Pith. sign in

REVIEW 1 major objections 9 minor 39 references

APVI-SLAM: Real-Time Acoustic-Pressure-Visual-Inertial Localization and Photorealistic Mapping System in Complex Underwater Environment

T0 review · 1 major / 9 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Underwater SLAM survives visual blackouts by swapping in acoustic and pressure estimates

desk verdict Solid engineering, but the core contribution is not isolated from the sensor advantage. read the letter →

arxiv 2607.06222 v1 pith:XK6DRGWC submitted 2026-07-07 cs.RO

Hanwen Zhang , Yipeng Zhu , Xiaopeng Guo , Huajian Huang , Sai-Kit Yeung This is my paper · ORCID
classification cs.RO
keywords underwaterSLAMmulti-sensorfusion3DGaussianSplattingDVLvisual-inertialodometryreliability-awareestimationmappingwater-mediummodeling
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

APVI-SLAM is a real-time system for underwater robot localization and photorealistic mapping that solves a core problem: when underwater visibility degrades, visual-inertial SLAM systems lose tracking, drift, or require expensive re-initialization. The paper's central mechanism is a reliability-aware fusion framework that runs two parallel estimators — a Visual-Inertial Odometry (VIO) subsystem using camera and IMU data, and a DVL-Inertial-Pressure Odometry (DIPO) subsystem using acoustic velocity logs, IMU, and pressure depth readings. A coarse-to-fine self-discrimination algorithm continuously evaluates which estimator is trustworthy by checking feature counts, inter-estimator agreement, and Hessian eigenvalue distributions. When VIO degrades, the system either freezes its state in a sliding window (preserving scale and prior information for fast recovery once features return) or drops it entirely, falling back on DIPO output. This prevents corrupted visual data from poisoning the overall state estimate. For mapping, the system uses a quadtree-guided 3D Gaussian Splatting scheme with online water-medium modeling: it partitions images by intensity similarity to guide depth densification in texture-poor regions, and jointly optimizes underwater optical parameters (attenuation and backscatter) to correct color distortion from turbidity. The paper also contributes a coral reef surveying dataset with synchronized multi-modal sensor data for benchmarking underwater mapping. Experiments on a public tank dataset, a new simulation benchmark, and the real coral reef dataset show that APVI-SLAM maintains sub-30cm trajectory accuracy in conditions where all visual-only baselines fail, while producing higher-fidelity renderings (PSNR 31-32 on simulation sequences) at real-time speeds.

What carries the argument

Coarse-to-fine self-discrimination algorithm (Alg. 1): feature count >50 and stable ratio >0.2 as coarse gate; inter-estimator offset <30cm and angular offset <10° as medium gate; 6-DOF Hessian eigenvalue thresholds from semicircle distribution as fine gate. Sliding-window freezing strategy: preserves VIO state variables and scale during visual blackout for rapid re-initialization. Quadtree-guided densification: 3-level quadtree (16×16, 8×8, 6×6) on intensity images to guide depth sampling via photometric cost and belief propagation. Water-medium model: jointly optimized attenuation (β_D) and backscatter (β_B) parameters with doubled learning rate during VIO failure.

What would settle it

If the coarse-to-fine self-discrimination thresholds (feature count, offset bounds, Hessian eigenvalue bounds) fail to generalize to a new underwater environment with different turbidity, lighting, or sensor mounting — causing the system to either trust a degraded VIO or unnecessarily discard a functional one — the reliability-aware fusion would produce worse localization than a simple always-fuse or always-DIPO baseline.

Watch

Extended reading notes

Core claim

The paper demonstrates that the failure mode of underwater visual-inertial SLAM — persistent visual degradation corrupting the fused state estimate — can be addressed by running a parallel acoustic-inertial-pressure estimator (DIPO) alongside the visual-inertial estimator (VIO), and dynamically switching between them based on a multi-level reliability assessment. The coarse-to-fine self-discrimination mechanism (Algorithm 1) is the load-bearing component: it uses feature point counts and stability as a coarse gate, inter-estimator positional/angular agreement as a medium gate, and Hessian eigenvalue thresholds (derived from semicircle distribution modeling on simulation data) as a fine gate.

