Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Dense-depth map guided deep Lidar-Visual Odometry with Sparse Point Clouds and Images

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A LiDAR-visual odometry system that fills sparse point clouds into dense depth maps and uses those maps to scale optical flow and refine pose estimates achieves accuracy competitive with state-of-the-art visual and LiDAR odometry on the…

desk verdict A plausible dense-depth-guided fusion pipeline whose ablation is solid, but the 'superior on KITTI' claim relies on averaging over training sequences; on held-out 09-10, DVL-SLAM beats it. read the letter →

arxiv 2507.15496 v1 pith:3BKO43CN submitted 2025-07-21 cs.CV cs.LGcs.RO

JunYing Huang , Ao Xu , DongSun Yong , KeRen Li , YuanFeng Wang , Qi Qin This is my paper
classification cs.CVcs.LGcs.RO
keywords LiDAR-visualodometrydepthcompletionopticalflowposeestimationmulti-scalefeatureextractionattentionmechanismKITTIdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes D3LVO, a deep LiDAR-visual odometry system that turns sparse LiDAR scans and RGB images into dense depth maps and feeds those maps throughout the pose-estimation pipeline. It claims that this dense depth guidance — scaling optical flow magnitudes and conditioning hierarchical pose updates — resolves scale ambiguity and occlusion errors that hurt RGB-only or sparse-depth systems. On the KITTI odometry benchmark, the method reports a mean translation error of 0.97 percent on sequences 07–10 and 0.86 percent on sequences 00–10, the lowest among the compared learning-based methods, with strong generalization to unseen sequences 09 and 10. The intended significance is that dense depth from completion is a simple, effective bridge between camera and LiDAR modalities for odometry.

What carries the argument

The central object is the dense depth map $D_i$ produced by PENet depth completion, which enters the pipeline twice: as a fourth channel appended to RGB, and as the input to a depth-modulation function $G(D_i)$ that multiplies the raw flow prediction, $u_i = F(F_1^i, F_2^i, C_i) \cdot G(D_i)$, and scales residual flow in refinement, $u_i = u_{i+1}^{\uparrow} + G(D_i)\Delta u_i$. This depth-scaled flow, combined with a residual pose update $T_{i+1} = T_i + \Delta T_i$ where $\Delta T_i$ is regressed from flow, depth, and prior pose by an MLP, is what carries the accuracy gains. The machinery is a PWC-Net-style four-level pyramid with a $9 \times 9$ search-window cost volume, plus uncertainty-weighted fusion of level predictions.

What would settle it

Run the same architecture on a driving sequence with heavy rain, night lighting, or reflective surfaces where PENet depth completion is known to produce artifacts, and compare its trajectory error against the RGB+sparse-depth ablation; if the dense-depth version does not beat the sparse-depth version, the central claim that dense completion guidance is the source of the gains fails. Alternatively, train and test on a different benchmark such as KITTI-360 to check whether the reported advantage persists outside the training distribution.

Watch

Extended reading notes

Core claim

The paper shows that dense depth maps produced by a pretrained depth-completion network, concatenated with RGB as a four-channel input and used to modulate optical flow and pose refinement at every pyramid level, yield more accurate and robust pose estimation than RGB-only, sparse-depth, or single-modality baselines. The network combines multi-scale feature extraction with channel, spatial, and cross attention; a depth-aware flow predictor that scales flow by a learned depth-dependent weight map; and a hierarchical pose refinement module that regresses residual pose updates from flow, depth, and previous estimates. Evaluated on KITTI sequences 00–10, the method achieves mean translational RMSE of 0.86–0.97 percent depending on the comparison table, outperforming the compared visual, LiDAR, and multimodal methods on most sequences.

Load-bearing premise

The dense depth maps from the pretrained PENet depth-completion network are accurate enough on the test sequences to provide reliable guidance; the paper's own conclusion concedes the method relies on depth-completion performance, and if completion fails in low-texture, reflective, or out-of-distribution scenes, the flow scaling and pose refinement degrade.

Editorial extensions

If this is right

  • If correct, dense depth completion offers a drop-in way to improve deep visual odometry without changing the sensor suite.
  • The reported 88% lower translation error on KITTI 09–10 relative to RGB-only input suggests depth is not a minor refinement but the dominant cue.
  • The hierarchical uncertainty-weighted pose refinement implies that coarse-to-fine depth-guided estimates can be made robust across scales.
  • The method's generalization to unseen sequences 09–10 (trained on 00–08) supports the claim that depth-aware features transfer across driving environments.

