Pith. sign in

REVIEW 4 major objections 5 minor 24 references

Enhancing Situational Awareness in Underwater Robotics with Multi-modal Spatial Perception

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A work-class ROV can map poor-visibility underwater scenes in real time by fusing cameras with inertial and velocity sensors.

desk verdict A useful new underwater field dataset wrapped in an overclaiming abstract; the SLAM extension is modest and the evaluation is qualitative, so treat it as a teaser until the numbers and data appear. read the letter →

arxiv 2506.06476 v1 pith:6MS3PLBE submitted 2025-06-06 cs.RO

Pushyami Kaveti , Ambjorn Grimsrud Waldum , Hanumant Singh , Martin Ludvigsen This is my paper
classification cs.RO
keywords underwaterSLAMmulti-camerastateestimationIMU-DVLsensorfusionROVfielddatasetslearning-basedvisualodometrysemantic3Dreconstructionperception
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

Underwater visibility is so poor that ordinary monocular or stereo SLAM often stops tracking, so the paper asks a practical question: can fusing several synchronized cameras with an IMU and a Doppler velocity log keep an ROV aware of where it is and what it sees? It reports field campaigns of a work-class ROV in a Norwegian fjord, covering a shipwreck, a subsea pipeline module, a shallow turbid seafloor loop, and a pipeline, and applies three perception methods to those recordings. The central result is a demonstration that a multi-camera SLAM back-end extended with IMU and DVL factors runs in real time and produces loop-closed trajectories and recognizable sparse 3D structure in conditions that defeat single-camera pipelines. Alongside it, the paper shows that a learning-based dense SLAM pipeline can reconstruct the shipwreck in detail but drifts because it lacks loop closure, and that a semantic segmentation model can tag reconstructed surfaces as pipeline, pipeline support, or background. If the demonstration holds, a multi-camera rig with inertial and velocity fusion is a viable route to real-time situational awareness in visually degraded underwater operations.

What carries the argument

The carrying mechanism is the rigid three-camera rig plus the augmented SLAM back-end. The rig's cameras are leader-follower frame-synchronized; their relative poses are calibrated in air and their individual focal properties in a saltwater pool. The back-end is a generalized multi-camera SLAM framework (MCSLAM) that models the camera array as a collection of rays, and the paper augments it with custom IMU and DVL factors in a factor graph, so vehicle motion measurements and depth and velocity readings constrain the pose optimization and enable loop closures. Against this baseline, a learning-based dense SLAM pipeline with a dense bundle adjustment layer and learned optical flow supplies dense depth and a detailed reconstruction, while a deep semantic segmentation network supplies pixel-level labels that are projected onto the reconstruction using the estimated camera poses and depths.

What would settle it

Run the same multi-camera rig over an object whose 3D geometry has been accurately surveyed, in turbid low-light water at working depth, and compute both the inter-camera reprojection error and the distance between the reconstructed model and the surveyed geometry. If those errors grow beyond the calibration uncertainty as pressure, temperature, or salinity increase, the pool-calibration assumption fails and the claimed transfer to field conditions is disproven for that environment.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-modal spatial perception—three synchronized cameras, an IMU, and a DVL fused in a factor-graph SLAM back-end, an optimization over poses, landmarks, and sensor measurements—enables real-time state estimation and high-quality 3D reconstruction in visually challenging underwater conditions. On the shipwreck and subsea-module surveys, the extended multi-camera SLAM produces continuous trajectories, visible loop closures, and point clouds in which the hull and the module's grid structure are recognizable. On the shipwreck, a learning-based dense SLAM pipeline yields a dense reconstruction with clear circular survey trajectories, but only after its global refinement stage, and it loses track on the sequence where the ROV leaves the scene and returns. The paper's comparison positions the geometric multi-sensor approach as the one that closes loops and fuses sensors, the learning-based approach as the one that tolerates visual degradation but accumulates drift, and semantic segmentation as a layer that gives the reconstruction object-level meaning.

Load-bearing premise

The load-bearing assumption is that the camera calibration measured in air and in a saltwater pool remains valid during field dives; if pressure, temperature, or refraction change the camera parameters, the multi-camera fusion and the maps built from it are no longer valid.