Load-bearing premise

The coarse-to-fine self-discrimination mechanism uses fixed numerical thresholds — 50 feature points, 20% stable ratio, 30cm positional offset, 10-degree angular offset, and Hessian eigenvalue bounds from simulation — to decide when to trust VIO versus DIPO. These thresholds are assumed to work across different underwater environments, camera setups, and sensor configurations, but the paper provides no sensitivity analysis or adaptive mechanism for them.

Editorial extensions

If this is right

  • Autonomous underwater vehicles could operate in turbid, low-texture environments (coral reefs, shipwrecks, seabed infrastructure) without human intervention for tracking recovery, since the freezing strategy bridges visual gaps of seconds to minutes.
  • The parallel-estimator-with-reliability-switching architecture could generalize to other domains where one sensor modality is intermittently unreliable — e.g., GPS-denied urban navigation with vision/LiDAR fallback, or space robotics where lighting changes cause visual dropout.
  • The coral reef surveying dataset fills a gap in underwater mapping benchmarks and could accelerate development of marine robotics systems for environmental monitoring and reef health assessment.
  • Online water-medium parameter optimization during VIO failure suggests that acoustic-only navigation phases can still contribute to map quality, making the mapping process continuous rather than intermittent.

Reading between the lines

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

  • The hardcoded thresholds in Algorithm 1 (feature count >50, stable ratio >0.2, 30cm/10° offsets, Hessian eigenvalue bounds) are derived from specific simulation and tank environments. Whether these transfer to different water types (open ocean vs. coastal vs. turbid estuary), different camera configurations, or different DVL mounting geometries is untested. An adaptive or learned threshold mechani
  • The semicircle law approximation for Hessian eigenvalue distributions assumes the Hessian behaves like a random symmetric matrix. In structured environments with repetitive textures (e.g., pipeline inspection), the eigenvalue distribution may deviate from semicircle, potentially causing false positives or negatives in the fine-grained VIO stability check.
  • The system's mapping quality depends on DIPO providing sufficiently accurate poses during VIO blackout. Since DIPO uses an Error-State Kalman Filter with DVL velocity and pressure depth observations, its drift characteristics over extended visual gaps (minutes rather than seconds) are not characterized in the paper but would determine the maximum operational visual blackout duration.
  • The quadtree-based densification assumes intensity similarity implies depth similarity. In environments with strong but non-depth-correlated intensity variation (e.g., caustics from surface light, biological color variation), this assumption may produce incorrect depth priors in specific regions.
Share X Bluesky LinkedIn Reddit HN

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

1 major / 9 minor

Summary. This paper presents APVI-SLAM, a real-time underwater SLAM system that fuses monocular visual-inertial odometry (VIO) with DVL-inertial-pressure odometry (DIPO) through a reliability-aware framework. The system introduces a coarse-to-fine self-discrimination mechanism to switch between or reweight the two estimators, a sliding-window freezing strategy for rapid VIO recovery after visual dropout, and a quadtree-guided 3D Gaussian Splatting mapping module with online water-medium modeling. The authors also contribute a coral reef surveying dataset. Evaluation is conducted on a public Tank dataset, a new simulation benchmark, and the new real-world coral reef dataset, with comparisons against several SLAM baselines.