Reading between the lines

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

  • A natural extension is to make depth completion trainable end-to-end with the odometry head; the paper treats completion as a fixed pretrained module, so joint fine-tuning could push accuracy further and may be where part of the reported gains originate.
  • The depth-modulation mechanism could generalize to scene-flow estimation or video frame interpolation, where depth-scaled flow faces the same scale-ambiguity problem.
  • The reliance on PENet means the method inherits PENet's generalization boundary; testing on out-of-distribution depth-completion failures is the fastest way to probe the method's real-world robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes D3LVO, a deep LiDAR-visual odometry framework that (i) completes sparse LiDAR depth with a pretrained PENet module, (ii) fuses RGB and dense depth into a four-channel input, (iii) extracts multi-scale attention-weighted features, (iv) estimates optical flow with depth-aware scaling at multiple pyramid levels, and (v) refines poses hierarchically with uncertainty weighting. The loss is a scale-aware multi-level pose loss. Experiments are on KITTI odometry sequences 00-10, with training on 00-08 and evaluation reported on all of 00-10; Tables I-III compare against visual, LiDAR, and multimodal baselines, and Table IV ablates RGB-only vs sparse-depth vs completed-depth input on held-out sequences 09-10. The authors claim accuracy and robustness similar or superior to state-of-the-art visual and LiDAR odometry methods.

Significance. The direction is sensible, and the paper has concrete strengths: it evaluates on a standard benchmark, reports held-out 09-10 results, includes an ablation isolating the effect of dense depth, and uses standard odometry metrics. The depth-aware flow modulation and hierarchical pose refinement are plausible architectural choices that align with the ablation evidence. However, the headline claim of 'similar or superior accuracy' is not supported as presented: the comparative tables include training sequences in the reported averages, several baselines are self-supervised while D3LVO is supervised with ground-truth poses, and no error bars or repeated runs are given. The reliance on a fixed pretrained depth-completion module is acknowledged in Section V but not stress-tested. With a corrected evaluation protocol and more careful claims, the work could be a useful engineering contribution; in its current form the central comparative claim is overstated.

major comments (3)
  1. [IV-D-2 / Table II] The claim that D3LVO 'achieves the lowest RMSE values in five out of the 11 sequences and outperforms all baselines in terms of average RMSE' is computed over sequences 00-10, which include the training sequences 00-08. On the held-out pair 09-10, DVL-SLAM reports trel = 0.66/0.70 (mean 0.68), PL-LOAM reports 1.06/0.83 (mean 0.945), and D3LVO reports 0.97/0.95 (mean 0.96). Thus DVL-SLAM is substantially more accurate and PL-LOAM is slightly more accurate on unseen data, which directly contradicts the abstract's 'similar or superior accuracy' claim. Please report held-out means separately and restrict comparative claims to the evidence, or retrain/evaluate on a proper split.
  2. [IV-D-1 / Table I] Table I compares D3LVO against mostly self-supervised or single-sensor baselines while D3LVO is trained with ground-truth poses on sequences 00-08, and the 'Mean 07-10' column includes sequences 07-08, which are also training sequences. This protocol does not demonstrate superior generalization, and no error bars or multiple-seed statistics are provided for any method. Please mark the supervision status of each baseline, add held-out-only statistics, and report variance or repeated-run results.
  3. [III-A / V] The paper states in Section V that 'our method relies on the performance of depth-completion' and uses PENet without fine-tuning, yet no experiment quantifies sensitivity to depth-completion errors. Since the depth maps multiply flow estimates in Eqs. (6)-(7) and condition the pose refinement, a systematic failure in low-texture, reflective, or out-of-distribution scenes could invalidate the robustness claim. Please add a robustness analysis (e.g., corrupting depth maps, substituting a different completion model, or testing on non-KITTI data) or explicitly temper the robustness claim.
minor comments (6)
  1. [IV-B] Training hyperparameters are incomplete: please report the optimizer, learning rate schedule, batch size, number of epochs, and the loss weights alpha_l used in Eq. (12).
  2. [III-F / Eq. (11)] Equation (11) is missing explicit norm delimiters and parentheses; please write the translation and rotation terms unambiguously and define all symbols, including the predicted/ground-truth quaternion normalization.
  3. [IV-D-2] The phrase '8.5% decline in the mean translation error' should be '8.5% reduction' and should state the comparison baseline and sequence range; as written it is ambiguous whether the direction is favorable.
  4. [IV-D-3] The text says 'an 73% lower'; it should read 'a 73% lower', and the comparison target (An et al. on sequences 07 and 10) should be stated in the same sentence.
  5. [III-D] The paper claims design 'for efficiency and robustness in real-time settings' but reports no runtime, parameter count, or inference speed; please add these numbers.
  6. [General] No code, trained checkpoints, or evaluation scripts are released, which limits reproducibility; consider providing them in the revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: pose supervision is external ground truth, depth completion is an external pretrained module, and held-out 09-10 results are reported; any evaluation-validity concern is not a derivation-circle.

