Pith. sign in

REVIEW 3 major objections 5 minor 63 references

PoseIDON: 6DoF Pose Estimation with Foundation Model Features for Marine Sediment Burial Mapping

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

Pith's one-line read From ordinary ROV footage, PoseIDON reports seafloor burial depths of barrels and munitions with a mean error of about 10 centimeters.

desk verdict A competent, clearly-written engineering pipeline that combines DINOv2/FoundPose with COLMAP and ICP to estimate burial depth of seafloor objects; the 10 cm claim is plausible as internal consistency, but the shared-COLMAP benchmark means true accuracy is not yet established. read the letter →

arxiv 2506.10386 v2 pith:AFVFAM5O submitted 2025-06-12 cs.CV

classification cs.CV
keywords 6DoFposeestimationmarinesedimentburialunderwatercomputervisionfoundationmodelfeaturesphotogrammetryROVsurveydepthmappingzero-shotobject
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

The paper argues that burial depth of human-made objects on the seafloor can be recovered from ordinary ROV video, without retraining on underwater data, given only a CAD model of each object. The proposed PoseIDON pipeline combines strong visual features from a self-supervised vision foundation model with classical multiview photogrammetry: it estimates a coarse monocular pose per frame, uses the object's known dimensions to scale the reconstruction, refines and filters the poses geometrically, fits a plane to the seafloor, and reads burial depth from the lowest point of the CAD mesh below that plane. Tested on footage of 54 barrels and munitions at a historic deep-ocean dumpsite, the paper reports a mean burial depth error of about 10 cm, or 0.11 relative to oriented object height, and burial-depth maps whose spatial pattern tracks known sedimentation. If correct, the method turns archived survey footage into a non-invasive tool for sedimentation monitoring, contamination-risk assessment, and munitions-recovery planning.

What carries the argument

The load-bearing object is a scaled, untextured CAD model aligned to a photogrammetric reconstruction whose scale is otherwise ambiguous. PoseIDON gets monocular pose hypotheses by matching frozen patch features from a self-supervised vision transformer between real images and rendered CAD templates; each hypothesis fixes the reconstruction's scale through a closed-form minimization of the variance of scaled object centroids. The surviving poses are then refined by iterative closest-point registration against the object's point cloud, filtered and averaged with RANSAC, and the seafloor is modeled as a plane fit to the remaining points; burial depth is the distance from the lowest CAD-mesh point to that plane. This chain is what converts arbitrary ROV flybys into metric burial depth, and the spatial burial patterns are produced by mapping that scalar per object location.

What would settle it

Physically measure burial depth on a subset of the same 54 objects using a method independent of photogrammetry, for example sediment probing or excavation from an ROV, or a pre-surveyed ground-control frame around the object, and compare those values to PoseIDON's predictions. If the mean absolute difference is much larger than 10 cm, or if the underestimate bias disappears or reverses, the reported accuracy reflects agreement with a biased benchmark rather than true burial depth.

Watch

Extended reading notes

Core claim

The central claim is that zero-shot, CAD-driven 6DoF pose estimation can survive the extreme underwater conditions of occlusion, marine growth, haze, and unknown camera intrinsics if monocular foundation-feature pose guesses are filtered through multiview geometry. On 54 selected objects in the San Pedro Basin, the full pipeline achieves a mean burial-depth error of roughly 10 cm (an 11% error relative to oriented object height), with most errors under 0.15 m and only a few catastrophic failures. The paper also reports that the model systematically underestimates burial depth, which it attributes to fouling and sediment making the real object physically larger than its CAD model, and that errors grow for deeper objects with fewer visible features. Validated against manual labels, the recovered depth field distinguishes a central barrel cluster with mean depth about 0.39 m from an eastern cluster at about 0.29 m, matching the sediment-transport story of the site.

Load-bearing premise

The load-bearing premise is that the manually labeled burial depths, which are themselves produced from the same photogrammetric reconstruction the pipeline uses, are accurate enough to serve as ground truth for the claimed 10 cm error.

Editorial extensions

If this is right

  • Archived ROV footage of legacy dumpsites can be reprocessed into burial-depth maps without new data collection or task-specific training, as long as a CAD model exists for each object type.
  • A mean error near 10 cm is small enough to resolve region-scale burial differences such as the 0.39 m versus 0.29 m cluster contrast, so the method can support qualitative and coarse quantitative sediment-transport assessments.
  • The reported underestimate bias and poor performance on deeply buried objects identify a clear operating envelope: objects with most of their surface visible, and the systematic bias should be corrected before trusting absolute depths.
  • For munitions and barrels with known dimensions, the pipeline provides a non-invasive alternative to sediment coring for estimating local accumulation, and it can be combined with physics-based impact-burial models to cross-check object age or sedimentation rates.

