Pith. sign in

REVIEW 3 major objections 8 minor 61 references

Using overlapping seafloor footprints as ground truth, this paper shows that visual place recognition in dynamic benthic environments is substantially harder than existing benchmarks indicate.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 18:55 UTC pith:62RXLKUQ

load-bearing objection Useful first photic-zone multi-site underwater VPR benchmark, with a plausible footprint-based ground truth whose main weakness is missing independent validation of the footprint geometry; the core result that benthic VPR is hard is probably robust. the 3 major comments →

arxiv 2603.04056 v1 pith:62RXLKUQ submitted 2026-03-04 cs.CV cs.RO

Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark

Martin Kvisvik Larsen , Oscar Pizarro This is my paper
classification cs.CV cs.RO
keywords benthic environmentsvisual place recognitionlong-term visual localizationimage footprintsground truthAUV imageryunderwater roboticsrepeat surveys
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the standard way of judging whether an underwater robot has recognized a place — measuring how close two image locations are — is misleading when the seafloor is rugged or the robot's altitude varies. It introduces a dataset of five seafloor sites revisited over up to six years and a ground-truthing method that computes each image's 3D footprint on the seafloor and declares a match only when footprints overlap. Using this ground truth, it benchmarks eight visual place recognition models and finds Recall@K far below what terrestrial and deep-sea benchmarks report. The paper also shows that switching from a distance threshold to footprint overlap lowers reported recall, especially at sites with terrain relief, because spatial proximity is a poor proxy for shared visual content.

Core claim

The central claim is that overlap-aware ground truth is the right way to evaluate long-term visual place recognition in near-nadir underwater imagery, and that the field has been underestimating task difficulty. The authors estimate the 3D footprint of each down-facing image on the seafloor by combining stereo-derived metric range with a monocular relative-depth map scaled by a single per-image affine transform, project the four image corners onto the seafloor, and link a query image to a database image when their footprints' intersection-over-union exceeds 0.07. On the resulting ground truth, eight state-of-the-art visual place recognition methods achieve Recall@10 of roughly 12-51% dependi

What carries the argument

The load-bearing mechanism is the image footprint: the quadrilateral on the seafloor obtained by inverse-projecting the four image corners through a calibrated pinhole camera, using per-corner range from a fused range map. The range map for each image is built by taking stereo-computed metric disparity and rescaling a complete monocular relative-depth prediction with one global affine transform (z_fused = a*z_rel + b) fitted by robust regression. Footprint overlap, measured by intersection over union with a conservative threshold tau_f = 0.07 derived from altitude, field of view, and the 0.16 m registration error, converts geometric registration into content-based ground-truth links.

Load-bearing premise

The ground truth assumes that each image's fused range map is metrically accurate — specifically, that a monocular network's relative depth is related to true stereo-measured distance by one global scale-and-offset per image; if light attenuation or backscatter breaks that linear relationship, the estimated seafloor footprints and the links built from their overlap would be wrong.

What would settle it

Take a site visit, place or identify a flat, texture-rich patch of seafloor with known dimensions and altitude from the vehicle's Doppler velocity log, then compare the estimated footprint corners against the analytic projection onto that plane. If the corner error consistently exceeds the roughly 0.16 m registration error used to set the IoU threshold, the footprint-based ground-truth links — and the recall numbers computed from them — are not trustworthy.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Under footprint-based ground truth, long-term visual place recognition in dynamic benthic environments is much harder than previous benchmarks suggest: best models reach only roughly 20-51% Recall@10 depending on site.
  • Distance-threshold ground truth overestimates Recall@K, and the overestimation grows when terrain relief or altitude differences make footprint width exceed or misalign with a fixed spatial threshold.
  • Transformer-based place recognition models consistently outperform convolutional models across all five sites.
  • Recall@10 declines as revisit interval grows, with the sharpest drop between one and two years and a slower decline afterward, although the authors caution the temporal analysis is qualitative.
  • Successful recognitions cluster spatially around visually distinctive, persistent features such as coral colonies and rock-sand transitions, not uniformly across the seafloor.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If footprint overlap is the right standard of 'place', then prior underwater and some terrestrial VPR benchmarks built on distance thresholds are likely optimistic; re-evaluating them with footprint-style ground truth could change model rankings.
  • The per-image affine fusion of monocular and stereo range is testable: comparing fused footprint corners against a known flat area of seafloor with measured altitude would directly quantify footprint error, and such validation could strengthen or correct the benchmark.
  • The footprint IoU is a continuous quantity; treating it as a soft label or weighting retrievals by IoU could yield more informative metrics than a binary threshold at 0.07.
  • Because reliable recognition clusters in distinctive patches, a robot could plan surveys to maximize overlap with previously recognized persistent features rather than relying on uniform single-image retrieval.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The paper presents a curated AUV image dataset for long-term visual localization in photic-zone benthic environments, comprising five reference sites revisited over up to six years, with raw and color-corrected stereo imagery, calibrations, and poses registered across visits. The central methodological contribution is a footprint-based ground-truthing procedure: for each image, a fused range map (stereo HitNet metric depth aligned with Depth Anything V2 relative depth via a per-image global affine transform) is used to estimate 3D seafloor footprint corners; camera-view links are then defined by footprint overlap with an IoU threshold. The authors benchmark eight visual place recognition (VPR) models on this ground truth, reporting substantially lower Recall@K than on terrestrial and previous underwater benchmarks, and compare footprint-based with location-based ground truth, arguing that distance thresholds overestimate VPR recall on rugged sites.