full rationale

I examined the derivation chain. Pose regression is supervised by ground-truth poses through the scale-aware loss (Eqs. 11-12), depth maps come from a fixed external pretrained PENet module [32], and evaluation on held-out sequences 09-10 is reported (Tables III-IV). No formula equates a predicted quantity to a fitted parameter, no self-citation is load-bearing, and no uniqueness or ansatz is imported from the authors' own prior work. The paper's own Section V concedes reliance on depth-completion performance, which is a limitation rather than a circular reduction. The possible over-reliance on training sequences 00-08 in the 'best average' claim is an evaluation-validity concern, not a circularity of the derivation chain. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central result rests on a pretrained depth-completion module and standard KITTI supervision rather than on a self-contained derivation. The listed free parameters are loss and network hyperparameters whose values are not reported, so the exact numbers in Tables I-III are not reproducible from the paper alone.

free parameters (3)
  • learnable loss scales s_t, s_q = not reported
    Equation (11) introduces s_t and s_q to balance translation and rotation losses; they are learned from data and their final values are not given, so the reported errors are contingent on these fitted scalars.
  • level weights alpha_l = not reported
    Equation (12) weights the per-level losses alpha_l; these values are chosen or learned but not specified, affecting the trained model.
  • optical flow search range S = 4
    The cost volume search range is fixed to S = 4 (Section III-C), a hand-chosen hyperparameter that bounds the maximum pixel displacement the network can estimate.
assumptions (4)
  • domain assumption PENet produces accurate dense depth on KITTI odometry sequences
    The entire pipeline uses completed depth from a pretrained PENet (Section III-A) without fine-tuning; inaccurate depth would propagate errors through flow scaling and pose refinement.
  • domain assumption KITTI ground-truth poses from GPS/IMU are accurate enough for supervision
    Section IV-A uses KITTI GPS/IMU poses as labels for training and evaluation; systematic ground-truth error would bias the reported RMSE values.
  • domain assumption Optical flow between consecutive frames is dominated by ego-motion and static scenes
    The depth-aware flow module estimates pixel motion and pose from RGB-D; dynamic objects are mentioned but no explicit dynamic-object handling is described in Section III-D.
  • standard math L2 distance between normalized quaternions approximates rotation error
    Equation (11) uses squared L2 of normalized quaternion difference instead of geodesic rotation distance; this is a common approximation but not exact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dense-depth map guided deep Lidar-Visual Odometry with Sparse Point Clouds and Images." pith.science (2026). https://pith.science/paper/3BKO43CN

@misc{pith2026250715496,
  author       = {Pith},
  title        = {Pith review of: Dense-depth map guided deep Lidar-Visual Odometry with Sparse Point Clouds and Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BKO43CN}},
  note         = {Machine review of arXiv:2507.15496}
}
read the original abstract

Odometry is a critical task for autonomous systems for self-localization and navigation. We propose a novel LiDAR-Visual odometry framework that integrates LiDAR point clouds and images for accurate and robust pose estimation. Our method utilizes a dense-depth map estimated from point clouds and images through depth completion, and incorporates a multi-scale feature extraction network with attention mechanisms, enabling adaptive depth-aware representations. Furthermore, we leverage dense depth information to refine flow estimation and mitigate errors in occlusion-prone regions. Our hierarchical pose refinement module optimizes motion estimation progressively, ensuring robust predictions against dynamic environments and scale ambiguities. Comprehensive experiments on the KITTI odometry benchmark demonstrate that our approach achieves similar or superior accuracy and robustness compared to state-of-the-art visual and LiDAR odometry methods.

Figures

Figures reproduced from arXiv: 2507.15496 by the authors.