Editorial extensions

If this is right

  • An ROV with calibrated synchronized cameras plus IMU and DVL data can keep tracking through visually degraded stretches and close loops on return, producing a consistent sparse map where a single camera would lose track.
  • Dense learning-based reconstruction can deliver a detailed shipwreck model, but because it lacks loop closure and multi-sensor fusion it drifts on long sequences and after leaving and revisiting a scene; using it online requires adding relocalization or a geometric back-end.
  • Semantic labels can be overlaid on the 3D reconstruction to distinguish infrastructure such as pipelines and pipeline supports from the seabed, which is directly useful for inspection and intervention.
  • The field datasets themselves are a contribution: they give the community test sequences with loops, featureless intervals, changing visibility, and turbid shallow water for benchmarking underwater perception systems.
  • The paper's comparison points toward a hybrid design—learned optical flow for robustness combined with a geometric back-end that closes loops and fuses sensors—as the natural next system to build.

Reading between the lines

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

  • An implication the paper leaves implicit is that the practical deployment limit is calibration drift: operational teams that change zoom, pan, and tilt during a dive invalidate the fixed extrinsics, so online auto-recalibration is the step between this feasibility demonstration and routine field use.
  • The paper validates its reconstructions qualitatively; a quantitative specification would come from re-running the pipeline over a surveyed site and reporting loop-closure error and map-to-truth distance, which would make the feasibility claim falsifiable.
  • The semantic relocalization sketch in the discussion could be tested directly: drive the ROV through a featureless water column and back, then measure whether recognizing a previously seen object closes the loop better than raw visual features.
  • The paper's own limitation about thruster-stirred sediment implies that learned optical flow and segmentation should be stress-tested with synthetic data that adds realistic scattering and turbidity, not just with the relatively clear field sequences.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper describes a multi-modal spatial perception system for a work-class ROV, combining a rigid multi-camera rig, IMU, DVL, and other sensors. It introduces field datasets collected in the Trondheim Fjord (shipwreck, pipeline, seafloor, and a synthetic Gazebo dataset), and reports qualitative results from three perception pipelines: a multi-camera multi-sensor SLAM system built by extending the authors' prior MCSLAM framework with IMU and DVL factors, a learning-based dense visual SLAM (DROID-SLAM), and a semantic segmentation module (DeepLabv3) whose labels are projected onto a DROID-SLAM point cloud. The central claim is that this demonstrates the feasibility of real-time reliable state estimation and high-quality 3D reconstruction in challenging underwater conditions. The paper also discusses limitations of the learning-based methods and calibration challenges.

Significance. If the central feasibility claim were quantitatively supported, the paper would be a useful contribution to underwater robotics, particularly because it applies recent multi-camera SLAM methods to real field data and identifies practical issues such as calibration drift and loop-closure limitations. The datasets, if released, could enable benchmarking of multi-modal SLAM in visually degraded underwater environments. The paper is honest in acknowledging several limitations: DROID-SLAM lacks loop closures, the semantic segmentation model is trained on similar pipelines, and the evaluation is qualitative. However, the absence of any quantitative accuracy, precision, or runtime measurement, and the lack of comparison against unmodified baselines, mean that the central claim of 'real-time reliable state estimation' is currently unsupported. The qualitative reconstructions are encouraging, but they are not sufficient to establish reliability in the sense claimed.

major comments (4)
  1. [IV.A (see also Abstract)] The central claim of 'real-time reliable state estimation' in the abstract is not supported by the evidence presented. Section IV.A explicitly states that because ground truth is difficult to obtain, the authors 'primarily rely on the continuity of tracking and the structure of the reconstructed point clouds as qualitative performance indicators.' No trajectory error metrics (ATE/RPE), no scale check against DVL, no loop-closure residual, and no measured per-frame processing time are reported. 'Real-time' is never quantified; the only runtime-related datum is that the cameras run at 30 Hz, which does not establish that the processing pipeline meets any real-time constraint on the actual ROV computer. Please add quantitative evaluation, even if it is limited to a few sequences: for example, compare the SLAM trajectory against DVL-integrated odometry, report ATE/RPE before and after loop closure, and measure per-frame runtime on the onboard computer.
  2. [II.B] The calibration procedure determines extrinsic parameters in air and intrinsic parameters in a saltwater pool, and these parameters are then used in field deployments where pressure, temperature, and refraction conditions differ. The validity of these calibrations under field conditions is load-bearing for the multi-camera fusion and the resulting maps, but the paper provides no evidence of calibration stability in the field. Please provide a quantitative validation (e.g., reprojection error statistics on field images, or a comparison of multi-camera SLAM trajectories against DVL-derived trajectories) or an explicit sensitivity analysis of the results to plausible calibration errors.
  3. [IV.A] The paper describes extending MCSLAM with custom IMU and DVL factors, but it does not compare the extended system against the original MCSLAM (or a vision-only variant) on the same datasets. Without this ablation, the reader cannot determine whether the multi-modal fusion actually improves robustness or accuracy, or whether the observed tracking continuity is attributable to the base multi-camera SLAM. Please include a baseline comparison, even if qualitative, reporting tracking success/failure and, where possible, trajectory error for both configurations.
  4. [IV.B] The DROID-SLAM reconstruction in Figure 5 is described as 'high-quality', but no quantitative measure of reconstruction quality or trajectory accuracy is provided. The paper also notes that DROID-SLAM lacks loop closure, causing it to lose track in the PLM dataset where the ROV leaves and later revisits the module. A quantitative comparison of drift between the geometric MCSLAM-based system and DROID-SLAM on the same loop-closure sequences would help substantiate the claimed relative advantages and limitations of the two approaches.
minor comments (5)
  1. [I] There are typographical errors, e.g., 'monitering' in the first paragraph and 'starboarrd' and 'SubFigther' in Section II.A. These should be corrected.
  2. [II.A] The ROV platform description mentions a 'Sperre 100K-SubFigther' and manipulators, but the spelling is inconsistent and should be checked against the manufacturer's naming.
  3. [III.A] Table II lists the TBS Seafloor dataset as '19 April 2024' and calls it a 'deck test' in the text; please clarify whether this is a dockside or field deployment, as the distinction matters for interpreting the environmental conditions.
  4. [V.C] The sentence 'the segmentation model is trained of nearby similar pipelines' is grammatically unclear and should be rewritten, e.g., 'trained on images of nearby, similar pipelines'.
  5. [III] The paper introduces datasets as a contribution but does not state whether they will be publicly released. If the datasets are central to the paper's value, a data-availability statement would be helpful.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the paper applies previously published methods with qualitative field evaluation; self-citations are not load-bearing.

