890, 2021. 8, 9

    work page 2021

Show all 52 references
  1. [9]

    Ushani, and Ryan M

    Nicholas Carlevaris-Bianco, Arash K. Ushani, and Ryan M. Eustice. University of Michigan North Campus long-term vision and lidar dataset.International Journal of Robotics Research, 35(9):1023–1035, 2015. 1, 3

    2015

  2. [10]

    University of Michigan North Campus long-term vi- sion and lidar dataset.The International Journal of Robotics Research, 35(9):1023–1035, 2016

    Nicholas Carlevaris-Bianco, Arash K Ushani, and Ryan M Eustice. University of Michigan North Campus long-term vi- sion and lidar dataset.The International Journal of Robotics Research, 35(9):1023–1035, 2016. Publisher: SAGE Publi- cations Ltd STM. 1

    2016

  3. [11]

    Rawseeds ground truth collection systems for indoor self- localization and mapping.Autonomous Robots, 27:353–371,

    Simone Ceriani, Giulio Fontana, Alessandro Giusti, Daniele Marzorati, Matteo Matteucci, Davide Migliore, Davide Rizzi, Domenico Giorgio Sorrenti, and Pierluigi Taddei. Rawseeds ground truth collection systems for indoor self- localization and mapping.Autonomous Robots, 27:353–371,

  4. [12]

    Leap-vo: Long-term effective any point tracking for visual odometry

    Weirong Chen, Le Chen, Rui Wang, and Marc Pollefeys. Leap-vo: Long-term effective any point tracking for visual odometry. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19844– 19853, 2024. 3, 8, 9

    2024

  5. [13]

    Ref-NeuS: Ambiguity-Reduced Neural Implicit Sur- face Learning for Multi-View Reconstruction with Reflec- tion, 2023

    Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-NeuS: Ambiguity-Reduced Neural Implicit Sur- face Learning for Multi-View Reconstruction with Reflec- tion, 2023. arXiv:2303.10840 [cs]. 2

    2023 arXiv

  6. [14]

    Are we ready for autonomous driving? the kitti vision benchmark suite

    Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. InConference on Computer Vision and Pattern Recog- nition (CVPR), 2012. 1, 3, 18

    2012

  7. [15]

    On averaging rotations.Journal of Mathe- matical Imaging and Vision, 15(1):7–16, 2001

    Claus Gramkow. On averaging rotations.Journal of Mathe- matical Imaging and Vision, 15(1):7–16, 2001. 14

    2001

  8. [16]

    Insta360 x4.https://www.insta360.com/ product/insta360-x4, 2024

    Insta360. Insta360 x4.https://www.insta360.com/ product/insta360-x4, 2024. 5

    2024

  9. [17]

    Large scale multi-view stereopsis eval- uation

    Rasmus Jensen, Anders Dahl, George V ogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis eval- uation. In2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406–413. IEEE, 2014. 3

    2014

  10. [18]

    W. Kabsch. A solution for the best rotation to relate two sets of vectors.Acta Crystallograph- ica Section A, 32(5):922–923, 1976. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1107/S0567739476001873. 7 10

    1976 doi

  11. [19]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  12. [20]

    G2o: A general framework for graph optimization

    Rainer K ¨ummerle, Giorgio Grisetti, Hauke Strasdat, Kurt Konolige, and Wolfram Burgard. G2o: A general framework for graph optimization. In2011 IEEE International Confer- ence on Robotics and Automation, pages 3607–3613, 2011. 6, 16

    2011

  13. [21]

    A method for the solution of certain non-linear problems in least squares.Quarterly of Applied Mathematics, 2(2):164–168, 1944

    Kenneth Levenberg. A method for the solution of certain non-linear problems in least squares.Quarterly of Applied Mathematics, 2(2):164–168, 1944. 7

    1944

  14. [22]

    Deep Patch Vi- sual SLAM, 2024

    Lahav Lipson, Zachary Teed, and Jia Deng. Deep Patch Vi- sual SLAM, 2024. arXiv:2408.01654. 1, 3, 8, 9

    2024 arXiv

  15. [23]

    The zurich urban micro aerial vehicle dataset.Int

    Andr ´as L Majdik, Charles Till, and Davide Scaramuzza. The zurich urban micro aerial vehicle dataset.Int. J. Rob. Res., 36(3):269–273, 2017. 3

    2017

  16. [24]

    Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines.ACM Transac- tions on Graphics (ToG), 38(4):1–14, 2019

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines.ACM Transac- tions on Graphics (ToG), 38(4):1–14, 2019. 3

    2019

  17. [25]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 3, 8

    2021

  18. [26]

    Raul Mur-Artal, J. M. M. Montiel, and Juan D. Tardos. ORB- SLAM: a Versatile and Accurate Monocular SLAM System,

  19. [27]

    Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

  20. [28]

    OpenCV: Detection of aruco boards

    OpenCV Team. OpenCV: Detection of aruco boards. https : / / docs . opencv . org / 4 . x / db / da9 / tutorial_aruco_board_detection.html, 2024. Accessed: 2024-11-15. 14

    2024

  21. [29]

    OpenCV: Detection of charuco boards

    OpenCV Team. OpenCV: Detection of charuco boards. https : / / docs . opencv . org / 3 . 4 / df / d4a / tutorial_charuco_detection.html, 2024. Ac- cessed: 2024-11-15. 14

    2024

  22. [30]