Significance. The dataset directly addresses a real need: there is no existing multi-site, multi-year photic-zone benthic benchmark for long-term VPR, and the footprint-overlap idea is a principled alternative to fixed distance thresholds for near-nadir imagery. The benchmark itself is useful, and the R@10 vs. IR@10 comparison exposes a metric-pitfall that deserves attention. However, the trustworthiness of all reported benchmark numbers rests on the geometric validity of the footprint links. The current manuscript does not provide independent validation of the affine range-fusion assumption or of the absolute accuracy of the registered poses; both are load-bearing. If those validations are added, the dataset and protocol would be a solid contribution to underwater visual localization.

major comments (3)
  1. [§2.4.2, Eq. (6)] The entire footprint-based ground truth relies on the assumption that Depth Anything V2 relative range is related to metric stereo range by a single global affine transform per image, z_fused = a*z_rel + b. This is never validated against any independent seafloor geometry. The paper reports no fit residuals, no distribution of (a,b) values, and no comparison of estimated footprint polygons with, for example, the MVS point clouds produced in §2.3 or with known ground control features. Underwater attenuation and backscatter, and the several-meter range variation within a single field of view noted for Site 4 (Fig. 4), can break this linearity, particularly near image borders where the footprint corners are defined. Since all camera-view links (Eqs. 9–11) and all downstream Recall@K values inherit the corner-range errors, this is the weakest load-bearing link in the paper. Please add a quan
  2. [§2.3, Eq. (1) and §3.1] The registration-error statistic e_ij is computed on the correspondences from the final colored-ICP stage. These are the same correspondences used to estimate the transform, so the reported errors are fitting residuals, not an independent measure of pose accuracy. The abstract claims “sub-decimeter registered camera poses,” but the manuscript only demonstrates that the ICP residuals are small (99th percentile < 0.16 m). Real pose errors could be larger due to low-frequency warping, unmodeled scale drift, or scene change, none of which is captured by this metric. Because the footprint-IoU threshold τ_f is designed to tolerate a specific registration error t_e = 0.16 m, an underestimation of true pose error would make the ground-truth links less reliable. Please validate pose accuracy with an independent method, or at least rephrase the claims to state explicitly that they refer to point-c
  3. [§3.4, Table 4] The comparison between footprint-based and location-based ground truth sets the location threshold to the 95th percentile of distances between footprint-linked camera views. With this construction, the location-based ground truth is by design a superset of the footprint-based links, and R@K is non-decreasing in the number of links. The observation that R@10 increases is therefore partly a consequence of the threshold choice. The paper acknowledges this in the limitations but still states as a main conclusion that “distance-threshold ground-truthing can overestimate VPR Recall@K.” To support that general claim, the authors should sweep over a range of distance thresholds (or use a physically motivated fixed threshold) and show that the qualitative conclusion is not an artifact of the single 95th-percentile choice. At minimum, the conclusion should be framed as a demonstration that R@K is
minor comments (8)
  1. [Abstract and §3.1] The phrase “sub-decimeter registered camera poses” is stronger than what is supported. The 99th-percentile registration residual is 0.16 m. Please harmonize the wording with the actual metric.
  2. [§2.4.2, Eq. (5)] The Huber loss ρ is not defined. Please define it or provide a reference.
  3. [Eq. (6) and surrounding text] The fused-range notation appears as “zf used” with a spacing artifact. Please use z_fused consistently.
  4. [§2.5.2, Eq. (12)] The description of the simplified scenario for τ_f is ambiguous. The text says both cameras observe a flat patch, but Eq. (12) corresponds to the apparent IoU induced between non-overlapping footprints under a translation error. Please clarify the geometry in words and define whether t_e is the translation error or the produced overlap width.
  5. [Table 3 and references] CosPlace is cited as [48], which is the distributed-training paper, not the original CosPlace paper [41]. Please cite the original architecture when reporting the model name.
  6. [§2.5.2 and Table 2] The terms “valid query” and “query coverage overlap” are used without formal definitions in the main text. A short definition next to Eq. (14) or in Table 2 would improve reproducibility.
  7. [Dataset availability] The paper describes a curated dataset but gives no URL, DOI, or detailed access instructions. A public dataset is essential for a benchmark paper; please add a data-availability statement.
  8. [§2.4.3] The 2D projection of 3D footprints is justified qualitatively. Given the several-meter range variation at Site 4, a sentence quantifying the expected projection error or a sensitivity test would be useful.