Reading between the lines

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

  • Because the manual labels are produced from the same photogrammetric reconstruction that the pipeline uses, shared systematic errors in camera poses or plane orientation would make the reported 10 cm error an underestimate of error against true burial; an independent physical measurement campaign is the natural check.
  • The planar-seafloor assumption is the most fragile geometric link: sediment mounds against object bases mean the fitted plane need not coincide with the true sediment surface, so depth errors will be correlated with local sediment build-up rather than random.
  • The same scale-recovery trick, using a known CAD dimension to metrize an unscaled reconstruction, could transfer to other marine infrastructure monitoring tasks such as pipeline scour, cable exposure, or mooring-block burial wherever a reference model exists.
  • A testable extension would be to compare PoseIDON estimates against acoustic sub-bottom profiles or against deliberately planted objects of known burial depth in a controlled underwater test site.
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 PoseIDON, a pipeline for estimating 6DoF pose and burial depth of anthropogenic objects on the seafloor from ROV video. The method combines DINOv2/FoundPose monocular pose estimates with COLMAP/OpenMVS photogrammetry, ICP refinement, RANSAC filtering, and planar seafloor fitting. The authors validate on 54 objects from the San Pedro Basin, reporting a mean burial depth error of approximately 10 cm (0.11 relative to oriented object height) and spatial burial patterns consistent with sedimentation. Ablation experiments compare the full multiview pipeline with monocular-only estimates and with alternative reconstruction backbones (Fast3r, VGGT). The paper includes a Limitations section and releases source code.

Significance. If the accuracy claim holds, the contribution is practically significant: it offers a retraining-free, scalable method for mapping seafloor burial from standard ROV footage, which is relevant for environmental monitoring and legacy waste assessment. The paper's strengths include a clearly described pipeline, explicit parameter settings, an ablation against deep-learning reconstruction baselines, and an honest discussion of failure modes. However, the central quantitative claim rests on an evaluation benchmark that is not independent of the method: manual labels are created inside the same COLMAP reconstruction that PoseIDON uses, so reconstruction errors are shared by both sides. Without independent ground truth or a sensitivity analysis, the reported 10 cm error measures agreement with a biased benchmark rather than absolute accuracy. This is a load-bearing limitation that affects the headline result, the spatial-pattern claim, and the comparison with alternative reconstructions.

major comments (3)
  1. [Evaluation, Manually Labeling Poses] The ground truth used for the central accuracy claim is not independent of the method. As stated in the paper, 'For each video of an object, photogrammetry is used to reconstruct camera positions and manually orient the object and seafloor in the scene,' and the authors acknowledge that 'This process ties the quality of the labeling to photogrammetry performance.' Since PoseIDON also uses COLMAP/OpenMVS to recover camera poses, intrinsics, and the dense point cloud, any systematic error in the reconstruction—particularly in the orientation of the seafloor plane or in the metric scale—enters both the manual labels and the model predictions. The reported mean burial depth error of approximately 10 cm therefore quantifies agreement between two procedures sharing a common source of bias, not accuracy against an external standard. Rerunning photogrammetry on different frames estimates label variance, not bias, so it does not address this concern. The authors should either provide independent validation (e.g., physical measurements of exposed object height, acoustic ground truth, or synthetic scenes with known poses) or explicitly reframe the claims as consistency with manual photogrammetric labels and quantify how sensitive the burial-depth error is to plausible reconstruction errors.
  2. [Results, Model Performance] The headline result 'mean error of around 10 cm' is reported without a confidence interval or an uncertainty propagation from the per-object estimates. Figure 2 shows that both manual labels and model predictions carry standard deviations on the order of 6-18 cm across n=6 repeated runs, and Figure 4 shows a Pearson correlation of only ρ=0.49 between predicted and labeled burial depths. With this scatter, the precision of the 10 cm estimate is unclear, and it is not possible to assess whether the difference between the central and eastern clusters (0.39 m vs. 0.29 m in Figure 5) is meaningful. The authors should report the mean error with a confidence interval or standard error, and ideally a per-object error distribution, to support the quantitative claims.
  3. [Mapping Burial Depths] The claim that the model 'resolves spatial burial patterns that reflect underlying sediment transport processes' is not statistically supported. Figure 5 compares two clusters of objects (n=20 and n=8) whose reported standard deviations overlap substantially, and no significance test or effect-size measure is provided. The observed difference could arise from object type, initial burial depth, or random variation rather than sediment transport. The authors should either add a statistical comparison or temper the conclusion to state that the predicted depths reproduce a qualitative spatial trend visible in the manual labels.
minor comments (5)
  1. [Evaluation, Manually Labeling Poses] The manual labeling procedure in Blender is described only briefly; please provide more detail on how the object and seafloor orientations are initialized and constrained, how the CAD model scale is imposed, and whether any inter-operator or inter-run variability was assessed beyond the n=6 resampling.
  2. [Table 2] The first two rows of Table 2 report ARVSD/ARMSSD/ARMSPD for monocular FoundPose with and without featuremetric refinement, but no burial-depth error is given for these configurations; this makes it difficult to see how much of the final accuracy improvement comes from the multiview stages versus the monocular baseline.
  3. [Results, Model Performance] The text states that 'the model has a tendency to underestimate the burial depth' and attributes this to biological growth and sediment volume; this explanation is plausible but not tested, and it would be helpful to quantify the bias (mean signed error) separately from the absolute error.
  4. [Figure 5] The caption reports regional means with '±' values but does not define whether these are standard deviations or standard errors; please clarify, and consider showing the underlying object-level points to make the cluster comparison more transparent.
  5. [Implementation Details] The RANSAC thresholds and ICP cutoff are stated clearly, but the sensitivity of the final burial-depth estimate to these parameters is not explored; a short sensitivity analysis would strengthen the reproducibility of the pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: PoseIDON's pose and burial-depth estimates are derived from independent photogrammetric, feature-matching, ICP, and plane-fitting stages, and the manual-label benchmark, though sharing a photogrammetric reconstruction, is not an input to the predictions.