Figure 1
Figure 1. Overview of our proposed D3LVO framework. The network consists of depth completion, multi-scale feature extraction with attention mechanisms, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our attention-guided feature extraction/fusion module. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Overview of Depth-Aware Pose Estimation: The proposed Pose Warp [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: 2D trajectory comparison on KITTI sequences 09 and 10. The [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,

    T. Shan, B. Englot, D. Meyers, W. Wang, C. Ratti, and D. Rus, “Lio-sam: Tightly-coupled lidar inertial odometry via smoothing and mapping,” in 2020 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2020, pp. 5135–5142

    work page 2020

  2. [2]

    Visual-lidar odometry and mapping: Low-drift, robust, and fast,

    J. Zhang and S. Singh, “Visual-lidar odometry and mapping: Low-drift, robust, and fast,” in 2015 IEEE international conference on robotics and automation (ICRA). IEEE, 2015, pp. 2174–2181

    work page 2015

  3. [3]

    Self-supervised visual- lidar odometry with flip consistency,

    B. Li, M. Hu, S. Wang, L. Wang, and X. Gong, “Self-supervised visual- lidar odometry with flip consistency,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 3844– 3852

    work page 2021

  4. [4]

    Cnn-slam: Real-time dense monocular slam with learned depth prediction,

    K. Tateno, F. Tombari, I. Laina, and N. Navab, “Cnn-slam: Real-time dense monocular slam with learned depth prediction,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 6243–6252

    work page 2017

  5. [5]

    Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume,

    D. Sun, X. Yang, M.-Y . Liu, and J. Kautz, “Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 8934–8943

    work page 2018

  6. [6]

    Lidar odometry and mapping based on semantic information for outdoor environment,

    S. Du, Y . Li, X. Li, and M. Wu, “Lidar odometry and mapping based on semantic information for outdoor environment,”Remote Sensing, vol. 13, no. 15, p. 2864, 2021

    work page 2021

  7. [7]

    Loam: Lidar odometry and mapping in real- time

    J. Zhang, S. Singh et al., “Loam: Lidar odometry and mapping in real- time.” in Robotics: Science and systems , vol. 2, no. 9. Berkeley, CA, 2014, pp. 1–9

    work page 2014

  8. [8]

    Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain,

    T. Shan and B. Englot, “Lego-loam: Lightweight and ground-optimized lidar odometry and mapping on variable terrain,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2018, pp. 4758–4765

    work page 2018

Show all 50 references
  1. [9]

    Vision meets robotics: The kitti dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The kitti dataset,” The International Journal of Robotics Research , vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  2. [10]

    Depth completion from sparse lidar data with depth-normal constraints,

    Y . Xu, X. Zhu, J. Shi, G. Zhang, H. Bao, and H. Li, “Depth completion from sparse lidar data with depth-normal constraints,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 2811–2820

    2019

  3. [11]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems, 2017

    2017

  4. [12]

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

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in 2012 IEEE conference on computer vision and pattern recognition . IEEE, 2012, pp. 3354–3361

    2012

  5. [13]

    Orb-slam: a versatile and accurate monocular slam system,

    R. Mur-Artal, J. M. M. Montiel, and J. D. Tardos, “Orb-slam: a versatile and accurate monocular slam system,” IEEE transactions on robotics , vol. 31, no. 5, pp. 1147–1163, 2015

    2015

  6. [14]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,

    R. Mur-Artal and J. D. Tard ´os, “Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras,” IEEE transactions on robotics, vol. 33, no. 5, pp. 1255–1262, 2017

    2017

  7. [15]

    Direct sparse odometry,

    J. Engel, V . Koltun, and D. Cremers, “Direct sparse odometry,” IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 3, pp. 611–625, 2017

    2017

  8. [16]

    Dtam: Dense tracking and mapping in real-time,

    R. A. Newcombe, S. J. Lovegrove, and A. J. Davison, “Dtam: Dense tracking and mapping in real-time,” in 2011 international conference on computer vision. IEEE, 2011, pp. 2320–2327. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 8

    2011

  9. [17]

    Lsd-slam: Large-scale direct monocular slam,

    J. Engel, T. Sch ¨ops, and D. Cremers, “Lsd-slam: Large-scale direct monocular slam,” in European conference on computer vision. Springer, 2014, pp. 834–849

    2014

  10. [18]

    Deepvo: Towards end-to- end visual odometry with deep recurrent convolutional neural networks,

    S. Wang, R. Clark, H. Wen, and N. Trigoni, “Deepvo: Towards end-to- end visual odometry with deep recurrent convolutional neural networks,” in 2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017, pp. 2043–2050

    2017

  11. [19]

    Df-vo: What should be learnt for visual odometry?