full rationale

The paper makes no new formal derivation and fits no free parameters to its own outputs. Its multi-camera SLAM system is MCSLAM, cited to the authors' prior IEEE RA-L paper (ref. [15]); that prior work is an external, published result rather than a premise defined in terms of the present claim, so the self-citation is informative but not circular. The claimed contribution is a dataset collection and qualitative field evaluation, with DROID-SLAM and a DeepLabV3-based semantic projection used as published off-the-shelf tools. The paper explicitly limits its evidence: Section IV.A states that due to the difficulty of obtaining ground truth, the authors 'primarily rely on the continuity of tracking and the structure of the reconstructed point clouds as qualitative performance indicators.' This is a weakness in support for the abstract's 'feasibility of real-time reliable state estimation,' but it is an evidence gap, not a circularity: no metric is computed from a quantity already assumed by the method. The Discussion also candidly acknowledges calibration drift (Section V.B), the lack of loop closure in DROID-SLAM (Section V.A), and that the semantic segmentation model 'is trained of nearby similar pipelines' (Section V.C), which reduces the external-generalization claim rather than smuggling in a conclusion. No equation is shown to equal its own input, no fitted parameter is renamed as a prediction, and the cited prior work is independently published and falsifiable. Consistent with the reader's assessment, a score of 1 reflects only the presence of minor self-citations that are not load-bearing in a circular sense.

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

The central claim rests on the assumption that existing SLAM and deep-learning methods transfer to the underwater domain, and that the calibration performed in benign conditions holds in the field. No new free parameters are fitted in this paper.

assumptions (4)
  • domain assumption MCSLAM's generalized camera model correctly represents the three-camera rig
    The paper builds on MCSLAM [15] without re-deriving the model; if the generalized camera model fails for this rig, the multi-camera SLAM results are invalid.
  • domain assumption DROID-SLAM's pretrained weights generalize to underwater imagery
    DROID-SLAM is used off-the-shelf (Section IV.B); if the learned optical flow fails underwater, the reconstructions would not be produced.
  • domain assumption The calibration from air and pool transfers to field conditions
    Section II.B describes air and pool calibration; field pressure, temperature, and refraction may change the calibration, affecting the multi-camera fusion.
  • domain assumption The semantic segmentation model, trained on similar pipelines, is applicable to the pipeline dataset
    Section V.C admits the model is trained on nearby similar pipelines, so its success on the pipeline data is expected and not a test of generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Situational Awareness in Underwater Robotics with Multi-modal Spatial Perception." pith.science (2026). https://pith.science/paper/6MS3PLBE

@misc{pith2026250606476,
  author       = {Pith},
  title        = {Pith review of: Enhancing Situational Awareness in Underwater Robotics with Multi-modal Spatial Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6MS3PLBE}},
  note         = {Machine review of arXiv:2506.06476}
}
read the original abstract