full rationale

The paper's derivation chain is self-contained rather than circular. The pipeline takes ROV images and a CAD model as inputs, obtains monocular poses with FoundPose/DINOv2, recovers scale via centroid-variance minimization (Eq. 6-7), refines with ICP, aggregates poses with RANSAC, fits a seafloor plane, and measures burial depth from the CAD mesh. None of these steps fits a parameter to the manual labels or defines the target in terms of the output. The manual labels are produced independently by orienting the object and seafloor in a Blender scene; the paper explicitly notes the limitation that 'This process ties the quality of the labeling to photogrammetry performance' (Evaluation, Manually Labeling Poses). This is a benchmark-validity caveat, not a derivation circularity: the model's predictions are not constructed from the labels, and the reported ~10 cm error could in principle be large even with shared photogrammetry because the manual orientations and the model's feature/ICP stages are independent. The self-citations ([1], [11]) are not load-bearing: [1] supplies the survey dataset and [11] is only contextual related work. The ablation comparing COLMAP with Fast3r/VGGT is weakened by the manual labels being placed in a photogrammetric reconstruction, but this affects external validity rather than making any prediction equivalent to its input by construction. No equation, fitted parameter, or uniqueness claim reduces to an input, so the circularity score is 0.

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

The pipeline is an engineering integration of existing components (DINOv2, FoundPose, COLMAP, OpenMVS, ICP). No new physical entities are introduced. The main assumptions are domain-specific: the seafloor is locally planar, the CAD model is dimensionally faithful, DINOv2 features transfer underwater, and COLMAP can recover camera geometry from degraded ROV footage. Several hand-set RANSAC thresholds and the ICP cutoff are free parameters; no sensitivity analysis is provided.