Significance. The paper addresses a practically important problem—robust underwater SLAM and photorealistic mapping under severe visual degradation—and proposes a complete, real-time system. The sliding-window freezing strategy (Table IV: 100% vs 21% re-initialization success rate) and the quadtree densification with water-medium modeling (Table V: PSNR gains on hard sequences) are ablated and show clear benefits. The coral reef dataset with synchronized multi-modal data is a useful contribution to the community. The system is implemented in C++/CUDA and runs at real-time speeds. However, the significance of the reliability-aware fusion framework itself is not fully isolated from the benefit of simply having additional sensors, as discussed below.

major comments (1)
  1. The central claim—that the reliability-aware fusion framework (dynamic reweighting, self-discrimination, sliding-window freezing) achieves state-of-the-art performance—is not adequately isolated from the benefit of simply adding DVL and pressure sensors. In Table II (simulation), APVI-SLAM uses M+I+D+P (4 modalities) while every baseline uses only M or M+I (1–2 modalities). The large gap on Hard sequences (5.9m vs 185–633m ATE) is likely dominated by DVL dead-reckoning maintaining pose during visual failure, not by the reweighting logic. Table III's ablation partially addresses this by showing incremental gains from adding sensors (I+D: 0.38m, I+D+P: 0.31m, M+I+D+P: 0.22m on Tank), but it does not include a naive fusion baseline using the same 4 sensors (e.g., a standard ESKF or fixed-weight factor graph with M+I+D+P). Without this comparison, the contribution of the reliability-aware re
minor comments (9)
  1. Section IV.A.b states the coral reef evaluation is 'restricted to qualitative comparisons, as quantitative metrics may not reliably reflect reconstruction quality,' yet Section IV.D reports PSNR values (24.14 vs 29.05). Please reconcile this inconsistency—either clarify what quantitative metrics were computed and how, or remove the PSNR numbers if they were not rigorously evaluated.
  2. Algorithm 1 uses hardcoded thresholds (feature count > 50, stable ratio > 0.2, positional offset < 30cm, angular offset < 10°). While these are common in practice, a brief discussion of how they were selected and whether they are expected to generalize across different camera configurations, sensor setups, and environments would strengthen the paper. A sensitivity analysis on at least one or two thresholds would be valuable.
  3. Section III-B.0.c: The Hessian eigenvalue thresholds (λ_thre_1, λ_thre_2) are derived from simulation data using semicircle distribution theory (Ref [31]). The generalization of these thresholds to real-world environments is not discussed. Please clarify whether these thresholds were validated on real data or adjusted for the coral reef dataset.
  4. Table I: The sensor column for 'Ours' lists M+I+D+P, but the paper's VIO subsystem uses a monocular camera. Please clarify whether stereo configurations were also tested or if the system is strictly monocular.
  5. Figure 2 contains Chinese text labels (e.g., 'DVL 残差构建', '压⼒计残差构建', '⾃判别模块', '滑窗冻结'). These should be translated to English for consistency.
  6. Section III-C.0.b: The quadtree levels are described as '16×16, 8×8, 6×6' but 6×6 is not a standard quadtree subdivision of 16×16. Please clarify whether this is a typo (perhaps 4×4?) or explain the non-standard division.
  7. Equation (8) references 'd_0, ..., d_63' (64 samples) but the text states '128 depth hypothesis.' Please reconcile this discrepancy.
  8. Table IV: The 'Reference' row for Re-init. Time is unclear—what is the reference method? Please clarify.
  9. The paper would benefit from a brief discussion of failure cases or limitations beyond the one-line mention in the conclusion (Section V).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. The referee raises a valid and important point: the current evaluation does not isolate the contribution of the reliability-aware fusion framework from the benefit of simply having additional sensors. We agree that a naive fusion baseline using the same M+I+D+P sensor suite is needed and will add it in the revision.

read point-by-point responses

  1. Referee: The central claim—that the reliability-aware fusion framework (dynamic reweighting, self-discrimination, sliding-window freezing) achieves state-of-the-art performance—is not adequately isolated from the benefit of simply adding DVL and pressure sensors. In Table II (simulation), APVI-SLAM uses M+I+D+P (4 modalities) while every baseline uses only M or M+I (1–2 modalities). The large gap on Hard sequences (5.9m vs 185–633m ATE) is likely dominated by DVL dead-reckoning maintaining pose during visual failure, not by the reweighting logic. Table III's ablation partially addresses this by showing incremental gains from adding sensors (I+D: 0.38m, I+D+P: 0.31m, M+I+D+P: 0.22m on Tank), but it does not include a naive fusion baseline using the same 4 sensors (e.g., a standard ESKF or fixed-weight factor graph with M+I+D+P). Without this comparison, the contribution of the reliability-aware re

