Pith. sign in

REVIEW 3 major objections 6 minor 64 references

DVLO4D: Deep Visual-Lidar Odometry with Sparse Spatial-temporal Fusion

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read DVLO4D claims that sparse LiDAR-query fusion with temporal memory banks achieves state-of-the-art visual-LiDAR odometry, reporting 0.73% translation error and 0.37°/100m rotation on KITTI 07-10 at 82 ms per frame.

desk verdict Solid engineering paper with a meaningful component integration, but the headline SOTA numbers are weakened by test-set hyperparameter selection and the 'global optimization' claim outruns what the 3-frame loss actually does. read the letter →

arxiv 2509.06023 v1 pith:KGAZGEKM submitted 2025-09-07 cs.CV

classification cs.CV
keywords visual-LiDARodometrysparsequeryfusiontemporalposememorycollectiveaveragelossdeformablecross-attentionKITTIArgoversereal-timelocalization
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

DVLO4D is a learning-based visual-LiDAR odometry system that fuses camera image features into sparse LiDAR point queries rather than densely aligning the two modalities. It adds a temporal interaction module with two memory banks—one for past fused features and one for past poses—that supplies a better starting pose and refines each estimate using history. Training uses video clips and a collective average loss over short sub-clips, which the paper says reduces scale drift on long sequences. On KITTI 07-10 the paper reports a mean translation error of 0.73% and rotation error of 0.37°/100m, a 12% translation improvement over DVLO and 18% over EfficientLO, with 0.089 ATE on Argoverse and 82 ms inference. If these numbers hold, the system is both more accurate and fast enough for real-time onboard use.

What carries the argument

The load-bearing machinery is a set of three interacting modules. Sparse Query Fusion (SQF) treats each LiDAR point as a query, projects it into the image plane, samples visual features at learned offsets, and applies multi-head cross-attention, which replaces dense fusion with a sparse point-to-pixel correspondence. The Temporal Interaction and Update (TIU) module stores recent fused features and poses in a Memory Feature Bank and a Memory Pose Bank, encodes them with a transformer-plus-LSTM, and uses them to initialize and refine the current pose. Collective Average Loss (CAL) averages the pose loss over sub-clips of three frames during training, which the paper argues lets the network opt

What would settle it

Run the trained DVLO4D in one continuous pass over a KITTI sequence longer than the 6s training clips (e.g., sequence 00 or 02) and compute translation/rotation drift per kilometer. If the error per kilometer is no better than an ablated model trained with T_s=1 (frame-to-frame loss), the central claim that Collective Average Loss suppresses long-sequence scale drift is falsified. A second check: split a long sequence into 3-frame clips, compare DVLO4D's accumulated pose against ground truth at the end; if drift accumulates linearly with no periodic correction, the temporal memory is not contr

Watch

Extended reading notes

Core claim

The central claim is that sparse, query-based cross-modal fusion combined with explicit temporal memory yields state-of-the-art visual-LiDAR odometry. LiDAR features act as queries whose 3D positions become positional embeddings; sampled image features act as keys and values in deformable cross-attention, so the model learns where in the image to look for each point. A Temporal Interaction and Update module then combines a Memory Feature Bank and a Memory Pose Bank with the current cost volume, and a temporal encoder (self-attention plus LSTM) produces an initial pose that is iteratively refined. To fight drift, Temporal Clip Training splits each 6-second clip into 0.3-second sub-clips and a

Load-bearing premise

The method's drift control rests on averaging the training loss over only 0.3-second sub-clips and keeping a 3-second pose memory; the paper does not show that this short-horizon mechanism bounds errors over the long trajectories it claims to optimize.

