Pith. sign in

REVIEW 4 major objections 5 minor 28 references

REMOTE: Real-time Ego-motion Tracking for Various Endoscopes via Multimodal Visual Feature Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A multimodal feature-learning network called REMOTE estimates an endoscope's motion directly from video and reports the most accurate ego-motion tracking on three endoscopic datasets while running above 30 frames per second.

desk verdict A plausible, well-ablated engineering contribution to endoscopic ego-motion tracking whose headline numbers on the authors' own dataset need a transparent split, code, and significance checks before they can be trusted. read the letter →

arxiv 2501.18124 v2 pith:V36ZX3D2 submitted 2025-01-30 cs.CV cs.AI

Liangjing Shao , Benshuang Chen , Shuting Zhao , Xinrong Chen This is my paper · ORCID
classification cs.CVcs.AI
keywords ego-motiontrackingendoscopyvisualodometryopticalflowmultimodalfeaturelearningattentionmechanismrelativeposeestimationreal-time
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 REMOTE, a deep learning framework that estimates the relative pose between consecutive endoscopic video frames and chains these estimates into an absolute endoscope trajectory. The central claim is that by fusing three visual modalities — scene features from each frame, motion features from optical flow, and a joint feature from the concatenated frame pair — with an attention-based feature extractor and a specialized pose decoder, the network achieves the most accurate ego-motion tracking reported on three endoscopic datasets (nasal endoscopy, simulated colonoscopy, and ex-vivo intestine endoscopy) while running above 30 frames per second. If correct, this would make vision-only real-time navigation of endoscopes practical without external optical or magnetic tracking hardware.

What carries the argument

The load-bearing mechanism is the multimodal feature fusion pipeline: a shared feature extractor processes the two frames and the optical flow to give scene features and a motion feature; a separate joint-feature branch concatenates the two frames and runs them through the first two ResNet-34 layers, each followed by an attention module that pools feature maps along height, width, and channel orientations, mixes max- and average-pooled attention weights with learnable scalars, and sums the three reweighted views. All features are concatenated and fed to a pose decoder built from a 1×1 convolution, layer normalization, and two depthwise-separable-convolution residual blocks; the decoder outputs a 7-dimensional relative pose (3 translation, 4 quaternion) trained with a geometric loss on translation and quaternion logarithm. Optical flow ground truth for finetuning is generated by the Dual TVL1 algorithm.

What would settle it

A reproduction that runs the same three datasets with strictly separated subject/trajectory splits, identical hyperparameter tuning for every method, and released training code would settle the claim: if NEPose ATE stays within noise of PoseCorrNet's 8.63 mm instead of 4.01 mm, the stated advantage would disappear.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that REMOTE outperforms seven prior ego-motion methods on all three datasets: on NEPose, absolute translation error drops to 4.01±2.92 mm from the best previous 8.63±5.49 mm; on SimCol, ATE is 5.43±3.89 cm versus 6.08±4.87 cm; on EndoSLAM, ATE is 9.86±4.28 cm versus 10.04±5.68 cm. The framework predicts the relative pose transformation between a frame and an earlier frame four timesteps back, using optical flow computed by a finetuned FastFlowNet, scene features and a joint feature from the concatenated frame pair, then recovers the absolute pose by multiplying the chain of relative transformations from a known initial pose. The authors present ablations showing each modality and each proposed module contributes to the final accuracy.

Load-bearing premise

The accuracy claim assumes the test frames come from subjects or trajectories that were not seen during training and that all baseline methods were tuned with equal care, but the paper does not document subject-exclusive splits, shared hyperparameter search, or released code.