    Authors: The referee is correct that the current evaluation does not fully isolate the contribution of the reliability-aware fusion logic from the benefit of having additional sensors. We acknowledge this gap and agree it must be addressed. We will add a naive fusion baseline that uses the same M+I+D+P sensor suite but with a standard fixed-weight approach (e.g., a tightly-coupled ESKF or fixed-weight factor graph that does not perform dynamic reweighting, self-discrimination, or sliding-window freezing). This will be included as a new row in both Table II (simulation) and Table I (Tank dataset). We expect this baseline to show that while having DVL and pressure sensors helps, the reliability-aware mechanisms—particularly the self-discrimination that prevents corrupted VIO estimates from contaminating the fused output, and the sliding-window freezing that enables rapid VIO recovery—provide additional gains beyond naive fusion, especially on Hard sequences where visual degradation is intermittent rather than total. We note that Table IV already provides partial evidence: without the sliding-window freezing strategy, the re-initialization success rate drops from 100% to 21%, and Table III shows that the full M+I+D+P configuration (0.22m) outperforms I+D+P (0.31m), suggesting the fusion logic extracts value from visual measurements beyond what DIPO alone provides. However, we agree that a direct naive-fusion comparison is the cleanest way to make this case, and we will add it. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity found; the derivation chain is self-contained with one minor non-load-bearing self-citation.

full rationale

The paper's derivation chain does not exhibit circularity. The core fusion formula (Eq. 6) is a standard ESKF update — the fused state is computed from VIO and DIPO estimates via Kalman gain, which is not equivalent to either input by construction. The water-medium model (Eq. 7) is externally cited from Akkaynak and Treibitz [32], not derived by the authors. The quadtree densification (Eqs. 8–9) uses standard photometric cost and belief propagation, building on externally cited work [33]. The Hessian eigenvalue thresholds in Sec. III-B.0.c are calibrated on a simulation dataset and then applied to real data — this is a calibration step, not a circular reduction (the threshold values are not the same quantity as the predictions they enable). Algorithm 1's hardcoded thresholds are design parameters, not self-definitional. The only self-citation is Photo-SLAM [11] (co-authored by Huang and Yeung), but it is used as a baseline for comparison, not as a load-bearing premise for any derivation in this paper. The experimental evaluation is against external baselines on public and new datasets. The skeptic's concerns about sensor-suite confounding and threshold generalization are correctness and experimental-design risks, not circularity.

Assumptions & free parameters 11 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities, particles, or forces. It combines existing sensor modalities and algorithmic components. The free parameters are primarily engineering thresholds and coefficients that govern the switching and fusion logic. Many are not numerically specified, which limits reproducibility. The axioms are domain assumptions about the applicability of statistical approximations and the generalizability of hardcoded thresholds.