    H. Zhan, C. S. Weerasekera, J.-W. Bian, R. Garg, and I. Reid, “Df-vo: What should be learnt for visual odometry?” arXiv preprint arXiv:2103.00933, 2021

    2021 arXiv

  12. [20]

    Raft: Recurrent all-pairs field transforms for op- tical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for op- tical flow,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 2020, pp. 402–419

    2020

  13. [21]

    Unsupervised learning of monocular depth estimation and visual odom- etry with deep feature reconstruction,

    H. Zhan, R. Garg, C. S. Weerasekera, K. Li, H. Agarwal, and I. Reid, “Unsupervised learning of monocular depth estimation and visual odom- etry with deep feature reconstruction,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 340– 349

    2018

  14. [22]

    D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry,

    N. Yang, L. v. Stumberg, R. Wang, and D. Cremers, “D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 1281–1292

    2020

  15. [23]

    Method for registration of 3-d shapes,

    P. J. Besl and N. D. McKay, “Method for registration of 3-d shapes,” in Sensor fusion IV: control paradigms and data structures , vol. 1611. Spie, 1992, pp. 586–606

    1992

  16. [24]

    Tightly coupled 3d lidar inertial odometry and mapping,

    H. Ye, Y . Chen, and M. Liu, “Tightly coupled 3d lidar inertial odometry and mapping,” in 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 3144–3150

    2019

  17. [25]

    Lo- net: Deep real-time lidar odometry,

    Q. Li, S. Chen, C. Wang, X. Li, C. Wen, M. Cheng, and J. Li, “Lo- net: Deep real-time lidar odometry,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2019, pp. 8473–8482

    2019

  18. [26]

    Deeplo: Geometry-aware deep lidar odometry,

    Y . Cho, G. Kim, and A. Kim, “Deeplo: Geometry-aware deep lidar odometry,” arXiv preprint arXiv:1902.10562 , 2019

    1902 arXiv

  19. [27]

    Pwclo-net: Deep lidar odometry in 3d point clouds using hierarchical embedding mask optimization,

    G. Wang, X. Wu, Z. Liu, and H. Wang, “Pwclo-net: Deep lidar odometry in 3d point clouds using hierarchical embedding mask optimization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 910–15 919

    2021

  20. [28]

    Lodonet: A deep neural network with 2d keypoint matching for 3d lidar odometry estimation,

    C. Zheng, Y . Lyu, M. Li, and Z. Zhang, “Lodonet: A deep neural network with 2d keypoint matching for 3d lidar odometry estimation,” in Proceedings of the 28th ACM international conference on multimedia , 2020, pp. 2391–2399

    2020

  21. [29]

    Illumination invariant imaging: Applications in robust vision-based localisation, mapping and classification for autonomous vehicles,

    W. Maddern, A. Stewart, C. McManus, B. Upcroft, W. Churchill, and P. Newman, “Illumination invariant imaging: Applications in robust vision-based localisation, mapping and classification for autonomous vehicles,” in Proceedings of the Visual Place Recognition in Changing Envir...

    2014

  22. [30]

    Dvlo: Deep visual-lidar odometry with local-to-global feature fusion and bi- directional structure alignment,

    J. Liu, D. Zhuo, Z. Feng, S. Zhu, C. Peng, Z. Liu, and H. Wang, “Dvlo: Deep visual-lidar odometry with local-to-global feature fusion and bi- directional structure alignment,” in European Conference on Computer Vision. Springer, 2024, pp. 475–493

    2024

  23. [31]

    Visual-lidar slam based on unsu- pervised multi-channel deep neural networks,

    Y . An, J. Shi, D. Gu, and Q. Liu, “Visual-lidar slam based on unsu- pervised multi-channel deep neural networks,” Cognitive Computation, vol. 14, no. 4, pp. 1496–1508, 2022

    2022

  24. [32]

    Penet: Towards precise and efficient image guided depth completion,

    M. Hu, S. Wang, B. Li, S. Ning, L. Fan, and X. Gong, “Penet: Towards precise and efficient image guided depth completion,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 13 656–13 662

    2021

  25. [33]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  26. [34]

    Cbam: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “Cbam: Convolutional block attention module,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 3–19

    2018

  27. [35]

    Liteflownet: A lightweight convo- lutional neural network for optical flow estimation,

    T.-W. Hui, X. Tang, and C. C. Loy, “Liteflownet: A lightweight convo- lutional neural network for optical flow estimation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 8981–8989

    2018

  28. [36]