Editorial extensions

If this is right

  • If the claimed accuracy holds, endoscope navigation could be driven purely by the video stream, removing the occlusion and electromagnetic-interference problems of optical and magnetic trackers.
  • The same framework runs on nasal, colon, and intestine imagery, suggesting one trained model can serve multiple endoscopic procedures instead of a per-domain tracker.
  • Because the network predicts relative pose without recurrent layers, it can be paired with existing SLAM back-ends or loop-closure methods to reduce long-sequence drift.
  • The reported inference speed (31–94 fps on an RTX 4090) positions the method for intraoperative use rather than offline analysis.
  • The finetuning recipe for optical flow on endoscopic data is a reusable component for other self-motion estimators in low-texture, specular scenes.

Reading between the lines

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

  • Beyond the reported comparisons, the multimodal fusion design (separate scene and motion branches plus a concatenation branch) could be tested on other monocular odometry tasks in texture-poor environments, such as inside the urinary tract or sinuses, where appearance changes are dominated by motion and specularity.
  • The paper does not demonstrate that the NEPose test subjects are disjoint from the training subjects; a strict subject-exclusive split would clarify how much of the ATE gain is genuine generalization rather than memorization of individual anatomy.
  • One extension the authors leave implicit is combining the relative pose chain with global optimization or a lightweight SLAM loop closure, which could reduce the drift visible in long trajectories like the NEPose 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 proposes REMOTE, a supervised multi-modal network for real-time endoscopic ego-motion tracking. Given a pair of adjacent frames, the method concatenates scene features from the two frames, a motion feature from an optical-flow branch, and a joint feature from an attention-based extractor, then predicts the relative pose with a custom decoder; absolute trajectories are obtained by chaining relative poses. Experiments are reported on three endoscopic datasets (NEPose, SimCol, EndoSLAM) with comparisons to seven prior methods. The central claim, stated in Section IV-D, is that the proposed method achieves the most accurate relative pose estimation and ego-motion tracking in these diverse endoscopic scenes while running above 30 frames per second. The paper also includes ablations on input modalities and on the proposed feature extractor and pose decoder.

Significance. If the reported results hold, the paper would provide a meaningful advance: a real-time endoscopic ego-motion tracker that improves accuracy on three datasets of different anatomical scenes, with ablations suggesting that each input modality and each proposed module contributes to performance. The authors are to be credited for testing on three datasets, for including qualitative trajectory comparisons, and for reporting inference speed on all three datasets. However, the evidence for the central claim currently rests on one very large improvement on the authors' own NEPose dataset, while the improvements on SimCol and EndoSLAM are small relative to the reported standard deviations. The lack of a verified subject-disjoint split, the absence of significance tests or multiple-seed results, and the incomplete specification of baseline training conditions make the magnitude of the claimed advantage difficult to assess. The method itself is straightforward supervised pose regression, so the main novelty is architectural; the value of that novelty depends on the reliability of the comparison.