free parameters (11)
  • alpha = not specified
    Linear coefficient scaling feature tracking consistency in the VIO information matrix (Sec. III-B.0.a). Value not reported.
  • beta = not specified
    Linear coefficient in the VIO information matrix (Sec. III-B.0.a). Value not reported.
  • n_min, n_max = not specified
    Minimum and maximum tracking thresholds for feature point consistency weighting (Sec. III-B.0.a). Values not reported.
  • Feature count threshold (50) = 50
    Hardcoded threshold in Alg. 1 Step 1 for coarse VIO stability judgment.
  • Stable feature ratio (0.2) = 0.2
    Hardcoded ratio in Alg. 1 Step 1 requiring >20% stable features.
  • Positional offset threshold (30cm) = 30cm
    Hardcoded threshold in Alg. 1 Step 2 for DIPO-VIO divergence.
  • Angular offset threshold (10 degrees) = 10
    Hardcoded threshold in Alg. 1 Step 2 for DIPO-VIO angular divergence.
  • lambda thresholds (lambda_thre_1, lambda_thre_2) = determined from simulation
    Hessian eigenvalue thresholds for VIO failure detection (Sec. III-B.0.c). Derived from simulated Hessian distributions but specific values not reported.
  • Water-medium learning rate multiplier (2x) = 2
    Learning rate doubled during VIO failure for water-medium parameters (Sec. III-C.0.a). Chosen ad hoc.
  • Quadtree levels (16x16, 8x8, 6x6) = [16,8,6]
    Quadtree division levels for densification (Sec. III-C.0.b). Note: 6x6 is not a standard quadtree subdivision.
  • Depth hypothesis range (0.10-25m, 128 samples) = 0.10-25, 128
    Depth sampling range and count for photometric cost evaluation (Sec. III-C.0.b).
assumptions (4)
  • domain assumption The semicircle law approximation for Hessian eigenvalue distributions holds for the VIO factor graph structure
    Sec. III-B.0.c invokes Ref [31] on random symmetric matrices to justify eigenvalue thresholds. The VIO Hessian is not a random matrix; its structure depends on the factor graph topology and measurement noise. The applicability of the semicircle law is assumed without proof.
  • ad hoc to paper Hardcoded thresholds in Alg. 1 generalize across environments and sensor configurations
    The 50-feature, 30cm, 10-degree thresholds are assumed to be valid across all tested environments. No sensitivity analysis or adaptation mechanism is provided.
  • domain assumption DVL and pressure sensor measurements are sufficiently accurate and free of systematic bias
    The DIPO subsystem (Sec. III-B.0.b) treats DVL velocity and pressure depth as direct measurements. DVL accuracy depends on bottom-track conditions and water column homogeneity, which are not discussed.
  • domain assumption The underwater image formation model (Eq. 7) with two attenuation parameters adequately captures water-medium effects
    The model from Ref [32] is used with per-pixel depth from Gaussians. The model assumes uniform water properties along the light path, which may not hold in stratified water columns.

how reviews work

0 comments
Cite this review

Pith. "Pith review of APVI-SLAM: Real-Time Acoustic-Pressure-Visual-Inertial Localization and Photorealistic Mapping System in Complex Underwater Environment." pith.science (2026). https://pith.science/paper/XK6DRGWC

@misc{pith2026260706222,
  author       = {Pith},
  title        = {Pith review of: APVI-SLAM: Real-Time Acoustic-Pressure-Visual-Inertial Localization and Photorealistic Mapping System in Complex Underwater Environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XK6DRGWC}},
  note         = {Machine review of arXiv:2607.06222}
}
read the original abstract

Extreme subsea environments often cause severe feature de-gradation and estimator divergence in underwater visual-inertial SLAM. Although sensors like Doppler Velocity Logs (DVL) and pressure gauges provide auxiliary constraints, robust multi-sensor fusion during intermittent visual failure remains challenging. To address this, we present APVI-SLAM, a real-time multi-sensor fusion SLAM system that achieves both accurate underwater localization and photorealistic mapping. Our approach introduces a reliability-aware localization framework that dynamically reweights sensor estimators and employs a sliding-window freezing strategy to recover from tracking failures, substantially enhancing system robustness. Furthermore, for high-fidelity scenes reconstruction, we propose an efficient quadtree-guided mapping module that facilitates incremental water-medium modeling and 3D Gaussian optimization. Recognizing the lack of benchmark for underwater mapping evaluation, we also contribute a coral reef surveying dataset with synchronized multi-modality data. Extensive experiments on public and our proposed benchmarks demonstrate that APVI-SLAM achieves state-of-the-art localization and reconstruction quality at real-time speeds.