Circularity Check

0 steps flagged

No significant circularity: the footprint-based ground truth is geometric and independent of the VPR descriptors being benchmarked; self-citations are non-load-bearing.

full rationale

The paper's core derivation chain is self-contained rather than circular. Ground-truth links are generated geometrically: inverse projection (Eqs. 4a-4b) with median corner ranges from fused stereo/monocular range maps (Sec. 2.4.2), footprint overlap tests (Eq. 9), and an IoU threshold derived from a simplified footprint model and registration-error statistics (Eq. 12, Sec. 3.2). None of these steps is fitted to, or optimized against, the VPR descriptors or Recall@K results; the affine range fusion in Eq. 6 is fit to stereo ranges, not to retrieval outcomes. The paper cites prior work by the authors for dataset clustering [7], footprint-based loop-closure ideas [8,9], and color correction [23], but these citations do not carry the central benchmark claim or forbid alternative methods. The location-based comparison is explicitly framed as a consequence of the recall definition: the paper states that R@K is non-decreasing as ground-truth links are added, so the direction of the R@10 difference is mathematically expected; the magnitudes and IR@10 patterns remain data-dependent. The main weakness is empirical, not circular: the global affine alignment between monocular relative range and stereo metric range (Eq. 6) is not independently validated against known seafloor geometry, so footprint accuracy and hence the ground-truth links could be biased in underwater conditions. That is a correctness/robustness risk, not a reduction of the derivation to its own inputs. Overall, no load-bearing circular step was found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central benchmark depends on a chain of geometric estimates: camera poses from SFM/registration, range maps from stereo+monocular fusion, and footprints projected to 2D. The most load-bearing free parameters are the per-image scale/offset (a,b) and the hand-derived IoU threshold tau_f. No new physical entities are introduced. The gray-world intensity targets and registration voxel sizes are additional hand-chosen constants.

free parameters (5)
  • Gray-world target mean/std (mu_y, sigma_y) = 0.35 / 0.12
    Chosen empirically in Section 2.2 to balance brightness/contrast; changes all input images to the VPR models.
  • Stereo range mask interval = [0.2, 6.0] m
    Hand-chosen in Section 2.4.2 to remove close-range artifacts and far-range noise before fitting the scale/offset.
  • Per-image range fusion scale/offset (a, b) = not reported
    Robust regression fitted to stereo ranges (Eq. 5-6); directly determines footprint sizes but no fitted values or validation are reported.
  • Footprint IoU threshold tau_f = 0.07
    Derived in Section 3.2 from Eq. 12 using FOV_y=34 deg, a=2.0 m, te=0.16 m; te is the 99th percentile registration error, making the threshold data-dependent.
  • Location-based distance threshold = 1.33-2.48 m per visit pair
    Set to the 95th percentile of distances between footprint-linked views (Section 3.4); this choice shapes the GT comparison and the overestimation result.
axioms (5)
  • standard math Ideal pinhole camera model without skew (Eq. 2) with intrinsics from SFM calibration.
    Standard projective model; calibration quality is assumed from Metashape and stereo rig calibration.
  • domain assumption Monocular relative depth (Depth Anything V2) is related to metric stereo depth by a single global scale and offset per image (Eq. 6).
    This is the core of range fusion in Section 2.4.2; if the affine assumption fails in turbid or non-linearly attenuated underwater imagery, footprint corners and all ground-truth links are wrong. Not independently validated.
  • domain assumption Projecting 3D footprints to a 2D horizontal plane by discarding height is acceptable (Section 2.4.3).
    Relies on max site elevation difference ~9 m and near-nadir views; the authors note it breaks down for cliffs, steep walls, and caves.
  • domain assumption Registration errors are representable by a horizontal translation below te=0.16 m in the simplified IoU model (Eq. 12).
    Used to set tau_f≈0.07. Real errors include rotation and local warping, so the threshold is an approximation.
  • domain assumption Selected sites are representative of dynamic benthic environments despite excluding kelp-dominated, highly dynamic cover (Section 2.1).
    Sites were chosen for reconstruction quality and excluded the most dynamic habitats, limiting the generality of the 'dynamic benthic' claim.