free parameters (6)
  • RANSAC scale-correction inlier threshold = 0.15 m
    Hand-chosen threshold for selecting inlier object centroids during scale recovery; affects the estimated scale and thus all burial depths.
  • RANSAC plane-fitting inlier threshold = 0.05 m
    Hand-chosen distance threshold for seafloor plane inliers; directly affects plane orientation and burial depth.
  • RANSAC rotation-averaging inlier threshold = 0.2 rad
    Hand-chosen angular threshold for selecting quaternions before averaging; affects final object orientation.
  • ICP outlier standard-deviation cutoff = t = 2
    Points registering more than 2 standard deviations from the mean registration distance are excluded; affects the refined pose.
  • FoundPose pose hypotheses per image = H = 5
    Number of template hypotheses kept from FoundPose; larger H increases robustness but also computation; set by the FoundPose reference.
  • Point-cloud mask projection threshold
    A point is kept as object point if it projects into the segmented mask in at least a threshold number of images; the value is not specified in the paper.
assumptions (6)
  • domain assumption The seafloor can be modeled as a local plane.
    The pipeline fits a plane to the point cloud around the object and defines burial depth relative to this plane. The paper acknowledges non-planarity and sediment buildup next to objects but relies on this approximation for the burial-depth estimate. See 'Discussion' and 'Plane Fitting and Burial Estimation'.
  • domain assumption The provided CAD models match the real objects in dimensions and overall shape despite degradation and biological growth.
    The CAD model is used as the reference for pose estimation and scale recovery; the paper itself attributes the observed underestimation bias to extra sediment and biological volume between the real object and the CAD model. See 'Results, Model Performance'.
  • domain assumption DINOv2 features transfer from synthetic CAD renderings to real underwater imagery without retraining.
    The monocular pose stage and the synthetic-to-real matching depend entirely on this transfer. The paper provides qualitative evidence and cites FoundPose, but no formal guarantee for degraded underwater images. See 'Robust Feature Detection and Matching'.
  • domain assumption COLMAP can recover camera intrinsics and extrinsics from underwater ROV footage despite unknown camera intrinsics, changing zoom, and visual degradation.
    Camera intrinsics were not recorded and the footage is recorded with a wide-angle camera under poor visibility; COLMAP must estimate these. Both the model and the manual labeling rely on this reconstruction. See 'Data Collection' and 'Segmented 3D Reconstruction'.
  • standard math The pinhole camera model with square pixels and zero skew is adequate.
    Stated in 'Preliminaries, Representing Object Poses and Cameras' as the projection model used throughout.
  • ad hoc to paper The scale factor that minimizes the variance of object centroids is the correct scene scale.
    Equations (6)-(7) define the scale estimate based on the assumption that correctly scaled camera positions bring all FoundPose object translations into agreement. If monocular translations are systematically biased, this scale will be biased too.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PoseIDON: 6DoF Pose Estimation with Foundation Model Features for Marine Sediment Burial Mapping." pith.science (2026). https://pith.science/paper/AFVFAM5O

@misc{pith2026250610386,
  author       = {Pith},
  title        = {Pith review of: PoseIDON: 6DoF Pose Estimation with Foundation Model Features for Marine Sediment Burial Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFVFAM5O}},
  note         = {Machine review of arXiv:2506.10386}
}
read the original abstract

The burial state of anthropogenic objects on the seafloor provides insight into localized sedimentation dynamics and is also critical for assessing ecological risks, potential pollutant transport, and the viability of recovery or mitigation strategies for hazardous materials such as munitions. Accurate burial depth estimation from remote imagery remains difficult due to partial occlusion, poor visibility, and object degradation. This work introduces a computer vision pipeline, called PoseIDON, which combines deep foundation model features with multiview photogrammetry to estimate six degrees of freedom object pose and the orientation of the surrounding seafloor from ROV video. Burial depth is inferred by aligning CAD models of the objects with observed imagery and fitting a local planar approximation of the seafloor. The method is validated using footage of 54 objects, including barrels and munitions, recorded at a historic ocean dumpsite in the San Pedro Basin. The model achieves a mean burial depth error of approximately 10 centimeters and resolves spatial burial patterns that reflect underlying sediment transport processes. This approach enables scalable, non-invasive mapping of seafloor burial and supports environmental assessment at contaminated sites.

Figures

Figures reproduced from arXiv: 2506.10386 by the authors.

Figure 1
Figure 1. To-scale untextured renderings of the CAD models fed into FoundPose. From left to right: drum barrel, depth [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Five example results from footage across multiple ROV dives. Each observed object (column) shows the video [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The distributions of ground truth, predictions, and errors for burial depth of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: A histogram comparing the manually labeled versus the predicted burial depth of [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The ground truth, predictions, and errors for burial depth of [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: A comparison of feature matching between hand-crafted SIFT features [ [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: A toy example of how texture inconsistencies can affect the performance of FoundPose [ [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Video frames are shown from the blue camera (top) and red camera (bottom), and the camera’s movement [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: An overview of PoseIDON’s pipeline. Given a series of RGB underwater images, inpainting is performed on [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: An example video frame, with the original image (top) and the image with the text overlay inpainted [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: An example video frame with the barrel bounding box (orange) found by Grounding DINO [ [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: A visualization of the camera trajectory recovered by COLMAP (purple) and the dense 3D reconstruction [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Example FoundPose [26] intermediary steps and results on a barrel. The input image is shown with its mask overlaid in white (top right), the top five most similar rendered templates (middle row), and the DINOv2 [12] patch descriptor matches between the real barrel (bo…
Figure 14
Figure 14. Figure 14: A visualization of the scale correction operation on the photogrammetry 3D reconstruction. Cameras [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]
Figure 15
Figure 15. Figure 15: After scale correction, ICP is run on all coarse poses (top), with the original pose (red) and pose after ICP [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]
Figure 16
Figure 16. Figure 16: The final render of the predicted barrel pose relative to the fitted plane. The blue part of the barrel mesh is [PITH_FULL_IMAGE:figures/full_fig_p017_16.png]
Figure 17
Figure 17. Figure 17: An example failure case with a deeply buried mousetrap, where only the fin is unburied (top left). A decent [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 52 canonical work pages