Editorial extensions

If this is right

  • At 82 ms per frame, the method clears the 100 ms real-time threshold for 10 Hz KITTI-style operation, so the architecture is a candidate for live autonomous driving stacks.
  • Robustness experiments indicate that temporal memory makes the estimator substantially less sensitive to LiDAR downsampling and sensor noise than EfficientLO, which matters when sensors degrade in the field.
  • The same sparse query mechanism, using 3D points as queries, can be applied to other projected sensor modalities without changing the image-side encoder, as the paper hints in its discussion of radar.
  • Because CAL aggregates loss over sub-clips, the training procedure can be reused by any frame-to-frame pose regressor that outputs per-frame poses; it is not specific to the query fusion design.
  • The paper reports 0.089 ATE on Argoverse under the DSLO protocol, indicating the design transfers to another dataset beyond KITTI.

Reading between the lines

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

  • The paper's long-sequence claim is only tested on KITTI sequences 07-10 with training clips of 6 s; a stringent check that goes beyond the paper would be a full traversal of KITTI 00-10 and reporting drift per kilometer. If CAL's 0.3 s averaging does not bound correlated errors, the 'global optimization' wording overstates what is shown.
  • The temporal memory bank may be a learned replacement for a fixed-lag smoother or Kalman filter. One could test this by replacing the bank with a simple moving average of poses and measuring the gap, isolating what the attention mechanism contributes.
  • SQF's sparse querying could be extended to 4D radar or ultrasonic sensors by using their 3D positions as queries, a direction the paper mentions but does not evaluate. A concrete test would be training the same fusion head on radar-LiDAR pairs and comparing pose error on the same sequences.
  • The reported 3% error increase under 5 Hz LiDAR suggests the temporal pose bank is doing predictive interpolation; a direct experiment would compare against a version that drops the Memory Pose Bank and simply repeats the last pose, to see whether the learned update rather than the memory content is responsible.
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 paper proposes DVLO4D, a learning-based visual-LiDAR odometry system that combines (1) Sparse Query Fusion, where LiDAR points serve as queries to sample and attend to camera features; (2) a Temporal Interaction and Update module with Memory Feature and Pose Banks to refine pose estimates using historical frames; and (3) Temporal Clip Training with a Collective Average Loss to train on short clips and aggregate losses over multiple frames. The method is evaluated on KITTI odometry (train 00-06, test 07-10) and Argoverse, reporting 0.73% mean translation error and 0.37°/100m rotation error on KITTI 07-10, 0.089 ATE on Argoverse, and an inference time of 82 ms. The paper claims state-of-the-art accuracy and robustness relative to DVLO, EfficientLO, and other published visual, LiDAR, and multi-modal odometry methods.

Significance. If the empirical claims hold, DVLO4D would be a meaningful advance: it improves on strong learned LiDAR and multi-modal odometry baselines, maintains real-time speed, and shows favorable robustness to LiDAR rate reduction and noise. The ablation structure is clear, the individual module gains are internally consistent, and the Argoverse result suggests some generalization. However, the headline claim is currently under-supported because the final hyperparameters (T_s, T_h) appear to be chosen by evaluating directly on the KITTI test sequences 07-10, with no separate validation split; the reported margins over baselines are therefore at risk of being inflated by test-set selection. In addition, the claimed 'global optimization' for long sequences is not established by the actual 3-frame Collective Average Loss. No code or checkpoints are provided, so the empirical results are not independently reproducible.