pith-pipeline@v1.3.0-alltime-deepseek · 26963 in / 13020 out tokens · 125441 ms · 2026-08-02T18:55:45.977879+00:00 · methodology

0 comments
read the original abstract

Long-term visual localization has the potential to reduce cost and improve mapping quality in optical benthic monitoring with autonomous underwater vehicles (AUVs). Despite this potential, long-term visual localization in benthic environments remains understudied, primarily due to the lack of curated datasets for benchmarking. Moreover, limited georeferencing accuracy and image footprints necessitate precise geometric information for accurate ground-truthing. In this work, we address these gaps by presenting a curated dataset for long-term visual localization in benthic environments and a novel method to ground-truth visual localization results for near-nadir underwater imagery. Our dataset comprises georeferenced AUV imagery from five benthic reference sites, revisited over periods up to six years, and includes raw and color-corrected stereo imagery, camera calibrations, and sub-decimeter registered camera poses. To our knowledge, this is the first curated underwater dataset for long-term visual localization spanning multiple sites and photic-zone habitats. Our ground-truthing method estimates 3D seafloor image footprints and links camera views with overlapping footprints, ensuring that ground-truth links reflect shared visual content. Building on this dataset and ground truth, we benchmark eight state-of-the-art visual place recognition (VPR) methods and find that Recall@K is significantly lower on our dataset than on established terrestrial and underwater benchmarks. Finally, we compare our footprint-based ground truth to a traditional location-based ground truth and show that distance-threshold ground-truthing can overestimate VPR Recall@K at sites with rugged terrain and altitude variations. Together, the curated dataset, ground-truthing method, and VPR benchmark provide a stepping stone for advancing long-term visual localization in dynamic benthic environments.

Figures

Figures reproduced from arXiv: 2603.04056 by Martin Kvisvik Larsen, Oscar Pizarro.

Figure 1
Figure 1. Figure 1: Map with overview of the benthic reference sites in our dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (A) Raw (left) and color corrected (right) image from the 2010 visit to Site 1. (B) Raw (left) and color corrected [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Single line diagram of the geometric reconstruction and registration workflow for an example with two visits. In the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An example showing the fusion of relative and metric stereo-derived range maps for an image from 2009 visit to Site [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of a 2D simplified model of our footprint estimation method for two scenarios. Panel (A) shows a scenario [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Distribution of geometric registration errors across the benthic reference sites in the dataset. The legends are on the [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean Recall@K across all visit pairs for each site in the dataset. Results are based on the footprint-based ground truth [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual place recognition results at Site 1 rendered on top of an orthomosaic derived from the 2010 images, using [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual place recognition results at Site 2 rendered on top of an orthomosaic derived from the 2011 images, using [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Recall@10 for all VPR models for varying revisit intervals to the sites in the dataset. The overall trend shows a [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Scatter plots of (A) R@10 and (B) IR@10 for the footprint-based and location-based ground truth. The R@10 is [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visual place recognition results at Site 3 rendered on top of an orthomosaic derived from the 2010 images, using [PITH_FULL_IMAGE:figures/full_fig_p031_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visual place recognition results at Site 4 rendered on top of an orthomosaic derived from the 2009 images, using [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Visual place recognition results at Site 5 rendered on top of an orthomosaic derived from the 2010 images, using [PITH_FULL_IMAGE:figures/full_fig_p033_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Examples of very hard, hard, and easy query images from each benthic reference site. The very hard queries are not [PITH_FULL_IMAGE:figures/full_fig_p034_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Examples illustrating disagreement between footprint-based and location-based linking of query–database camera [PITH_FULL_IMAGE:figures/full_fig_p035_16.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 6 linked inside Pith

  1. [1]

    Benjamin Misiuk and Craig J. Brown. Benthic habitat mapping: A review of three decades of mapping biological patterns on the seafloor.Estuarine, Coastal and Shelf Science, 296:108599, January 2024

    2024

  2. [2]

    Survey of underwater robot positioning navigation.Applied Ocean Research, 90:101845, September 2019

    Yinghao Wu, Xuxiang Ta, Ruichao Xiao, Yaoguang Wei, Dong An, and Daoliang Li. Survey of underwater robot positioning navigation.Applied Ocean Research, 90:101845, September 2019

    2019

  3. [3]

    In-situ alignment calibration of attitude and ultra short baseline sensors for precision underwater positioning.Ocean Engineering, 35(14):1448–1462, October 2008

    Hsin-Hung Chen. In-situ alignment calibration of attitude and ultra short baseline sensors for precision underwater positioning.Ocean Engineering, 35(14):1448–1462, October 2008