Figures

Figures reproduced from arXiv: 2607.06222 by the authors.

Figure 1
Figure 1. The proposed APVI-SLAM enables real-time multi-sensor fusion [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed APVI-SLAM framework. Multi-sensor data, including RGB images, IMU, DVL, and pressure measurements, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Sliding-window freezing strategy for efficient re-initialization. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative results on the Tank dataset. Our method achieves [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Left: Qualitative results on the easy sequences of simulation dataset. Our method demonstrates superior texture clarity and finer geometric details [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on the coral reef surveying dataset. Our method [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages

  1. [1]

    A review of underwater slam technologies,

    H. Li, G. Wang, X. Li, and Y . Lian, “A review of underwater slam technologies,” in2023 5th International conference on robotics, intelligent control and artificial intelligence (RICAI). IEEE, 2023, pp. 215–225

    work page 2023

  2. [2]

    An overview of key slam technologies for underwater scenes,

    X. Wang, X. Fan, P. Shi, J. Ni, and Z. Zhou, “An overview of key slam technologies for underwater scenes,”Remote Sensing, vol. 15, no. 10, p. 2496, 2023

    work page 2023

  3. [3]

    Underwater slam meets deep learning: challenges, multi-sensor integration, and future directions,

    M. Heshmat, L. Saad Saoud, M. Abujabal, A. Sultan, M. Elmezain, L. Seneviratne, and I. Hussain, “Underwater slam meets deep learning: challenges, multi-sensor integration, and future directions,”Sensors, vol. 25, no. 11, p. 3258, 2025

    work page 2025

  4. [4]

    Underwater visual acoustic slam with extrinsic calibration,

    S. Xu, T. Luczynski, J. S. Willners, Z. Hong, K. Zhang, Y . R. Petillot, and S. Wang, “Underwater visual acoustic slam with extrinsic calibration,” in2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2021, pp. 7647–7652

    work page 2021

  5. [5]

    Svin2: A multi-sensor fusion-based underwater slam system,

    S. Rahman, A. Quattrini Li, and I. Rekleitis, “Svin2: A multi-sensor fusion-based underwater slam system,”The International Journal of Robotics Research, vol. 41, no. 11-12, pp. 1022–1042, 2022

    work page 2022

  6. [6]

    Turtlmap: Real-time localization and dense mapping of low-texture underwater environments with a low-cost unmanned underwater vehi- cle,

    J. Song, O. Bagoren, R. Andigani, A. Sethuraman, and K. A. Skinner, “Turtlmap: Real-time localization and dense mapping of low-texture underwater environments with a low-cost unmanned underwater vehi- cle,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 1191–1198

    work page 2024

  7. [7]

    Aqua-slam: Tightly-coupled un- derwater acoustic-visual-inertial slam with sensor calibration,

    S. Xu, K. Zhang, and S. Wang, “Aqua-slam: Tightly-coupled un- derwater acoustic-visual-inertial slam with sensor calibration,”IEEE Transactions on Robotics, 2025

    work page 2025

  8. [8]

    Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,

    T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monoc- ular visual-inertial state estimator,”IEEE transactions on robotics, vol. 34, no. 4, pp. 1004–1020, 2018

    work page 2018

Show all 39 references
  1. [9]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, G. Drettakis,et al., “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  2. [10]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 18 039–18 048

    2024

  3. [11]

    Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,

    H. Huang, L. Li, H. Cheng, and S.-K. Yeung, “Photo-slam: Real-time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 21 584– 21 593

    2024

  4. [12]

    Vings-mono: Visual-inertial gaussian splatting monocular slam in large scenes,

    K. Wu, Z. Zhang, M. Tie, Z. Ai, Z. Gan, and W. Ding, “Vings-mono: Visual-inertial gaussian splatting monocular slam in large scenes,” IEEE Transactions on Robotics, 2025

    2025

  5. [13]