major comments (3)
  1. [§IV-B, Table VIII] The hyperparameters T_s and T_h are selected by sweeping on KITTI 07-10, the same sequences used for the headline SOTA claim. Table VIII reports mean trel/rrel on 07-10 for T_s ∈ {1,3} and T_h ∈ {15,30,45}, and the chosen configuration (T_s=3, T_h=30) is exactly the one reported as the final result (0.73, 0.37). No validation split is mentioned anywhere in the paper, and the same test sequences are used for the ablations (Table VI), robustness checks (Table V), and final comparison (Table I). Under this protocol the reported numbers are the result of test-set optimization and may overestimate performance. Please add a held-out validation split for hyperparameter and ablation decisions, report results over multiple random seeds with mean and standard deviation, and clearly state which decisions were made before seeing 07-10.
  2. [§III-E, Eq. (10), §IV-B] The Collective Average Loss in Eq. (10) averages losses only over sub-clips of length T_s=3 frames (0.3 s), even though the abstract and conclusion claim 'global optimization' that reduces 'scale drift over long sequences.' A 3-frame loss average, combined with a 30-frame memory bank, does not by construction bound errors over trajectories of hundreds of meters, and no experiment measures drift as a function of sequence length or shows that gradients are propagated across the full clip. This is a load-bearing gap for one of the three claimed contributions. Please either provide a formal argument relating the local clip loss to long-term drift, or add an experiment on long sequences (e.g., full KITTI 09/10 trajectories) that directly evaluates the claimed long-horizon benefit.
  3. [§IV-C, Tables I and III] The comparative SOTA claim is not fully controlled because training/evaluation protocols differ across methods. For example, Table III explicitly notes that DVLO4D is trained on 00-06 while other multi-modal methods are trained on 00-08; in Table I, the footnote about training sequences covers only the visual odometry methods, leaving the LiDAR and multi-modal baselines unspecified. Single-run results without error bars further weaken the comparison: the reported margin over DVLO is 0.09 percentage points in mean trel. Please state precisely the training splits for every baseline, or re-run the strongest baselines under a common protocol, and provide variance estimates or at least multiple-seed results.
minor comments (6)
  1. [§III-E, Table VI] The heading 'Collective Averge Loss' and the corresponding table entry contain a typo; should be 'Average.'
  2. [§III-B, Eq. (2)] In Eq. (2), the text says 'F_cam acts as both the key and value,' but the arguments of MHCA are F_sample; please fix the notation for clarity.
  3. [Fig. 2] Some symbols in Fig. 2 (e.g., t_t^3, q_t^3, E_ego) are not formally defined in the caption or in the body text. Please add definitions or a legend.
  4. [Table VI] The rows with two check marks do not make explicit which module is removed. Use a clear 'w/o' notation or a column for the omitted module so the reader does not have to infer the ablation from the text.
  5. [§IV-E] The phrase 'demonstrating CAL's essence' is awkward; 'essential role' would be clearer.
  6. [Table V] The row labels are inconsistent: '10HZ', '5HZ', and 'Add noise' vs 'noise'. Please unify formatting and state exactly what noise is added (mean/variance) and how the 5Hz data are generated.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the SOTA claim rests on external KITTI/Argoverse benchmarks; self-citations are building blocks, not load-bearing reductions.

full rationale

DVLO4D's central claim is an empirical accuracy/robustness result measured on KITTI 07-10 and Argoverse, against external baselines (DVLO, EfficientLO, H-VLO, DSLO, etc.). The method sections define SQF, TIU, and CAL through explicit equations (Eqs. 1-10); these are ordinary network/loss definitions and are not fitted to the reported test errors. The paper inherits standard components from prior work, including the authors' own DVLO [19] and TransLO [23], but these are used as modular building blocks (projection, global adaptive fusion, cost volume) and the headline improvement is measured against DVLO itself, so the result is not equivalent to the cited inputs by construction. The Collective Average Loss (Eq. 10) is a training objective, not a prediction; it cannot by itself force the reported test-set numbers. The main validity concern is that hyperparameters T_s and T_h appear to be selected using the same test sequences 07-10 (Table VIII) with no separate validation split, and no error bars are reported; this is a benchmark-integrity / statistical issue rather than circular reasoning. No circularity steps meet the quoted-equivalence standard.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on domain assumptions about calibration, projection, memory reliability, and the sufficiency of the 3-frame loss. The hyperparameters T_s, T_h, alpha, beta, and query counts are hand-chosen or test-selected. No new physical entities are introduced.