Autonomous Underwater Vehicles (AUVs) and Remotely Operated Vehicles (ROVs) demand robust spatial perception capabilities, including Simultaneous Localization and Mapping (SLAM), to support both remote and autonomous tasks. Vision-based systems have been integral to these advancements, capturing rich color and texture at low cost while enabling semantic scene understanding. However, underwater conditions -- such as light attenuation, backscatter, and low contrast -- often degrade image quality to the point where traditional vision-based SLAM pipelines fail. Moreover, these pipelines typically rely on monocular or stereo inputs, limiting their scalability to the multi-camera configurations common on many vehicles. To address these issues, we propose to leverage multi-modal sensing that fuses data from multiple sensors-including cameras, inertial measurement units (IMUs), and acoustic devices-to enhance situational awareness and enable robust, real-time SLAM. We explore both geometric and learning-based techniques along with semantic analysis, and conduct experiments on the data collected from a work-class ROV during several field deployments in the Trondheim Fjord. Through our experimental results, we demonstrate the feasibility of real-time reliable state estimation and high-quality 3D reconstructions in visually challenging underwater conditions. We also discuss system constraints and identify open research questions, such as sensor calibration, limitations with learning-based methods, that merit further exploration to advance large-scale underwater operations.

Figures

Figures reproduced from arXiv: 2506.06476 by the authors.

Figure 1
Figure 1. Minerva II. The red-boxes from top to bottom shows [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Prototype of multi-camera rig consisting of three [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. ROV Control Room. Courtesy of Ole Martin Wold. [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Snapshots from both the synthetic (a) and real-world field datasets (b), alongside the multi-camera, multi-sensor [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]
Figure 5
Figure 5. Figure 5: The final dense 3D reconstruction of the Hercules Shipwreck captured from various viewing angles. The red triangles [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Factor graph of the multi-camera back-end with IMU [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Point cloud generated by DROID-SLAM with and [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 19 canonical work pages

  1. [1]

    Scientific challenges and present capabilities in underwater robotic vehicle design and navigation for oceanographic exploration under- ice,

    L. D. L. Barker, M. V . Jakuba, A. D. Bowen, C. R. German, T. Maksym, L. Mayer, A. Boetius, P. Dutrieux, and L. L. Whitcomb, “Scientific challenges and present capabilities in underwater robotic vehicle design and navigation for oceanographic exploration under- ice,” Remote Sensing , vol. 12, no. 16, 2020. [Online]. Available: https://www.mdpi.com/2072-42...

    work page 2020

  2. [2]

    Role of deep-sea equipment in promoting the forefront of studies on life in extreme environments,

    J. Liang, J.-C. Feng, S. Zhang, Y . Cai, Z. Yang, T. Ni, and H.-Y . Yang, “Role of deep-sea equipment in promoting the forefront of studies on life in extreme environments,” iScience, vol. 24, no. 11, p. 103299, 2021. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S2589004221012682

    work page 2021

  3. [3]

    Subsea infrastructure inspection: A review study,

    C. Mai, S. Pedersen, L. Hansen, K. L. Jepsen, and Z. Yang, “Subsea infrastructure inspection: A review study,” in 2016 IEEE International Conference on Underwater System Technology: Theory and Applica- tions (USYS) , 2016, pp. 71–76

    work page 2016

  4. [4]

    A subsea space race,

    E. Maslin, “A subsea space race,” Engineering & technology , vol. 15, no. 3, pp. 72–75, 2020

    work page 2020

  5. [5]

    Underwater photogrammetric mapping of an intact standing steel wreck with rov,

    S. M. Nornes, M. Ludvigsen, Øyvind Ødegard, and A. J. SØrensen, “Underwater photogrammetric mapping of an intact standing steel wreck with rov,” IF AC-PapersOnLine, vol. 48, no. 2, pp. 206–211, 2015, 4th IFAC Workshop onNavigation, Guidance and Controlof Underwater VehiclesNGCUV 2015. [Online]. Available: https: //www.sciencedirect.com/science/article/pii...

    work page 2015

  6. [6]

    Towards adaptive benthic habitat mapping,

    J. Shields, O. Pizarro, and S. B. Williams, “Towards adaptive benthic habitat mapping,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) , 2020, pp. 9263–9270

    work page 2020

  7. [7]

    Towards Automated Sample Collection and Return in Extreme Underwater Environments

    G. Billings, M. Walter, O. Pizarro, M. Johnson-Roberson, and R. Camilli, “Towards automated sample collection and return in extreme underwater environments,” 2021. [Online]. Available: https://arxiv.org/abs/2112.15127

    work page Pith review arXiv 2021

  8. [8]