  1. [1]

    Merrifield, Sean Celona, Ryan A

    Sophia T. Merrifield, Sean Celona, Ryan A. McCarthy, Andrew Pietruszka, Heidi Batchelor, Robert Hess, Andrew Nager, Raymond Young, Kurt Sadorf, Lisa A. Levin, David L. Valentine, James E. Conrad, and Eric J. Terrill. Wide-Area Debris Field and Seabed Characterization of a Deep Ocean Dump Site Surveyed by Autonomous Underwater Vehicles.Environmental Scienc...

  2. [2]

    Newly identified DDT-related compounds accumulating in Southern California bottlenose dolphins.Environmental science & technology, 50(22):12129–12137, 2016

    Susan A Mackintosh, Nathan G Dodder, Nellie J Shaul, Lihini I Aluwihare, Keith A Maruya, Susan J Chivers, Kerri Danil, David W Weller, and Eunha Hoh. Newly identified DDT-related compounds accumulating in Southern California bottlenose dolphins.Environmental science & technology, 50(22):12129–12137, 2016. Publisher: ACS Publications

  3. [3]

    Disentangling the History of Deep Ocean Disposal for DDT and Other Industrial Waste Off Southern California.Environmental Science & Technology, 58(9):4346–4356, 2024

    Jacob T Schmidt, Mong Sin Christine Wu, Hailie E Kittner, J Samuel Arey, Douglas E Hammond, Earth 182A Group, and David L Valentine. Disentangling the History of Deep Ocean Disposal for DDT and Other Industrial Waste Off Southern California.Environmental Science & Technology, 58(9):4346–4356, 2024. Publisher: ACS Publications. 17 PREPRINT- SEPTEMBER10, 20...

  4. [4]

    Chu, Chenwu Fan, Joseph Calantoni, and Alex Sheremet

    Peter C. Chu, Chenwu Fan, Joseph Calantoni, and Alex Sheremet. Prediction of Mobility and Burial of Objects on Sandy Seafloor.IEEE Journal of Oceanic Engineering, 47(1):111–125, 2022

  5. [5]

    Restreppo, Warren T

    Giancarlo A. Restreppo, Warren T. Wood, and Benjamin J. Phrampus. Oceanic sediment accumulation rates predicted via machine learning algorithm: towards sediment characterization on a global scale.Geo-Marine Letters, 40(5):755–763, October 2020

  6. [6]

    A new deep learning engine for coralnet

    Qimin Chen, Oscar Beijbom, Stephen Chan, Jessica Bouwmeester, and David Kriegman. A new deep learning engine for coralnet. InProceedings of the IEEE/CVF international conference on computer vision, pages 3693–3702, 2021

  7. [7]

    FathomNet: A global image database for enabling artificial intelligence in the ocean.Scientific reports, 12(1):15914, 2022

    Kakani Katija, Eric Orenstein, Brian Schlining, Lonny Lundsten, Kevin Barnard, Giovanna Sainz, Oceane Boulais, Megan Cromwell, Erin Butler, Benjamin Woodward, and others. FathomNet: A global image database for enabling artificial intelligence in the ocean.Scientific reports, 12(1):15914, 2022. Publisher: Nature Publishing Group UK London

  8. [8]

    Evaluation of PoseNet for 6-DOF Underwater Pose Estimation

    Mikkel Cornelius Nielsen, Mari Hovem Leonhardsen, and Ingrid Schjølberg. Evaluation of PoseNet for 6-DOF Underwater Pose Estimation. InOCEANS 2019 MTS/IEEE SEATTLE, pages 1–6, 2019

Show all 63 references
  1. [9]

    Ayoung Kim and Ryan M. Eustice. Real-Time Visual SLAM for Autonomous Underwater Hull Inspection Using Visual Saliency.IEEE Transactions on Robotics, 29(3):719–733, 2013