free parameters (3)
  • Temporal clip length T_s and history length T_h = T_s=3, T_h=30
    Chosen via ablation on KITTI test sequences 07-10 (Table VIII); the final model uses the best-performing setting, effectively tuning on the benchmark.
  • Layer loss weights alpha_l and beta = alpha=[1.6,0.8,0.4,0.8], beta=0.8
    Hand-set in Sec IV-B; Eq. (10) weights four layers plus the refined loss without a stated model-selection criterion.
  • LiDAR query counts per feature level = 116, 228, 904, 3600
    Chosen for the KITTI sensor geometry in Sec IV-B; this affects the sparsity-efficiency tradeoff and fusion quality.
assumptions (5)
  • domain assumption Camera-LiDAR calibration and synchronization are correct, and LiDAR points project to valid image locations via T_k(P_i).
    Eq. (1) in Sec III-B; if calibration or timing is off, sampled visual features are misaligned and the sparse fusion corrupts pose estimation. Calibration error is not quantified.
  • domain assumption Historical poses and features stored in MPB and MFB provide a reliable prior whose errors do not systematically compound through the temporal encoder.
    Sec III-D; if the memory poses themselves are drifting, the temporal update can reinforce drift rather than correct it.
  • domain assumption Spherical projection of LiDAR onto a cylindrical pseudo-image retains the geometric information needed for odometry.
    Sec III-A follows [22], [39]; the projection choice is assumed not to discard structure that the pose estimator needs.
  • ad hoc to paper A 3-frame collective average loss is a sufficient proxy for long-sequence trajectory consistency.
    Eq. (10), Sec III-E; no derivation or experiment shows that averaging losses over T_s=3 frames bounds scale drift over long sequences. This is the bridge between the loss and the long-odometry claim.
  • domain assumption The fusion mask from Sec IV-B correctly identifies which LiDAR queries overlap the camera image.
    The mask is described only as 'designed'; its construction and failure modes are not specified, yet it controls which queries can be fused with image features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DVLO4D: Deep Visual-Lidar Odometry with Sparse Spatial-temporal Fusion." pith.science (2026). https://pith.science/paper/KGAZGEKM

@misc{pith2026250906023,
  author       = {Pith},
  title        = {Pith review of: DVLO4D: Deep Visual-Lidar Odometry with Sparse Spatial-temporal Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGAZGEKM}},
  note         = {Machine review of arXiv:2509.06023}
}
read the original abstract

Visual-LiDAR odometry is a critical component for autonomous system localization, yet achieving high accuracy and strong robustness remains a challenge. Traditional approaches commonly struggle with sensor misalignment, fail to fully leverage temporal information, and require extensive manual tuning to handle diverse sensor configurations. To address these problems, we introduce DVLO4D, a novel visual-LiDAR odometry framework that leverages sparse spatial-temporal fusion to enhance accuracy and robustness. Our approach proposes three key innovations: (1) Sparse Query Fusion, which utilizes sparse LiDAR queries for effective multi-modal data fusion; (2) a Temporal Interaction and Update module that integrates temporally-predicted positions with current frame data, providing better initialization values for pose estimation and enhancing model's robustness against accumulative errors; and (3) a Temporal Clip Training strategy combined with a Collective Average Loss mechanism that aggregates losses across multiple frames, enabling global optimization and reducing the scale drift over long sequences. Extensive experiments on the KITTI and Argoverse Odometry dataset demonstrate the superiority of our proposed DVLO4D, which achieves state-of-the-art performance in terms of both pose accuracy and robustness. Additionally, our method has high efficiency, with an inference time of 82 ms, possessing the potential for the real-time deployment.

Figures

Figures reproduced from arXiv: 2509.06023 by the authors.