    2008

  4. [4]

    Study on Installation Error Analysis and Calibration of Acoustic Transceiver Array Based on SINS/USBL Integrated System.IEEE Access, 6:66923–66939, 2018

    Tong Jinwu, Xu Xiaosu, Zhang Tao, Zhang Liang, and Li Yao. Study on Installation Error Analysis and Calibration of Acoustic Transceiver Array Based on SINS/USBL Integrated System.IEEE Access, 6:66923–66939, 2018

    2018

  5. [5]

    A Calibration Method of USBL Installation Error Based on Attitude Determination.IEEE Transactions on Vehicular Technology, 69(8):8317–8328, August 2020

    Yongyun Zhu, Tao Zhang, Shouquan Xu, Hyo-Sang Shin, Peijuan Li, Bonan Jin, Liang Zhang, Chengcheng Weng, and Yao Li. A Calibration Method of USBL Installation Error Based on Attitude Determination.IEEE Transactions on Vehicular Technology, 69(8):8317–8328, August 2020

    2020

  6. [6]

    Automated registration for multi-year robotic surveys of marine benthic habitats

    Mitch Bryson, Matthew Johnson-Roberson, Oscar Pizarro, and Stefan Williams. Automated registration for multi-year robotic surveys of marine benthic habitats. In2013 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 3344–3349, November 2013

    2013

  7. [7]

    Geometric Registration of Benthic Imagery for Learning Appearance-Based Place Recognition over Multiple Sessions

    Martin Kvisvik Larsen, Oscar Pizarro, and Martin Ludvigsen. Geometric Registration of Benthic Imagery for Learning Appearance-Based Place Recognition over Multiple Sessions. In2023 IEEE Underwater Technology (UT), pages 1–10, March 2023

    2023

  8. [8]

    Williams, Oscar Pizarro, and Matthew Johnson-Roberson

    Ian Mahon, Stefan B. Williams, Oscar Pizarro, and Matthew Johnson-Roberson. Efficient View-Based SLAM Using Visual Loop Closures.IEEE Transactions on Robotics, 24(5):1002–1014, October 2008

    2008

  9. [9]

    Eustice, Oscar Pizarro, and Hanumant Singh

    Ryan M. Eustice, Oscar Pizarro, and Hanumant Singh. Visually Augmented Navigation for Autonomous Underwater Vehicles.IEEE Journal of Oceanic Engineering, 33(2):103–122, April 2008

    2008

  10. [10]

    Autonomous Underwater Vehicle navigation: A review

    Bingbing Zhang, Daxiong Ji, Shuo Liu, Xinke Zhu, and Wen Xu. Autonomous Underwater Vehicle navigation: A review. Ocean Engineering, 273:113861, April 2023

    2023

  11. [11]

    Williams, Maria Byrne, and Will Figueira

    Renata Ferrari, Mitch Bryson, Tom Bridge, Julie Hustache, Stefan B. Williams, Maria Byrne, and Will Figueira. Quantifying the response of structural complexity and community composition to environmental change in marine communities.Global Change Biology, 22(5):1965–1975, 2016. 26 Larsen et al. LONG-TERMVISUALLOCALIZATION INBENTHICENVIRONMENTS

    1965

  12. [12]

    Yifan Song, David Nakath, Mengkun She, and Kevin Köser. Optical Imaging and Image Restoration Techniques for Deep Ocean Mapping: A Comprehensive Survey.PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 90(3):243–267, June 2022

    2022

  13. [13]

    Long-Term Visual Localization Revisited.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2074–2088, April 2022

    Carl Toft, Will Maddern, Akihiko Torii, Lars Hammarstrand, Erik Stenborg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, Fredrik Kahl, and Torsten Sattler. Long-Term Visual Localization Revisited.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2074–2088, April 2022

    2074

  14. [14]

    Eiffel Tower: A deep-sea underwater dataset for long-term visual localization.The International Journal of Robotics Research, 42(9):689–699, August 2023

    Clémentin Boittiaux, Claire Dune, Maxime Ferrera, Aurélien Arnaubec, Ricard Marxer, Marjolaine Matabos, Loïc Van Audenhaege, and Vincent Hugel. Eiffel Tower: A deep-sea underwater dataset for long-term visual localization.The International Journal of Robotics Research, 42(9):689–699, August 2023

    2023

  15. [15]

    Rachel H. X. Wong, Jacquomo Monk, Nicholas R. Perkins, and Neville S. Barrett. A systematic review on the anthropogenic stressors on sessile benthic mesophotic reef communities: implications for temperate reef management in Australia. Frontiers in Marine Science, 10, December 2023

    2023

  16. [16]