  2. [10]

    Visual SLAM for underwater vehicles: A survey.Computer Science Review, 46:100510, 2022

    Song Zhang, Shili Zhao, Dong An, Jincun Liu, He Wang, Yu Feng, Daoliang Li, and Ran Zhao. Visual SLAM for underwater vehicles: A survey.Computer Science Review, 46:100510, 2022

  3. [11]

    Pose Estimation of Buried Deep-Sea Objects using 3D Vision Deep Learning Models

    Jerry Yan, Chinmay Talegaonkar, Nicholas Antipa, Eric Terrill, and Sophia Merrifield. Pose Estimation of Buried Deep-Sea Objects using 3D Vision Deep Learning Models. InOCEANS 2024 - Halifax, pages 1–6. IEEE, 2024

  4. [12]

    DINOv2: Learning Robust Visual Features without Supervision, 2024

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabb...

  5. [13]

    Structure-from-Motion Revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-Motion Revisited. InConference on Computer Vision and Pattern Recognition (CVPR), 2016. 18 PREPRINT- SEPTEMBER10, 2025

  6. [14]

    Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys

    Johannes L. Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors,Computer Vision – ECCV 2016, pages 501–518, Cham, 2016. Springer Intern...

  7. [15]

    Deep Learning-Based Object Pose Estimation: A Comprehensive Survey, 2024

    Jian Liu, Wei Sun, Hui Yang, Zhiwen Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, and Ajmal Mian. Deep Learning-Based Object Pose Estimation: A Comprehensive Survey, 2024. _eprint: 2405.07801

  8. [16]

    The MOPED framework: Object recognition and pose estimation for manipulation.I

    Alvaro Collet, Manuel Martinez, and Siddhartha Srinivasa. The MOPED framework: Object recognition and pose estimation for manipulation.I. J. Robotic Res., 30:1284–1306, September 2011

  9. [17]

    D.G. Lowe. Object recognition from local scale-invariant features. InProceedings of the Seventh IEEE Interna- tional Conference on Computer Vision, volume 2, pages 1150–1157 vol.2, 1999

  10. [18]

    Besl and Neil D

    P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes.IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):239–256, 1992

  11. [19]

    Scale-Adaptive ICP.Graphical Models, 116:101113, 2021

    Yusuf Sahillio ˘glu and Ladislav Kavan. Scale-Adaptive ICP.Graphical Models, 116:101113, 2021

  12. [20]

    Alvarez, and Mathieu Salzmann

    Shuxuan Guo, Yinlin Hu, Jose M. Alvarez, and Mathieu Salzmann. Knowledge Distillation for 6D Pose Estimation by Aligning Distributions of Local Predictions. pages 18633–18642, June 2023

  13. [21]

    Yan Xu, Kwan-Yee Lin, Guofeng Zhang, Xiaogang Wang, and Hongsheng Li. RNNPose: 6-DoF Object Pose Estimation via Recurrent Correspondence Field Estimation and Pose Optimization.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4669–4683, 2024

  14. [22]

    Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image

    Zhaoxin Fan, Zhenbo Song, Jian Xu, Zhicheng Wang, Kejian Wu, Hongyan Liu, and Jun He. Object level depth reconstruction for category level 6d object pose estimation from monocular rgb image. InEuropean Conference on Computer Vision, pages 220–236. Springer, 2022

  15. [23]

    RGB-based category-level object pose estimation via decoupled metric scale recovery

    Jiaxin Wei, Xibin Song, Weizhe Liu, Laurent Kneip, Hongdong Li, and Pan Ji. RGB-based category-level object pose estimation via decoupled metric scale recovery. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2036–2042. IEEE, 2024

  16. [24]

    GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence

    Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. GigaPose: Fast and Robust Novel Object Pose Estimation via One Correspondence. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9903–9913, June 2024

  17. [25]

    Zs6d: Zero-shot 6d object pose estimation using vision transformers

    Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages 463–469. IEEE, 2024

  18. [26]

    FoundPose: Unseen Object Pose Estimation with Foundation Features, 2024

    Evin Pınar Örnek, Yann Labbé, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomas Hodan. FoundPose: Unseen Object Pose Estimation with Foundation Features, 2024. _eprint: 2311.18809

  19. [27]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at...

  20. [28]

    Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S

    Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jare...

  21. [29]

    Emerging Properties in Self-Supervised Vision Transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630–9640, 2021. 19 PREPRINT- SEPTE...