Figure 1
Figure 1. The pipeline of DVLO4D. The Sparse Query Fusion (SQF) mechanism utilizes sparse LiDAR queries to effectively fuse multi-modal data, enabling the integration of camera and LiDAR features. The Temporal Interaction and Update (TIU) module refine pose estimates by leveraging the Memory Feature Bank (MFB) and the Memory Pose Bank (MPB) and integrating them with the current input, improving the accuracy and robustness of … view at source ↗
Figure 2
Figure 2. Temporal Interaction and Update module. To refine the pose, we initialize the ego instance feature Eego using the coarsest cost volume feature map L−1. It pro￾vides two key benefits: the coarsest feature map encapsulates the semantic and geometric context of the driving scene, and its dense representation complements sparse representation, particularly in handling dynamic objects and occlusions. For the historical q… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 49 canonical work pages

  1. [1]

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

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

  2. [2]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wanget al., “Planning-oriented autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 17 853–17 862

  3. [3]

    Og-gaussian: Occupancy based street gaussians for autonomous driving,

    Y . Shen, X. Zhang, Y . Duan, S. Zhang, H. Li, Y . Wu, J. Ji, and Y . Zhang, “Og-gaussian: Occupancy based street gaussians for autonomous driving,” 2025. [Online]. Available: https://arxiv.org/abs/2502.14235

  4. [4]

    Real-Time LiDAR Point Cloud Compression and Transmission for Resource-constrained Robots

    Y . Cao, Y . Wang, and H. Chen, “Real-time lidar point cloud compres- sion and transmission for resource-constrained robots,”arXiv preprint arXiv:2502.06123, 2025

  5. [5]

    Compact 3d gaussian splatting for dense visual slam,

    T. Deng, Y . Chen, L. Zhang, J. Yang, S. Yuan, J. Liu, D. Wang, H. Wang, and W. Chen, “Compact 3d gaussian splatting for dense visual slam,”arXiv preprint arXiv:2403.11247, 2024

  6. [6]

    Deep patch visual odometry,

    Z. Teed, L. Lipson, and J. Deng, “Deep patch visual odometry,” Advances in Neural Information Processing Systems, vol. 36, 2024

  7. [7]

    Multi-camera collaborative depth prediction via consistent structure estimation,

    J. Xu, X. Liu, Y . Bai, J. Jiang, K. Wang, X. Chen, and X. Ji, “Multi-camera collaborative depth prediction via consistent structure estimation,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 2730–2738

  8. [8]

    Sni-slam: Semantic neural implicit slam,

    S. Zhu, G. Wang, H. Blum, J. Liu, L. Song, M. Pollefeys, and H. Wang, “Sni-slam: Semantic neural implicit slam,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 21 167–21 177

Show all 64 references
  1. [9]

    Toward learning-based visuomotor navigation with neural radiance fields,

    Q. Liu, N. Chen, Z. Liu, and H. Wang, “Toward learning-based visuomotor navigation with neural radiance fields,”IEEE Transactions on Industrial Informatics, 2024

  2. [10]

    Boosting explore-exploit behavior for navigating vehicle by maintaining informative topological frontier,

    Q. Liu, X. Cui, Z. Liu, and H. Wang, “Boosting explore-exploit behavior for navigating vehicle by maintaining informative topological frontier,”IEEE Transactions on Intelligent Vehicles, 2024

  3. [11]

    Talk2radar: Bridging natural language with 4d mmwave radar for 3d referring expression com- prehension,

    R. Guan, R. Zhang, N. Ouyang, J. Liu, K. L. Man, X. Cai, M. Xu, J. Smith, E. G. Lim, Y . Yueet al., “Talk2radar: Bridging natural language with 4d mmwave radar for 3d referring expression com- prehension,”arXiv preprint arXiv:2405.12821, 2024

  4. [12]

    Detect closer surfaces that can be seen: New modeling and evaluation in cross-domain 3d object detection,

    R. Zhang, Y . Wu, J. Lee, X. Cai, and A. Prugel-Bennett, “Detect closer surfaces that can be seen: New modeling and evaluation in cross-domain 3d object detection,” inECAI 2024. IOS Press, 2024, pp. 65–72

  5. [13]