major comments (4)
  1. [IV-A.1 and IV-D] The NEPose split description is internally inconsistent. Section IV-A.1 states that NEPose contains 50 videos from 16 subjects but then says that 4328 frames from 30 different subjects are randomly chosen for training and validation, while Section IV-D reports evaluation on 1308 frames from 12 different subjects. Since 30 + 12 exceeds 16, the text cannot be correct as written. More importantly, the paper never establishes that the 12 test subjects are disjoint from the training subjects. Because NEPose is the authors' own dataset and yields by far the largest improvement in Table I (ATE 4.01±2.92 mm versus 8.63±5.49 mm for PoseCorrNet), the central claim depends critically on this split. The authors must correct the subject/video counts and provide a verifiable subject-disjoint protocol, for example by listing subject IDs in training and test sets and reporting per-subject errors.
  2. [Table I and IV-D] The claimed superiority over previous methods is not statistically supported. On SimCol, the ATE margin over PoseCorrNet is 5.43±3.89 cm versus 6.08±4.87 cm; on EndoSLAM, the margin over OffsetNet is 9.86±4.28 cm versus 10.04±5.68 cm. In both cases the difference is within one standard deviation of the two methods, and no significance tests, confidence intervals, or multiple-seed runs are reported. The narrative that the method is 'the most accurate' on all three datasets should be backed by paired statistical tests or by a clear statement of how many trajectories improved and by how much, rather than by mean values alone.
  3. [III-B.1 and IV-B] The optical-flow branch is finetuned on pseudo-labels computed by Dual TVL1, but the manuscript does not state whether these pseudo-labels are generated only from training frames or also from frames in the test splits of the three datasets. If any test-sequence frames are included in the optical-flow finetuning, the comparison is not fair to the baselines. Please specify the exact split used for optical-flow finetuning, and if necessary rerun the experiments with finetuning restricted to training frames only.
  4. [IV-B and IV-D] The text says that 'the same implementation described in IV-B is used in the experiments of all methods,' but the implementation details for the seven baseline methods are not provided. In particular, no information is given on the baselines' learning rates, number of epochs, loss weights, or early stopping criteria. Since the margins on SimCol and EndoSLAM are small, a fair comparison requires either releasing the baseline training configurations or following each baseline's original protocol as closely as possible and stating any modifications. Without this, the possibility that the reported differences reflect tuning effects rather than method quality cannot be excluded.
minor comments (5)
  1. [III-B.2, Eq. (2)] The notation in Eq. (2) is confusing: F0, F_i^0, and F are introduced without clearly defining the permutation and the pooling operations. Please define F_i^0 explicitly and state the dimensions of the intermediate tensors.
  2. [III-C, Eq. (5)] The quaternion logarithm formula is not written cleanly; the zero-vector case and the scalar component are mixed. Please use the standard expression log(q) = [0, v * arccos(q0)/||v||] when ||v|| > 0 and define the convention consistently with Eq. (6).
  3. [IV-A.1] The total frame count is ambiguous: the paper says 'over 30k frames' for NEPose, but then reports only 4328 training/validation frames and 1308 test frames. Clarify how many frames were used in total and how the remaining frames were discarded or used.
  4. [Tables II and III] The table legends are inconsistent: Table II uses symbols S, C, O without a full expansion in the caption, and Table III introduces 'MUD' in the first row without defining it in the caption. Please define all symbols and abbreviations directly beneath each table.
  5. [Table IV and IV-F] The inference speed test uses only 18 samples total, with 6 per dataset, and the FPS values vary widely (for example, 22.89 to 40.83 fps on NEPose). Reporting the mean FPS over more runs, with standard deviation, would be more informative than the current six numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pose predictor is supervised by ground-truth pose labels and evaluated on held-out frames; the optical-flow pseudo-label finetuning is auxiliary and does not define the pose output.

full rationale

I find no circularity in the derivation chain. The relative pose output is produced by a learned network whose loss (Eq. 6) compares the predicted translation and log-quaternion against ground-truth relative poses computed from Pi and Pi-k; the predicted quantity is therefore not equal to the training target by construction, and Eq. 7 is only a composition of relative transformations to obtain absolute poses. The optical-flow branch is finetuned on TVL1 pseudo-labels (Eq. 1), but this supervises an auxiliary input modality, not the pose prediction, and the final pose remains supervised by pose ground truth. No load-bearing uniqueness theorem or self-citation chain is invoked; references to FastFlowNet, MambaOut, and the geometric loss are standard external components. One experimental-validity concern is not circularity: Section IV-A.1 says NEPose has 50 videos from 16 subjects but then says '4328 frames from 30 different subjects are randomly chosen as the training set and the validation set,' while Section IV-D says 1308 frames from 12 subjects are used for comparison; the text never establishes a subject-disjoint test split. If test subjects overlap training subjects, the Table I margin on NEPose could be inflated, but that would be data leakage or a reporting error, not a definitional equivalence between the method's inputs and outputs. The central claim is an empirical comparison, not a derivation that reduces to its own assumptions.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method's central claim depends on standard supervised learning assumptions: accurate labels, domain transfer of pretrained features, and a fair, disjoint evaluation protocol. The most fragile entries are the NEPose split consistency and the use of TVL1 optical flow as pseudo-ground-truth.