  22. [30]

    Seitz, and Richard Szeliski

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M. Seitz, and Richard Szeliski. Building Rome in a day.Commun. ACM, 54(10):105–112, October 2011. Place: New York, NY , USA Publisher: Association for Computing Machinery

  23. [31]

    From oblique photogrammetry to a 3D model – Structural modeling of Kilen, eastern North Greenland.Computers & Geosciences, 83:120–126, 2015

    Kristian Svennevig, Pierpaolo Guarnieri, and Lars Stemmerik. From oblique photogrammetry to a 3D model – Structural modeling of Kilen, eastern North Greenland.Computers & Geosciences, 83:120–126, 2015

  24. [32]

    Underwater 3D Reconstruction from Video or Still Imagery: Matisse and 3DMetrics Processing and Exploitation Software.Journal of Marine Science and Engineering, 11(5), 2023

    Aurélien Arnaubec, Maxime Ferrera, Javier Escartín, Marjolaine Matabos, Nuno Gracias, and Jan Opderbecke. Underwater 3D Reconstruction from Video or Still Imagery: Matisse and 3DMetrics Processing and Exploitation Software.Journal of Marine Science and Engineering, 11(5), 2023

  25. [33]

    3D reconstruction based on underwater video from ROV Kiel 6000 considering underwater imaging conditions

    Anne Sedlazeck, Kevin Koser, and Reinhard Koch. 3D reconstruction based on underwater video from ROV Kiel 6000 considering underwater imaging conditions. InOCEANS 2009-EUROPE, pages 1–10, 2009

  26. [34]

    Pizarro, R

    O. Pizarro, R. Eustice, and H. Singh. Large area 3D reconstructions from underwater surveys. InOceans ’04 MTS/IEEE Techno-Ocean ’04 (IEEE Cat. No.04CH37600), volume 2, pages 678–687 V ol.2, 2004

  27. [35]

    Underwater programmetry for archaeology and marine biology: 40 years of experience in Marseille, France

    Pierre Drap, Djamel Merad, Julien Seinturier, Amine Mahiddine, Daniela Peloso, Jean-Marc Boï, Bertrand Chemisky, Luc Long, and Joaquim Garrabou. Underwater programmetry for archaeology and marine biology: 40 years of experience in Marseille, France. In2013 Digital Heritage Int...

  28. [36]

    Pixel-Perfect Structure-from- Motion with Featuremetric Refinement, 2021

    Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-Perfect Structure-from- Motion with Featuremetric Refinement, 2021. _eprint: 2108.08291

  29. [37]

    SfM-Net: Learning of Structure and Motion from Video, 2017

    Sudheendra Vijayanarasimhan, Susanna Ricco, Cordelia Schmid, Rahul Sukthankar, and Katerina Fragkiadaki. SfM-Net: Learning of Structure and Motion from Video, 2017. _eprint: 1704.07804

  30. [38]

    Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli

    Jianing Yang, Alexander Sax, Kevin J. Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass, 2025. _eprint: 2501.13928

  31. [39]

    VGGT: Visual Geometry Grounded Transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT: Visual Geometry Grounded Transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

  32. [40]

    BOP challenge 2020 on 6D object localization

    Tomáš Hodaˇn, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiˇrí Matas. BOP challenge 2020 on 6D object localization. InComputer Vision–ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 577–...

  33. [41]

    Blender, 2024

    Blender Foundation and Blender Community. Blender, 2024. Place: Amsterdam, The Netherlands

  34. [42]

    Alexander and Homa J

    Clark R. Alexander and Homa J. Lee. Sediment accumulation on the Southern California Bight continental margin during the twentieth century. InEarth Science in the Urban Ocean: The Southern California Continental Borderland. Geological Society of America, January 2009

  35. [43]

    Chu and Chenwu Fan

    Peter C. Chu and Chenwu Fan. Prediction of Falling Cylinder Through Air-Water- Sediment Columns.Journal of Applied Mechanics, 73(2):300–314, August 2005. _eprint: https://asmedigitalcollection.asme.org/appliedmechanics/article-pdf/73/2/300/5473708/300_1.pdf

  36. [44]

    Berelson, D.E

    W.M. Berelson, D.E. Hammond, G. Smith, F. Spagnoli, A. Sturz, D. Colodner, and F. Chavez. The flushing of two deep-sea basins, southern California borderland.Limnology and Oceanography, 36(9):1687–1705, 1991

  37. [45]

    Santschi, and Kimberly A

    Chih-An Huh, Peter H. Santschi, and Kimberly A. Roberts. Long-term accumulation of polychlorinated biphenyls and DDT compounds in Southern California Bight sediments.Toxicological and Environmental Chemistry, 29(1-2):19–32, 1990. Publisher: Taylor & Francis

  38. [46]

    Berelson, Douglas E

    Nathaniel Kemnitz, William M. Berelson, Douglas E. Hammond, Laura Morine, Maria Figueroa, Timothy W. Lyons, Simon Scharf, Nick Rollins, Elizabeth Petsios, Sydnie Lemieux, and Tina Treude. Evidence of changes in sedimentation rate and sediment fabric in a low-oxygen setting: Sa...