    Limo: Lidar-monocular visual odometry,

    J. Graeter, A. Wilczynski, and M. Lauer, “Limo: Lidar-monocular visual odometry,” in2018 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2018, pp. 7872–7879

  6. [14]

    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,” in2020 IEEE international conference on robotics and automation (ICRA). IEEE, 2020, pp. 1091–1097

  7. [15]

    Dv-loam: Direct visual lidar odometry and mapping,

    W. Wang, J. Liu, C. Wang, B. Luo, and C. Zhang, “Dv-loam: Direct visual lidar odometry and mapping,”Remote Sensing, vol. 13, no. 16, p. 3340, 2021

  8. [16]

    Sdv-loam: semi- direct visual–lidar odometry and mapping,

    Z. Yuan, Q. Wang, K. Cheng, T. Hao, and X. Yang, “Sdv-loam: semi- direct visual–lidar odometry and mapping,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 11 203– 11 220, 2023

  9. [17]

    Long-term visual simulta- neous localization and mapping: Using a bayesian persistence filter- based global map prediction,

    T. Deng, H. Xie, J. Wang, and W. Chen, “Long-term visual simulta- neous localization and mapping: Using a bayesian persistence filter- based global map prediction,”IEEE Robotics & Automation Magazine, vol. 30, no. 1, pp. 36–49, 2023

  10. [18]

    4drvo-net: Deep 4d radar–visual odometry using multi-modal and multi-scale adaptive fusion,

    G. Zhuoins, S. Lu, L. Xiong, H. Zhouins, L. Zheng, and M. Zhou, “4drvo-net: Deep 4d radar–visual odometry using multi-modal and multi-scale adaptive fusion,”IEEE Transactions on Intelligent Vehi- cles, 2023

  11. [19]

    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,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 475–493

  12. [20]

    Keyframe-based visual-inertial slam using nonlinear optimization,

    S. Leutenegger, P. Furgale, V . Rabaud, M. Chli, K. Konolige, and R. Siegwart, “Keyframe-based visual-inertial slam using nonlinear optimization,”Proceedings of Robotis Science and Systems (RSS) 2013, 2013

  13. [21]

    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

  14. [22]

    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

  15. [23]

    Translo: A window-based masked point transformer framework for large-scale lidar odometry,

    J. Liu, G. Wang, C. Jiang, Z. Liu, and H. Wang, “Translo: A window-based masked point transformer framework for large-scale lidar odometry,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 2, 2023, pp. 1683–1691

  16. [24]

    Deformable detr: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable detr: Deformable transformers for end-to-end object detection,”arXiv preprint arXiv:2010.04159, 2020

  17. [25]

    A real-time method for depth enhanced visual odometry,

    J. Zhang, M. Kaess, and S. Singh, “A real-time method for depth enhanced visual odometry,”Autonomous Robots, vol. 41, pp. 31–43, 2017

  18. [26]

    Sdge: Stereo guided depth estimation for 360° camera sets,

    J. Xu, W. Yin, D. Gong, J. Jiang, and X. Liu, “Sdge: Stereo guided depth estimation for 360° camera sets,” in2024 IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 11 179–11 186

  19. [27]

    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,” in2015 IEEE international conference on robotics and automation (ICRA). IEEE, 2015, pp. 2174–2181

  20. [28]

    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

  21. [29]

    Lip-loc: Lidar image pretraining for cross-modal localization,

    S. Shubodh, M. Omama, H. Zaidi, U. S. Parihar, and M. Krishna, “Lip-loc: Lidar image pretraining for cross-modal localization,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 948–957

  22. [30]

    Adafusion: Visual-lidar fusion with adaptive weights for place recognition,

    H. Lai, P. Yin, and S. Scherer, “Adafusion: Visual-lidar fusion with adaptive weights for place recognition,”IEEE Robotics and Automa- tion Letters, vol. 7, no. 4, pp. 12 038–12 045, 2022

  23. [31]