    Mobley, The Oceanic Optics Book , 01 2022

    C. Mobley, The Oceanic Optics Book , 01 2022

    work page 2022

Show all 24 references
  1. [9]

    Extending kalibr: Calibrating the extrinsics of multiple imus and of individual axes,

    J. Rehder, J. Nikolic, T. Schneider, T. Hinzmann, and R. Siegwart, “Extending kalibr: Calibrating the extrinsics of multiple imus and of individual axes,” in 2016 IEEE International Conference on Robotics and Automation (ICRA) , 2016, pp. 4304–4311

    2016

  2. [10]

    From virtual waters to real oceans: A simulation-driven approach to rov control system design,

    W. Ambjørn, Fossdal, M. Basso, Erlend, and L. Martin, “From virtual waters to real oceans: A simulation-driven approach to rov control system design,” Paper presented at the Underwater Technology 2025, Tapei, March 2025

    2025

  3. [11]

    A review of monocular visual odometry,

    M. He, C. Zhu, Q. Huang, B. Ren, and J. Liu, “A review of monocular visual odometry,” The Visual Computer, vol. 36, pp. 1053–1065, 2020

    2020

  4. [12]

    ORBSlam3: 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, “ORBSlam3: An Accurate Open-Source Library for Visual, Visual–Inertial, and Multimap SLAM,” IEEE Transactions on Robotics, 2021

    2021

  5. [13]

    Redesigning slam for arbitrary multi-camera systems,

    J. Kuo, M. Muglikar, Z. Zhang, and D. Scaramuzza, “Redesigning slam for arbitrary multi-camera systems,” in 2020 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2020

    2020

  6. [14]

    Multicol-slam-a modular real-time multi- camera slam system,

    S. Urban and S. Hinz, “Multicol-slam-a modular real-time multi- camera slam system,” arXiv preprint arXiv:1610.07336 , 2016

    2016 arXiv

  7. [15]

    Design and evaluation of a generic visual slam framework for multi camera systems,

    P. Kaveti, S. N. Vaidyanathan, A. T. Chelvan, and H. Singh, “Design and evaluation of a generic visual slam framework for multi camera systems,” IEEE Robotics and Automation Letters , vol. 8, no. 11, pp. 7368–7375, 2023

    2023

  8. [16]

    Flownet: Learning optical flow with convolutional networks,

    A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas, V . Golkov, P. Van Der Smagt, D. Cremers, and T. Brox, “Flownet: Learning optical flow with convolutional networks,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 2758–2766

    2015

  9. [17]

    Gmflow: Learning optical flow via global matching,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, and D. Tao, “Gmflow: Learning optical flow via global matching,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 8121–8130

    2022

  10. [18]

    Raft: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II

    2020

  11. [19]

    Springer, 2020, pp. 402–419. Presented at the 2025 IEEE ICRA Workshop on Field Robotics

    2020

  12. [20]

    DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras,

    ——, “DROID-SLAM: deep visual SLAM for monocular, stereo, and RGB-D cameras,” CoRR, vol. abs/2108.10869, 2021. [Online]. Available: https://arxiv.org/abs/2108.10869

    2021 arXiv

  13. [21]

    Multimodal semantic segmentation in autonomous driving: A review of current approaches and future perspectives,

    G. Rizzoli, F. Barbato, and P. Zanuttigh, “Multimodal semantic segmentation in autonomous driving: A review of current approaches and future perspectives,” Technologies, vol. 10, no. 4, 2022. [Online]. Available: https://www.mdpi.com/2227-7080/10/4/90

    2022

  14. [22]

    Deep semantic segmentation for automated driving: Taxonomy, roadmap and challenges,

    M. Siam, S. Elkerdawy, M. Jagersand, and S. Yogamani, “Deep semantic segmentation for automated driving: Taxonomy, roadmap and challenges,” 2017. [Online]. Available: https://arxiv.org/abs/1707. 02432

    2017

  15. [23]

    A survey on deep learning based approaches for scene understanding in autonomous driving,

    Z. Guo, Y . Huang, X. Hu, H. Wei, and B. Zhao, “A survey on deep learning based approaches for scene understanding in autonomous driving,” Electronics, vol. 10, no. 4, 2021. [Online]. Available: https://www.mdpi.com/2079-9292/10/4/471

    2021

  16. [24]

    Rethinking atrous convolution for semantic image segmentation,

    L. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,” CoRR, vol. abs/1706.05587, 2017. [Online]. Available: http://arxiv.org/abs/1706. 05587 Presented at the 2025 IEEE ICRA Workshop on Field Robotics

    2017 arXiv

Pith tools

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