free parameters (4)
  • Loss balance weights lambda1, lambda2 = initially 0 and -3, learned
    Equation (6) in Section III-C uses learnable weights to balance translation and rotation terms; the final pose predictions depend on these fitted values.
  • Attention pooling weights alpha, beta = initialized randomly in [0,1), learned
    Equation (2) in Section III-B.2 adaptively combines max and average pooling; these are trained with the network and influence the joint feature map.
  • Residual scale gamma = initialized 10^-6, learned
    Section III-B.3 scales the output of each depthwise separable block with a learnable gamma before the residual connection.
  • Frame sampling interval k = 4
    Section IV-B sets k=4 so that relative poses are estimated between frames four apart; this hand-chosen value sets the difficulty and motion scale of the task.
assumptions (5)
  • domain assumption Dual TVL1 optical flow is accurate enough to serve as ground truth for finetuning FastFlowNet.
    Section III-B.1 uses TVL1 results as pseudo-labels for optical flow; noisy flow would propagate into the motion feature f_m.
  • domain assumption ImageNet-pretrained MambaOut features transfer to endoscopic images.
    Section III-B.2 initializes the scene and motion feature extractor with ImageNet-1k pretrained weights; if the domain gap is large, the extracted features may be uninformative.
  • domain assumption The three datasets' ground-truth poses are accurate enough for supervised training and evaluation.
    Section IV-A relies on an optical tracking system (NEPose), a simulator (SimCol), and a robot arm (EndoSLAM); any systematic error in these labels directly limits the reported ATE, CE, and DE values.
  • domain assumption Training and test splits are disjoint and representative.
    Section IV-A.1 and IV-D describe random selection without seeds; the NEPose subject count inconsistency (16 vs 30) makes disjointness uncertain.
  • domain assumption The geometric loss from Brahmbhatt et al. [28] is a suitable objective for endoscopic pose regression.
    Section III-C adopts this published loss without re-derivation; its weighting of translation and rotation errors is assumed appropriate for the target domain.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REMOTE: Real-time Ego-motion Tracking for Various Endoscopes via Multimodal Visual Feature Learning." pith.science (2026). https://pith.science/paper/V36ZX3D2

@misc{pith2026250118124,
  author       = {Pith},
  title        = {Pith review of: REMOTE: Real-time Ego-motion Tracking for Various Endoscopes via Multimodal Visual Feature Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V36ZX3D2}},
  note         = {Machine review of arXiv:2501.18124}
}
read the original abstract

Real-time ego-motion tracking for endoscope is a significant task for efficient navigation and robotic automation of endoscopy. In this paper, a novel framework is proposed to perform real-time ego-motion tracking for endoscope. Firstly, a multi-modal visual feature learning network is proposed to perform relative pose prediction, in which the motion feature from the optical flow, the scene features and the joint feature from two adjacent observations are all extracted for prediction. Due to more correlation information in the channel dimension of the concatenated image, a novel feature extractor is designed based on an attention mechanism to integrate multi-dimensional information from the concatenation of two continuous frames. To extract more complete feature representation from the fused features, a novel pose decoder is proposed to predict the pose transformation from the concatenated feature map at the end of the framework. At last, the absolute pose of endoscope is calculated based on relative poses. The experiment is conducted on three datasets of various endoscopic scenes and the results demonstrate that the proposed method outperforms state-of-the-art methods. Besides, the inference speed of the proposed method is over 30 frames per second, which meets the real-time requirement. The project page is here: remote-bmxs.netlify.app

Figures

Figures reproduced from arXiv: 2501.18124 by the authors.