    Perkins, Jacquomo Monk, Rachel H

    Nicholas R. Perkins, Jacquomo Monk, Rachel H. X. Wong, and Neville S. Barrett. Temporal variability in temperate mesophotic ecosystems revealed with over a decade of monitoring with an autonomous underwater vehicle.Marine Environmental Research, 209:107179, August 2025

    2025

  17. [17]

    A. G. Glover, A. J. Gooday, D. M. Bailey, D. S. M. Billett, P. Chevaldonne, Ana Colaco, J. Copley, Daphne Cuvelier, Daniel Desbruyeres, V . Kalogeropoulou, M. Klages, N. Lampadariou, C. Lejeusne, Nelia Mestre, G. L. J. Paterson, T. Perez, H. Ruhl, Jozee Sarrazin, T. Soltwedel, E. H. Soto, S. Thatje, A. Tselepides, S. van Gaever, and A. Vanreusel. Temporal...

    2010

  18. [18]

    Image-Based Relocalization and Alignment for Long-Term Monitoring of Dynamic Underwater Environments

    Beverley Gorry, Tobias Fischer, Michael Milford, and Alejandro Fontan. Image-Based Relocalization and Alignment for Long-Term Monitoring of Dynamic Underwater Environments. In2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10749–10756, October 2025. ISSN: 2153-0866

    2025

  19. [19]

    Williams, Oscar R

    Stefan B. Williams, Oscar R. Pizarro, Michael V . Jakuba, Craig R. Johnson, Neville S. Barrett, Russell C. Babcock, Gary A. Kendrick, Peter D. Steinberg, Andrew J. Heyward, Peter J. Doherty, Ian Mahon, Matthew Johnson-Roberson, Daniel Steinberg, and Ariell Friedman. Monitoring of Benthic Reference Sites: Using an Autonomous Underwater Vehicle.IEEE Robotic...

    2012

  20. [20]

    Williams, Michael V

    Oscar Pizarro, Stefan B. Williams, Michael V . Jakuba, Matthew Johnson-Roberson, Ian Mahon, Mitch Bryson, Daniel Steinberg, Ariell Friedman, Donald Dansereau, Navid Nourani-Vatani, Daniel Bongiorno, Michael Bewley, Asher Bender, Nasir Ashan, and Bertrand Douillard. Benthic monitoring with robotic platforms — The experience of Australia. In2013 IEEE Intern...

    2013

  21. [21]

    Squidle+: a collaborative platform to manage, discover and annotate marine imagery.Frontiers in Marine Science, 12, November 2025

    Ariell Friedman, Jacquomo Monk, Oscar Pizarro, Dhugal John Lindsay, Elizabeth Oh, Blair Thornton, Andrew Gerard Carroll, Rachel Przeslawski, and Stefan Bernard Williams. Squidle+: a collaborative platform to manage, discover and annotate marine imagery.Frontiers in Marine Science, 12, November 2025

    2025

  22. [22]

    Australian Statistical Geography Standard (ASGS) Edition 3: Digital Boundary Files, 2021

    Australian Bureau of Statistics. Australian Statistical Geography Standard (ASGS) Edition 3: Digital Boundary Files, 2021

    2021

  23. [23]

    Williams

    Mitch Bryson, Matthew Johnson-Roberson, Oscar Pizarro, and Stefan B. Williams. Colour-Consistent Structure-from- Motion Models Using Underwater Imagery. InRobotics: Science and Systems VIII, pages 33–40. MIT Press, 2013

    2013

  24. [24]

    Buchsbaum

    G. Buchsbaum. A spatial processor model for object colour perception.Journal of the Franklin Institute, 310(1):1–26, July 1980

    1980

  25. [25]

    Agisoft Metashape Professional, 2022

    Agisoft LLC. Agisoft Metashape Professional, 2022

    2022

  26. [26]

    S. Umeyama. Least-squares estimation of transformation parameters between two point patterns.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(4):376–380, April 1991

    1991

  27. [27]

    Fast Point Feature Histograms (FPFH) for 3D registration

    Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast Point Feature Histograms (FPFH) for 3D registration. In2009 IEEE International Conference on Robotics and Automation, pages 3212–3217, May 2009. 27 Larsen et al. LONG-TERMVISUALLOCALIZATION INBENTHICENVIRONMENTS

    2009

  28. [28]

    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.Communications of the ACM, 24(6):381–395, June 1981

    1981

  29. [29]

    Colored Point Cloud Registration Revisited

    Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Colored Point Cloud Registration Revisited. InProceedings of the IEEE International Conference on Computer Vision, pages 143–152, 2017

    2017

  30. [30]

    Open3D: A Modern Library for 3D Data Processing, January 2018

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A Modern Library for 3D Data Processing, January 2018. arXiv:1801.09847 [cs]