    Maskflownet: Asymmetric feature matching with learnable occlusion mask,

    S. Zhao, Y . Sheng, Y . Dong, E. I. Chang, Y . Xu et al. , “Maskflownet: Asymmetric feature matching with learnable occlusion mask,” in Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 6278–6287

    2020

  29. [37]

    Just go with the flow: Self-supervised scene flow estimation,

    H. Mittal, B. Okorn, and D. Held, “Just go with the flow: Self-supervised scene flow estimation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 11 177–11 185

    2020

  30. [38]

    Rethinking optical flow from geometric matching consistent perspective,

    Q. Dong, C. Cao, and Y . Fu, “Rethinking optical flow from geometric matching consistent perspective,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 1337–1347

    2023

  31. [39]

    Depth- aware video frame interpolation,

    W. Bao, W.-S. Lai, C. Ma, X. Zhang, Z. Gao, and M.-H. Yang, “Depth- aware video frame interpolation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3703– 3712

    2019

  32. [40]

    Unsupervised learning of depth and ego-motion from video,

    T. Zhou, M. Brown, N. Snavely, and D. G. Lowe, “Unsupervised learning of depth and ego-motion from video,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 1851–1858

    2017

  33. [41]

    Generalizing to the open world: Deep visual odometry with online adaptation,

    S. Li, X. Wu, Y . Cao, and H. Zha, “Generalizing to the open world: Deep visual odometry with online adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 13 184–13 193

    2021

  34. [42]

    H-vlo: hybrid lidar-camera fusion for self-supervised odometry,

    E. Aydemir, N. Fetic, and M. Unel, “H-vlo: hybrid lidar-camera fusion for self-supervised odometry,” in 2022 IEEE/RSJ international confer- ence on intelligent robots and systems (IROS) . IEEE, 2022, pp. 3302– 3307

    2022

  35. [43]

    Dvl-slam: Sparse depth enhanced direct visual-lidar slam,

    Y .-S. Shin, Y . S. Park, and A. Kim, “Dvl-slam: Sparse depth enhanced direct visual-lidar slam,” Autonomous Robots, vol. 44, no. 2, pp. 115– 130, 2020

    2020

  36. [44]

    Lidar- monocular visual odometry using point and line features,

    S.-S. Huang, Z.-Y . Ma, T.-J. Mu, H. Fu, and S.-M. Hu, “Lidar- monocular visual odometry using point and line features,” in 2020 IEEE international conference on robotics and automation (ICRA) . IEEE, 2020, pp. 1091–1097

    2020

  37. [45]

    Efficient 3d deep lidar odometry,

    G. Wang, X. Wu, S. Jiang, Z. Liu, and H. Wang, “Efficient 3d deep lidar odometry,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 5, pp. 5749–5765, 2022

    2022

  38. [46]

    Selfvio: Self-supervised deep monocular visual–inertial odometry and depth estimation,

    Y . Almalioglu, M. Turan, M. R. U. Saputra, P. P. De Gusm ˜ao, A. Markham, and N. Trigoni, “Selfvio: Self-supervised deep monocular visual–inertial odometry and depth estimation,” Neural Networks , vol. 150, pp. 119–136, 2022

    2022

  39. [47]

    Un- supervised deep visual-inertial odometry with online error correction for rgb-d imagery,

    E. J. Shamwell, K. Lindgren, S. Leung, and W. D. Nothwang, “Un- supervised deep visual-inertial odometry with online error correction for rgb-d imagery,” IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 10, pp. 2478–2493, 2019

    2019

  40. [48]

    Self-supervised depth comple- tion from direct visual-lidar odometry in autonomous driving,

    Z. Song, J. Lu, Y . Yao, and J. Zhang, “Self-supervised depth comple- tion from direct visual-lidar odometry in autonomous driving,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 8, pp. 11 654–11 665, 2021

    2021

  41. [49]

    Recent advances in conventional and deep learning-based depth completion: A survey,

    Z. Xie, X. Yu, X. Gao, K. Li, and S. Shen, “Recent advances in conventional and deep learning-based depth completion: A survey,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 3, pp. 3395–3415, 2022

    2022

  42. [50]

    Multi-sensor fusion self-supervised deep odometry and depth estimation,

    Y . Wan, Q. Zhao, C. Guo, C. Xu, and L. Fang, “Multi-sensor fusion self-supervised deep odometry and depth estimation,” Remote Sensing, vol. 14, no. 5, p. 1228, 2022

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.