Figure 1
Figure 1. The pipelines of the proposed network and previous methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The pipeline of the proposed framework. FS represents the feature extractor to extract scene features and the motion feature from two adjacent endoscopic images and the corresponding optical flow respectively. FJ represents the feature extractor to extract joint feature from the concatenation of two frames. In the pipeline of FJ , ’Conv1’ and ’Conv2’ represent the first two layers of ResNet-34. the concatenation of … view at source ↗
Figure 3
Figure 3. The trajectories tracking based on different methods. A-E are trajectories from dataset NEPose, F-H are trajectories from dataset SimCol (three [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The visualization of directions predicted by different methods. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 23 canonical work pages

  1. [1]

    Advanced endoscopic navigation: surgical big data, methodology, and applications,

    X. Luo, K. Mori, and T. M. Peters, “Advanced endoscopic navigation: surgical big data, methodology, and applications,” Annual review of biomedical engineering, vol. 20, no. 1, pp. 221–251, 2018

    work page 2018

  2. [2]

    Comparative study of hand–eye calibration methods for augmented reality using an endoscope,

    S. Lee, H. Lee, H. Choi, S. Jeon, H. Ha, and J. Hong, “Comparative study of hand–eye calibration methods for augmented reality using an endoscope,” Journal of Electronic Imaging, vol. 27, no. 4, pp. 043 017– 043 017, 2018

    work page 2018

  3. [3]

    Magnetic tracking of wireless capsule endoscope in mobile setup based on differential signals,

    S. Song, S. Wang, S. Yuan, J. Wang, W. Liu, and M. Q.-H. Meng, “Magnetic tracking of wireless capsule endoscope in mobile setup based on differential signals,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–8, 2021

    work page 2021

  4. [4]

    Location recognition using prioritized feature matching,

    Y . Li, N. Snavely, and D. P. Huttenlocher, “Location recognition using prioritized feature matching,” in Computer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11 . Springer, 2010, pp. 791–804

    work page 2010

  5. [5]

    Fast localization in large-scale environ- ments using supervised indexing of binary features,

    Y . Feng, L. Fan, and Y . Wu, “Fast localization in large-scale environ- ments using supervised indexing of binary features,” IEEE Transac- tions on Image Processing , vol. 25, no. 1, pp. 343–358, 2015

    work page 2015

  6. [6]

    A critical analysis of image-based camera pose estimation techniques,

    M. Xu, Y . Wang, B. Xu, J. Zhang, J. Ren, Z. Huang, S. Poslad, and P. Xu, “A critical analysis of image-based camera pose estimation techniques,” Neurocomputing, vol. 570, p. 127125, 2024

    work page 2024

  7. [7]

    Offsetnet: Deep learning for localization in the lung using rendered images,

    J. Sganga, D. Eng, C. Graetzel, and D. Camarillo, “Offsetnet: Deep learning for localization in the lung using rendered images,” in 2019 international conference on robotics and automation (ICRA) . IEEE, 2019, pp. 5046–5052

    work page 2019

  8. [8]

    Bronchopose: an analysis of data and model configuration for vision-based bronchoscopy pose estimation,

    J. Borrego-Carazo, C. Sanchez, D. Castells-Rufas, J. Carrabina, and D. Gil, “Bronchopose: an analysis of data and model configuration for vision-based bronchoscopy pose estimation,” Computer Methods and Programs in Biomedicine , vol. 228, p. 107241, 2023

    work page 2023

Show all 28 references
  1. [9]

    Endoslam dataset and an unsupervised monocular visual odometry and depth estimation approach for endoscopic videos,

    K. B. Ozyoruk, G. I. Gokceler, T. L. Bobrow, G. Coskun, K. Incetan, Y . Almalioglu, F. Mahmood, E. Curto, L. Perdigoto, M. Oliveira,et al., “Endoslam dataset and an unsupervised monocular visual odometry and depth estimation approach for endoscopic videos,” Medical image analy...