    Seasplat: Representing un- derwater scenes with 3d gaussian splatting and a physically grounded image formation model,

    D. Yang, J. J. Leonard, and Y . Girdhar, “Seasplat: Representing un- derwater scenes with 3d gaussian splatting and a physically grounded image formation model,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 7632–7638

    2025

  6. [14]

    Watersplatting: Fast underwater 3d scene reconstruction using gaussian splatting,

    H. Li, W. Song, T. Xu, A. Elsig, and J. Kulhanek, “Watersplatting: Fast underwater 3d scene reconstruction using gaussian splatting,” in 2025 International Conference on 3D Vision (3DV). IEEE, 2025, pp. 969–978

    2025

  7. [15]

    Seathru-nerf: Neural radiance fields in scattering media,

    D. Levy, A. Peleg, N. Pearl, D. Rosenbaum, D. Akkaynak, S. Korman, and T. Treibitz, “Seathru-nerf: Neural radiance fields in scattering media,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 56–65

    2023

  8. [16]

    Underwater visual slam with depth uncertainty and medium modeling,

    R. Liu, S. Fan, W. Wang, and Y . Yang, “Underwater visual slam with depth uncertainty and medium modeling,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 970–980

    2025

  9. [17]

    Visually augmented navigation in an unstructured environment using a delayed state history,

    R. Eustice, O. Pizarro, and H. Singh, “Visually augmented navigation in an unstructured environment using a delayed state history,” in IEEE International Conference on Robotics and Automation, 2004. Proceedings. ICRA’04. 2004, vol. 1. IEEE, 2004, pp. 25–32

    2004

  10. [18]

    Improved mhe-based tightly coupled sins/usbl method for underwater navigation,

    D. Wang, S. Sun, H. Huang, and J. Wang, “Improved mhe-based tightly coupled sins/usbl method for underwater navigation,”IEEE Sensors Journal, 2025

    2025

  11. [19]

    A hybrid-kernel- based adaptive robust kalman filter for ins/dvl integrated underwater navigation,

    L. Kang, K. He, J. Zhao, X. Wang, and P. Tan, “A hybrid-kernel- based adaptive robust kalman filter for ins/dvl integrated underwater navigation,”Ocean Engineering, vol. 350, p. 124269, 2026

    2026

  12. [20]

    Advancing underwater robotic localization accuracy through multi-sensor data fusion and evolutionary optimization techniques,

    H. M. Ali, H. J. P. Williams, L. Arulmozhiselvan, F. T. Ayasrah, A. Smerat, S. Venkatesh, H. R. Suresh, and S. Sengan, “Advancing underwater robotic localization accuracy through multi-sensor data fusion and evolutionary optimization techniques,” 2025

    2025

  13. [21]

    Robust ins/gnss/dvl integrated navigation for mass based on gradient-adaptive factor graph optimization,

    M. Guo, B. Wang, L. Wei, M. Zhang, C. Zhang, and H. Lu, “Robust ins/gnss/dvl integrated navigation for mass based on gradient-adaptive factor graph optimization,”Electronics, vol. 15, no. 3, p. 634, 2026

    2026

  14. [22]

    Sonid: underwater sonar slam integrated with neural inertial-dvl odometry,

    X. Chen, Y . Feng, B. Nie, W. Lu, and X. Wang, “Sonid: underwater sonar slam integrated with neural inertial-dvl odometry,” inFifth International Conference on Artificial Intelligence, Automation, and Algorithms (AI2A 2025), vol. 14073. SPIE, 2026, pp. 28–36

    2025

  15. [23]

    Acoustic-vins: Tightly coupled acoustic- visual-inertial navigation system for autonomous underwater vehicles,

    J. Song, W. Li, and X. Zhu, “Acoustic-vins: Tightly coupled acoustic- visual-inertial navigation system for autonomous underwater vehicles,” IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1620–1627, 2023