    Penncosyvio: A challenging visual inertial odometry benchmark

    Bernd Pfrommer, Nitin Sanket, Kostas Daniilidis, and Jonas Cleveland. Penncosyvio: A challenging visual inertial odometry benchmark. In2017 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 3847–3854,

  23. [31]

    The newer college dataset: Handheld lidar, inertial and vision with ground truth

    Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The newer college dataset: Handheld lidar, inertial and vision with ground truth. In2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020. 3

    2020

  24. [32]

    The Newer College Dataset: Handheld LiDAR, Inertial and Vision with Ground Truth, 2022

    Milad Ramezani, Yiduo Wang, Marco Camurri, David Wisth, Matias Mattamala, and Maurice Fallon. The Newer College Dataset: Handheld LiDAR, Inertial and Vision with Ground Truth, 2022. arXiv:2003.05691. 1

    2022 arXiv

  25. [33]

    SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing, 2023

    Jianyuan Ruan, Bo Li, Yibo Wang, and Yuxiang Sun. SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing, 2023. arXiv:2303.05252. 1

    2023 arXiv

  26. [34]

    Schonberger and Jan-Michael Frahm

    Johannes L. Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR),

  27. [35]

    Bad slam: Bundle adjusted direct rgb-d slam.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134–144, 2019

    Thomas Sch ¨ops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam.2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 134–144, 2019. 3

    2019

  28. [36]

    BAD SLAM: Bundle adjusted direct RGB-D SLAM

    Thomas Sch ¨ops, Torsten Sattler, and Marc Pollefeys. BAD SLAM: Bundle adjusted direct RGB-D SLAM. InConfer- ence on Computer Vision and Pattern Recognition (CVPR),

  29. [37]

    The tum vi benchmark for evaluating visual-inertial odometry

    David Schubert, Thore Goll, Nikolaus Demmel, Vladyslav Usenko, Jorg Stuckler, and Daniel Cremers. The tum vi benchmark for evaluating visual-inertial odometry. In2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2018. 1, 3, 4

    2018

  30. [38]

    Estimating the Dimension of a Model.The Annals of Statistics, 6(2):461 – 464, 1978

    Gideon Schwarz. Estimating the Dimension of a Model.The Annals of Statistics, 6(2):461 – 464, 1978. 7, 15

    1978

  31. [39]

    Sch ¨onberger and Jan-Michael Frahm

    Johannes L. Sch ¨onberger and Jan-Michael Frahm. Structure- from-motion revisited. In2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 4104– 4113, 2016. 2, 3, 8

    2016

  32. [40]

    A benchmark for the eval- uation of rgb-d slam systems

    J ¨urgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the eval- uation of rgb-d slam systems. In2012 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems, pages 573–580, 2012. 1, 3

    2012

  33. [41]

    Nerfstudio: A modular framework for neural radiance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. InACM SIGGRAPH 2023 conference proceedings, pages 1– 12,...

    2023

  34. [42]

    DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras, 2022

    Zachary Teed and Jia Deng. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras, 2022. arXiv:2108.10869. 1, 3

    2022 arXiv

  35. [43]

    Deep Patch Vi- sual Odometry, 2023

    Zachary Teed, Lahav Lipson, and Jia Deng. Deep Patch Vi- sual Odometry, 2023. arXiv:2208.04726. 8, 9

    2023 arXiv

  36. [44]

    S. Umeyama. Least-squares estimation of transformation pa- rameters between two point patterns.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(4):376–380,

  37. [45]

    How Are the Terms ‘Accu- racy’ & ‘Precision’ defined?https://www.vicon

    Vicon Motion Systems. How Are the Terms ‘Accu- racy’ & ‘Precision’ defined?https://www.vicon. com / support / faqs / how - are - the - terms - accuracy-precision-defined/. Accessed: Jul. 27,

  38. [46]

    DUSt3R: Geometric 3D Vision Made Easy, 2024

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy, 2024. arXiv:2312.14132 [cs]. 3

    2024 arXiv

  39. [47]

    Nex: Real-time view synthesis with neural basis expansion

    Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. Nex: Real-time view synthesis with neural basis expansion. InProceedings 11 of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8534–8543, 2021. 3

    2021

  40. [48]

    GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction, 2023

    Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction, 2023. arXiv:2309.02436. 1

    2023 arXiv

  41. [49]

    Oswald, and Marc Polle- feys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hu- jun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Polle- feys. NICE-SLAM: Neural Implicit Scalable Encoding for SLAM, 2022. arXiv:2112.12130. 3

    2022 arXiv

  42. [50]

    Oswald, Andreas Geiger, and Marc Pollefeys

    Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R. Oswald, Andreas Geiger, and Marc Pollefeys. NICER-SLAM: Neural Implicit Scene Encoding for RGB SLAM, 2023. arXiv:2302.03594. 1 12 Appendix Figure 7. Comparison of our ground truth trajectory with COLMAP and Vicon...

    2023 arXiv

  43. [52]

    However, we used stride 10 for the indoor and outdoor scenes since increasing the buffer size for longer sequences still resulted in failures

    As for LEAPVO, we also used the default configuration for the scanning sequences at stride 1. However, we used stride 10 for the indoor and outdoor scenes since increasing the buffer size for longer sequences still resulted in failures. Lastly, we ran COLMAP at stride 1 with d...

  44. [2015]

    arXiv:1502.00956. 1, 3

    arXiv

Pith tools

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