    Pith/arXiv arXiv 2018

  31. [31]

    Fossen.Handbook of Marine Craft Hydrodynamics and Motion Control

    Thor I. Fossen.Handbook of Marine Craft Hydrodynamics and Motion Control. John Wiley & Sons, Chichester, 1 edition, May 2011

    2011

  32. [32]

    Cambridge University Press, Cambridge, 2 edition, 2003

    Richard Hartley and Andrew Zisserman.Multiple View Geometry in Computer Vision. Cambridge University Press, Cambridge, 2 edition, 2003

    2003

  33. [33]

    HITNet: Hierarchi- cal Iterative Tile Refinement Network for Real-time Stereo Matching

    Vladimir Tankovich, Christian Hane, Yinda Zhang, Adarsh Kowdle, Sean Fanello, and Sofien Bouaziz. HITNet: Hierarchi- cal Iterative Tile Refinement Network for Real-time Stereo Matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14362–14372, 2021

    2021

  34. [34]

    Depth Anything V2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth Anything V2. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 21875–21911. Curran Associates, Inc., 2024

    2024

  35. [35]

    HybridDepth: Robust Metric Depth Fusion by Leveraging Depth from Focus and Single-Image Priors

    Ashkan Ganj, Hang Su, and Tian Guo. HybridDepth: Robust Metric Depth Fusion by Leveraging Depth from Focus and Single-Image Priors. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 973–982, February 2025

    2025

  36. [36]

    Williams

    Hongjie Zhang, Gideon Billings, and Stefan B. Williams. SPADE: Sparsity Adaptive Depth Estimator for Zero-Shot, Real-Time, Monocular Depth Estimation in Underwater Environments, October 2025. arXiv:2510.25463 [cs] version: 1

    arXiv 2025

  37. [37]

    Shapely: Manipulation and analysis of geometric objects, 2023

    Sean Gillies and Shapely contributors. Shapely: Manipulation and analysis of geometric objects, 2023

    2023

  38. [38]

    Open Source Geospatial Foundation, 2025

    GEOS contributors.GEOS computational geometry library. Open Source Geospatial Foundation, 2025

    2025

  39. [39]

    NetVLAD: CNN Architecture for Weakly Supervised Place Recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN Architecture for Weakly Supervised Place Recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5297–5307, 2016

    2016

  40. [40]

    MixVPR: Feature Mixing for Visual Place Recognition

    Amar Ali-bey, Brahim Chaib-draa, and Philippe Giguère. MixVPR: Feature Mixing for Visual Place Recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2998–3007, 2023

    2023

  41. [41]

    Rethinking Visual Geo-Localization for Large-Scale Applications

    Gabriele Berton, Carlo Masone, and Barbara Caputo. Rethinking Visual Geo-Localization for Large-Scale Applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4878–4888, 2022

    2022

  42. [42]

    DINOv2: Learning Robust Visual Features without Supervision, February 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 Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick La...

    Pith/arXiv arXiv 2024

  43. [43]

    AnyLoc: Towards Universal Visual Place Recognition.IEEE Robotics and Automation Letters, 9(2):1286–1293, February 2024

    Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava Krishna, and Sourav Garg. AnyLoc: Towards Universal Visual Place Recognition.IEEE Robotics and Automation Letters, 9(2):1286–1293, February 2024

    2024

  44. [44]

    Close, But Not There: Boosting Geographic Distance Sensitivity in Visual Place Recognition

    Sergio Izquierdo and Javier Civera. Close, But Not There: Boosting Geographic Distance Sensitivity in Visual Place Recognition. In Aleš Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and Gül Varol, editors, Computer Vision – ECCV 2024, pages 240–257, Cham, 2025. Springer Nature Switzerland

    2024

  45. [45]

    Optimal Transport Aggregation for Visual Place Recognition

    Sergio Izquierdo and Javier Civera. Optimal Transport Aggregation for Visual Place Recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17658–17668, 2024. 28 Larsen et al. LONG-TERMVISUALLOCALIZATION INBENTHICENVIRONMENTS

    2024

  46. [46]

    MegaLoc: One Retrieval to Place Them All

    Gabriele Berton and Carlo Masone. MegaLoc: One Retrieval to Place Them All. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2861–2867, 2025

    2025

  47. [47]

    The Faiss library, October 2025

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The Faiss library, October 2025. arXiv:2401.08281 [cs]

    Pith/arXiv arXiv 2025

  48. [48]

    Distributed training of CosPlace for large-scale visual place recognition.Frontiers in Robotics and AI, 11:1386464, May 2024