    2023

  16. [24]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,

    Z. Teed and J. Deng, “Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras,”Advances in neural information processing systems, vol. 34, pp. 16 558–16 569, 2021

    2021

  17. [25]

    Factor graphs for robot perception,

    F. Dellaert and M. Kaess, “Factor graphs for robot perception,” Foundations and Trends® in Robotics, vol. 6, no. 1-2, pp. 1–139, 2017

    2017

  18. [26]

    The computation of optical flow,

    S. S. Beauchemin and J. L. Barron, “The computation of optical flow,” ACM computing surveys (CSUR), vol. 27, no. 3, pp. 433–466, 1995

    1995

  19. [27]

    Orb: An efficient alternative to sift or surf,

    E. Rublee, V . Rabaud, K. Konolige, and G. Bradski, “Orb: An efficient alternative to sift or surf,” in2011 International conference on computer vision. Ieee, 2011, pp. 2564–2571

    2011

  20. [28]

    Quaternion kinematics for the error-state kalman filter,

    J. Sola, “Quaternion kinematics for the error-state kalman filter,”arXiv preprint arXiv:1711.02508, 2017

    2017 arXiv

  21. [29]

    Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter,

    W. Xu and F. Zhang, “Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 3317–3324, 2021

    2021

  22. [30]

    An algorithm for least-squares estimation of non- linear parameters,

    D. W. Marquardt, “An algorithm for least-squares estimation of non- linear parameters,”Journal of the society for Industrial and Applied Mathematics, vol. 11, no. 2, pp. 431–441, 1963

    1963

  23. [31]

    Lectures on the local semicircle law for wigner matrices,

    F. Benaych-Georges and A. Knowles, “Lectures on the local semicircle law for wigner matrices,”arXiv preprint arXiv:1601.04055, 2016

    2016 arXiv

  24. [32]

    A revised underwater image formation model,

    D. Akkaynak and T. Treibitz, “A revised underwater image formation model,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6723–6732

    2018

  25. [33]

    Quadtree-accelerated real-time monocular dense mapping,

    K. Wang, W. Ding, and S. Shen, “Quadtree-accelerated real-time monocular dense mapping,” in2018 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS). IEEE, 2018, pp. 1–9

    2018

  26. [34]

    Semi-dense visual odometry for a monocular camera,

    J. Engel, J. Sturm, and D. Cremers, “Semi-dense visual odometry for a monocular camera,” inProceedings of the IEEE international conference on computer vision, 2013, pp. 1449–1456

    2013

  27. [35]

    Tank dataset: An underwater multi-sensor dataset for slam evaluation,

    S. Xu, J. Scharff Willners, J. Roe, S. Katagiri, T. Luczynski, Y . Petillot, and S. Wang, “Tank dataset: An underwater multi-sensor dataset for slam evaluation,”The International Journal of Robotics Research, p. 02783649251364904, 2025

    2025

  28. [36]

    Uuv simulator: A gazebo-based package for un- derwater intervention and multi-robot simulation,

    M. M. M. Manh ˜aes, S. A. Scherer, M. V oss, L. R. Douat, and T. Rauschenbach, “Uuv simulator: A gazebo-based package for un- derwater intervention and multi-robot simulation,” inOceans 2016 Mts/Ieee Monterey. Ieee, 2016, pp. 1–8

    2016

  29. [37]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  30. [38]

    Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,

    C. Campos, R. Elvira, J. J. G. Rodr ´ıguez, J. M. Montiel, and J. D. Tard ´os, “Orb-slam3: An accurate open-source library for visual, visual–inertial, and multimap slam,”IEEE transactions on robotics, vol. 37, no. 6, pp. 1874–1890, 2021

    2021

  31. [39]

    Go-slam: Global optimization for consistent 3d instant reconstruction,

    Y . Zhang, F. Tosi, S. Mattoccia, and M. Poggi, “Go-slam: Global optimization for consistent 3d instant reconstruction,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3727–3737

    2023

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.