    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,” in2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017, pp. 2043–2050

  24. [32]

    Particle video revisited: Tracking through occlusions using point trajectories,

    A. W. Harley, Z. Fang, and K. Fragkiadaki, “Particle video revisited: Tracking through occlusions using point trajectories,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 59–75

  25. [33]

    Difflow3d: Toward robust uncertainty-aware scene flow estimation with iterative diffusion-based refinement,

    J. Liu, G. Wang, W. Ye, C. Jiang, J. Han, Z. Liu, G. Zhang, D. Du, and H. Wang, “Difflow3d: Toward robust uncertainty-aware scene flow estimation with iterative diffusion-based refinement,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2...

  26. [34]

    3dsflabelling: Boosting 3d scene flow estimation by pseudo auto-labelling,

    C. Jiang, G. Wang, J. Liu, H. Wang, Z. Ma, Z. Liu, Z. Liang, Y . Shan, and D. Du, “3dsflabelling: Boosting 3d scene flow estimation by pseudo auto-labelling,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 173–15 183

  27. [35]

    Tap-vid: A benchmark for tracking any point in a video,

    C. Doersch, A. Gupta, L. Markeeva, A. Recasens, L. Smaira, Y . Aytar, J. Carreira, A. Zisserman, and Y . Yang, “Tap-vid: A benchmark for tracking any point in a video,”Advances in Neural Information Processing Systems, vol. 35, pp. 13 610–13 626, 2022

  28. [36]

    Boot- stap: Bootstrapped training for tracking-any-point,

    C. Doersch, Y . Yang, D. Gokay, P. Luc, S. Koppula, A. Gupta, J. Heyward, R. Goroshin, J. Carreira, and A. Zisserman, “Boot- stap: Bootstrapped training for tracking-any-point,”arXiv preprint arXiv:2402.00847, 2024

  29. [37]

    Tracking everything everywhere all at once,

    Q. Wang, Y .-Y . Chang, R. Cai, Z. Li, B. Hariharan, A. Holynski, and N. Snavely, “Tracking everything everywhere all at once,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 795–19 806

  30. [38]

    Leap-vo: Long-term effective any point tracking for visual odometry,

    W. Chen, L. Chen, R. Wang, and M. Pollefeys, “Leap-vo: Long-term effective any point tracking for visual odometry,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 844–19 853

  31. [39]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 8473–8482

  32. [40]

    Regformer: An efficient projection-aware transformer network for large-scale point cloud registration,

    J. Liu, G. Wang, Z. Liu, C. Jiang, M. Pollefeys, and H. Wang, “Regformer: An efficient projection-aware transformer network for large-scale point cloud registration,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8451–8460

  33. [41]

    Epnet: Enhancing point features with image semantics for 3d object detection,

    T. Huang, Z. Liu, X. Chen, and X. Bai, “Epnet: Enhancing point features with image semantics for 3d object detection,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16. Springer, 2020, pp. 35–52

  34. [42]

    Unibev- fusion: Unified radar-vision bevfusion for 3d object detection,

    H. Zhao, R. Guan, T. Wu, K. L. Man, L. Yu, and Y . Yue, “Unibev- fusion: Unified radar-vision bevfusion for 3d object detection,”arXiv preprint arXiv:2409.14751, 2024

  35. [43]

    Achelous++: Power-oriented water-surface panoptic perception framework on edge devices based on vision-radar fusion and pruning of heterogeneous modalities,

    R. Guan, H. Zhao, S. Yao, K. L. Man, X. Zhu, L. Yu, Y . Yue, J. Smith, E. G. Lim, W. Dinget al., “Achelous++: Power-oriented water-surface panoptic perception framework on edge devices based on vision-radar fusion and pruning of heterogeneous modalities,”arXiv preprint arXiv:2...