  39. [47]

    Zeng, C.L

    E.Y . Zeng, C.L. Vista, K. Tran, and J. Kim. Distribution of chlorinated hydrocarbons in sediments of the Southern California Bight.Science of The Total Environment, 179:199–209, 1996. Publisher: Elsevier

  40. [48]

    Huh, L.F

    C.A. Huh, L.F. Small, L. Grant Ludwig, and R. Gordon. Sedimentation dynamics in the Santa Monica–San Pedro Basin off Los Angeles: radiochemical, sediment trap and transmissometer studies.Marine Chemistry, 26(1-3):143–165, 1989. Publisher: Elsevier

  41. [49]

    Nittrouer and Larry D

    Charles A. Nittrouer and Larry D. Wright. Transport of particles across continental shelves.Continental Shelf Research, 31(10):1050–1063, 2011. 20 PREPRINT- SEPTEMBER10, 2025

  42. [50]

    Kivenson, M

    V . Kivenson, M. Paddock, K.A. Maruya, M.I. Venkatesan, and R. Montagna. Ocean dumping of containerized DDT waste was a sloppy process.Environmental Science & Technology, 53(7):3702–3705, 2019. Publisher: ACS

  43. [51]

    Environmental Protection Agency Region IX

    U.S. Environmental Protection Agency Region IX. Initial Findings Regarding Ocean Disposal of Montrose Chemical’s Acid Waste. Technical report, EPA, 2021

  44. [52]

    Executive Office of the President, Washington, D.C., 1970

    Council on Environmental Quality.Ocean Dumping: A National Policy—Report to the President. Executive Office of the President, Washington, D.C., 1970

  45. [53]

    David G. Lowe. Distinctive Image Features from Scale-Invariant Keypoints.International Journal of Computer Vision, 60(2):91–110, November 2004

  46. [54]

    Deep ViT Features as Dense Visual Descriptors

    Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep ViT Features as Dense Visual Descriptors. ECCVW What is Motion For?, 2022

  47. [55]

    Fischler and Robert C

    Martin A. Fischler and Robert C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Commun. ACM, 24(6):381–395, June 1981. Place: New York, NY , USA Publisher: Association for Computing Machinery

  48. [56]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and others. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023

  49. [57]

    SAM 2: Segment Anything in Images and Videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenh...

  50. [58]

    An Image Inpainting Technique Based on the Fast Marching Method.Journal of Graphics Tools, 9(1):23–34, 2004

    Alexandru Telea. An Image Inpainting Technique Based on the Fast Marching Method.Journal of Graphics Tools, 9(1):23–34, 2004. Publisher: Taylor & Francis _eprint: https://doi.org/10.1080/10867651.2004.10487596

  51. [59]

    OpenMVS: Multi-View Stereo Reconstruction Library, 2020

    Dan Cernea. OpenMVS: Multi-View Stereo Reconstruction Library, 2020

  52. [60]

    Fast and globally convergent pose estimation from video images

    C-P Lu, Gregory D Hager, and Eric Mjolsness. Fast and globally convergent pose estimation from video images. IEEE transactions on pattern analysis and machine intelligence, 22(6):610–622, 2000. Publisher: IEEE

  53. [61]

    EPnP: An accurate O(n) solution to the PnP problem

    Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. EPnP: An accurate O(n) solution to the PnP problem. International Journal of Computer Vision, 81, February 2009

  54. [62]

    A consistently fast and globally optimal solution to the perspective-n-point problem

    George Terzakis and Manolis Lourakis. A consistently fast and globally optimal solution to the perspective-n-point problem. InEuropean Conference on Computer Vision, pages 478–494. Springer, 2020

  55. [63]

    Averaging Quaternions.Journal of Guidance, Control, and Dynamics, 30:1193–1196, July 2007

    Landis Markley, Yang Cheng, John Crassidis, and Yaakov Oshman. Averaging Quaternions.Journal of Guidance, Control, and Dynamics, 30:1193–1196, July 2007. 21

Pith tools

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