    Riccardo Zaccone, Gabriele Berton, and Carlo Masone. Distributed training of CosPlace for large-scale visual place recognition.Frontiers in Robotics and AI, 11:1386464, May 2024

    2024

  49. [49]

    EigenPlaces: Training Viewpoint Robust Models for Visual Place Recognition

    Gabriele Berton, Gabriele Trivigno, Barbara Caputo, and Carlo Masone. EigenPlaces: Training Viewpoint Robust Models for Visual Place Recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 11080–11090, 2023

    2023

  50. [50]

    A review of non-rigid transformations and learning-based 3D point cloud registration methods.ISPRS Journal of Photogrammetry and Remote Sensing, 196:58–72, February 2023

    Sara Monji-Azad, Jürgen Hesser, and Nikolas Löw. A review of non-rigid transformations and learning-based 3D point cloud registration methods.ISPRS Journal of Photogrammetry and Remote Sensing, 196:58–72, February 2023

    2023

  51. [51]

    Strategies for large scale elastic and semantic LiDAR reconstruction.Robotics and Autonomous Systems, 155:104185, September 2022

    Yiduo Wang, Milad Ramezani, Matias Mattamala, Sundara Tejaswi Digumarti, and Maurice Fallon. Strategies for large scale elastic and semantic LiDAR reconstruction.Robotics and Autonomous Systems, 155:104185, September 2022

    2022

  52. [52]

    Where is your place, Visual Place Recognition? InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4416–4425, August 2021

    Sourav Garg, Tobias Fischer, and Michael Milford. Where is your place, Visual Place Recognition? InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence, pages 4416–4425, August 2021. arXiv:2103.06443 [cs]

    Pith/arXiv arXiv 2021

  53. [53]

    On the Relationship between Self-Attention and Convolu- tional Layers, January 2020

    Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. On the Relationship between Self-Attention and Convolu- tional Layers, January 2020. arXiv:1911.03584 [cs]

    Pith/arXiv arXiv 2020

  54. [54]

    Eustice, and Matthew Johnson-Roberson

    Jie Li, Ryan M. Eustice, and Matthew Johnson-Roberson. Underwater robot visual place recognition in the presence of dramatic appearance change. InOCEANS 2015 - MTS/IEEE Washington, pages 1–6, October 2015

    2015

  55. [55]

    Steinberg, Stefan B

    Daniel M. Steinberg, Stefan B. Williams, Oscar Pizarro, and Michael V . Jakuba. Towards autonomous habitat classification using Gaussian Mixture Models. In2010 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4424–4431, October 2010

    2010

  56. [56]

    Steinberg, Oscar Pizarro, and Stefan B

    Daniel M. Steinberg, Oscar Pizarro, and Stefan B. Williams. Hierarchical Bayesian models for unsupervised scene understanding.Computer Vision and Image Understanding, 131:128–144, February 2015

    2015

  57. [57]

    Online Clustering of Seafloor Imagery for Interpretation during Long-Term AUV Operations, September 2025

    Cailei Liang, Adrian Bodenmann, Sam Fenton, and Blair Thornton. Online Clustering of Seafloor Imagery for Interpretation during Long-Term AUV Operations, September 2025. arXiv:2509.06678 [cs]

    Pith/arXiv arXiv 2025

  58. [58]

    From images to scenes: Compressing an image cluster into a single scene model for place recognition

    Edward Johns and Guang-Zhong Yang. From images to scenes: Compressing an image cluster into a single scene model for place recognition. In2011 International Conference on Computer Vision, pages 874–881, November 2011

    2011

  59. [59]

    Using Image Sequences for Long-Term Visual Localization

    Erik Stenborg, Torsten Sattler, and Lars Hammarstrand. Using Image Sequences for Long-Term Visual Localization. In 2020 International Conference on 3D Vision (3DV), pages 938–948, November 2020

    2020

  60. [60]

    The localization problem for underwater vehicles: An overview of operational solutions.Ocean Engineering, 330:121173, June 2025

    Fredrik Fogh Sørensen, Christian Mai, Malte von Benzon, Jesper Liniger, and Simon Pedersen. The localization problem for underwater vehicles: An overview of operational solutions.Ocean Engineering, 330:121173, June 2025

    2025

  61. [61]

    Ford, Gabe Bolton, Iman Shames, and Michael Milford

    Owen Claxton, Connor Malone, Helen Carson, Jason J. Ford, Gabe Bolton, Iman Shames, and Michael Milford. Improving Visual Place Recognition Based Robot Navigation by Verifying Localization Estimates.IEEE Robotics and Automation Letters, 9(12):11098–11105, December 2024. 29 Larsen et al. LONG-TERMVISUALLOCALIZATION INBENTHICENVIRONMENTS Supplementary Mater...

    2024