  36. [44]

    Delflow: Dense efficient learning of scene flow for large-scale point clouds,

    C. Peng, G. Wang, X. W. Lo, X. Wu, C. Xu, M. Tomizuka, W. Zhan, and H. Wang, “Delflow: Dense efficient learning of scene flow for large-scale point clouds,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023, pp. 16 901–16 910

  37. [45]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”Transactions of the ASME–Journal of Basic Engineering, vol. 82, no. Series D, pp. 35–45, 1960

  38. [46]

    Motr: End-to-end multiple-object tracking with transformer,

    F. Zeng, B. Dong, Y . Zhang, T. Wang, X. Zhang, and Y . Wei, “Motr: End-to-end multiple-object tracking with transformer,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 659–675

  39. [47]

    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,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1851–1858

  40. [48]

    Tartanvo: A generalizable learning- based vo,

    W. Wang, Y . Hu, and S. Scherer, “Tartanvo: A generalizable learning- based vo,” inConference on Robot Learning. PMLR, 2021, pp. 1761–1772

  41. [49]

    High accuracy monocular sfm and scale correction for autonomous driving,

    S. Song, M. Chandraker, and C. C. Guest, “High accuracy monocular sfm and scale correction for autonomous driving,”IEEE transactions on pattern analysis and machine intelligence, vol. 38, no. 4, pp. 730– 743, 2015

  42. [50]

    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

  43. [51]

    Dynamic object-aware visual odometry (vo) estimation based on optical flow matching,

    H. M. Cho and E. Kim, “Dynamic object-aware visual odometry (vo) estimation based on optical flow matching,”IEEE Access, vol. 11, pp. 11 642–11 651, 2023

  44. [52]

    Delo: Deep evidential lidar odometry using partial optimal transport,

    S. A. Ali, D. Aouada, G. Reis, and D. Stricker, “Delo: Deep evidential lidar odometry using partial optimal transport,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 4517–4526

  45. [53]

    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,” in2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). IEEE, 2022, pp. 3302–3307

  46. [54]

    Argoverse: 3d tracking and forecasting with rich maps,

    M.-F. Chang, J. Lambert, P. Sangkloy, J. Singh, S. Bak, A. Hartnett, D. Wang, P. Carr, S. Lucey, D. Ramananet al., “Argoverse: 3d tracking and forecasting with rich maps,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 8748–8757

  47. [55]

    What’s in my lidar odometry toolbox?

    P. Dellenbach, J.-E. Deschaud, B. Jacquet, and F. Goulette, “What’s in my lidar odometry toolbox?” inIEEE/RSJ International Conference on Intelligent Robots and Systems, 2021, pp. 4429–4436

  48. [56]

    A-LOAM: advanced implementation of loam,

    T. Qin and S. Cao, “A-LOAM: advanced implementation of loam,” in https://github.com/HKUST-Aerial-Robotics/A-LOAM

  49. [57]

    Dslo: Deep sequence lidar odometry based on incon- sistent spatio-temporal propagation,

    H. Zhang, G. Wang, X. Wu, C. Xu, M. Ding, M. Tomizuka, W. Zhan, and H. Wang, “Dslo: Deep sequence lidar odometry based on incon- sistent spatio-temporal propagation,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 10 672–10 677

  50. [58]

    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,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021, pp. 3844–3852

  51. [59]

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

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

  52. [60]

    Unsu- pervised deep visual-inertial odometry with online error correction for rgb-d imagery,

    E. J. Shamwell, K. Lindgren, S. Leung, and W. D. Nothwang, “Unsu- pervised 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

  53. [61]

    Multi-modal feature constraint based tightly coupled monocular visual-lidar odometry and mapping,

    C. Shu and Y . Luo, “Multi-modal feature constraint based tightly coupled monocular visual-lidar odometry and mapping,”IEEE Trans- actions on Intelligent Vehicles, vol. 8, no. 5, pp. 3384–3393, 2022

  54. [62]

    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

  55. [63]

    Loam: Lidar odometry and mapping in real-time

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

  56. [64]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.