    2021

  2. [10]

    Self-supervised monocular depth estimation for gastrointestinal endoscopy,

    Y . Liu and S. Zuo, “Self-supervised monocular depth estimation for gastrointestinal endoscopy,” Computer Methods and Programs in Biomedicine, p. 107619, 2023

    2023

  3. [11]

    Self-supervised lightweight depth estimation in endoscopy combining cnn and trans- former,

    Z. Yang, J. Pan, J. Dai, Z. Sun, and Y . Xiao, “Self-supervised lightweight depth estimation in endoscopy combining cnn and trans- former,” IEEE Transactions on Medical Imaging , 2024

    2024

  4. [12]

    Image-based localization using lstms for structured fea- ture correlation,

    F. Walch, C. Hazirbas, L. Leal-Taixe, T. Sattler, S. Hilsenbeck, and D. Cremers, “Image-based localization using lstms for structured fea- ture correlation,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 627–637

    2017

  5. [13]

    Vidloc: A deep spatio-temporal model for 6-dof video-clip relocalization,

    R. Clark, S. Wang, A. Markham, N. Trigoni, and H. Wen, “Vidloc: A deep spatio-temporal model for 6-dof video-clip relocalization,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 6856–6864

    2017

  6. [14]

    Unsupervised-learning-based continuous depth and motion estimation with monocular endoscopy for virtual reality minimally invasive surgery,

    L. Li, X. Li, S. Yang, S. Ding, A. Jolfaei, and X. Zheng, “Unsupervised-learning-based continuous depth and motion estimation with monocular endoscopy for virtual reality minimally invasive surgery,” IEEE Transactions on Industrial Informatics , vol. 17, no. 6, pp. 3920–3928, 2020

    2020

  7. [15]

    Bimodal cam- era pose prediction for endoscopy,

    A. Rau, B. Bhattarai, L. Agapito, and D. Stoyanov, “Bimodal cam- era pose prediction for endoscopy,” IEEE Transactions on Medical Robotics and Bionics , 2023

    2023

  8. [16]

    Self-supervised learning for monocular depth estimation on mini- mally invasive surgery scenes,

    S. Shao, Z. Pei, W. Chen, B. Zhang, X. Wu, D. Sun, and D. Doermann, “Self-supervised learning for monocular depth estimation on mini- mally invasive surgery scenes,” in2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 7159–7165

    2021

  9. [17]

    Posenet: A convolutional network for real-time 6-dof camera relocalization,

    A. Kendall, M. Grimes, and R. Cipolla, “Posenet: A convolutional network for real-time 6-dof camera relocalization,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 2938–2946

    2015

  10. [18]

    Deep regression for monocular camera- based 6-dof global localization in outdoor environments,

    T. Naseer and W. Burgard, “Deep regression for monocular camera- based 6-dof global localization in outdoor environments,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 1525–1530

    2017

  11. [19]

    Atloc: Attention guided camera localization,

    B. Wang, C. Chen, C. X. Lu, P. Zhao, N. Trigoni, and A. Markham, “Atloc: Attention guided camera localization,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 34, no. 06, 2020, pp. 10 393–10 401

    2020

  12. [20]

    Adversarial networks for camera pose regression and refinement,

    M. Bui, C. Baur, N. Navab, S. Ilic, and S. Albarqouni, “Adversarial networks for camera pose regression and refinement,” in Proceedings of the IEEE/CVF International Conference on Computer Vision Work- shops, 2019, pp. 0–0

    2019

  13. [21]

    Vmloc: Variational fusion for learning-based multi- modal camera localization,

    K. Zhou, C. Chen, B. Wang, M. R. U. Saputra, N. Trigoni, and A. Markham, “Vmloc: Variational fusion for learning-based multi- modal camera localization,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 7, 2021, pp. 6165–6173

    2021

  14. [22]

    Learning multi-scene absolute pose regression with transformers,

    Y . Shavit, R. Ferens, and Y . Keller, “Learning multi-scene absolute pose regression with transformers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 2733–2742

    2021

  15. [23]

    Digging into self-supervised monocular depth estimation,

    C. Godard, O. Mac Aodha, M. Firman, and G. J. Brostow, “Digging into self-supervised monocular depth estimation,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 3828–3838

    2019

  16. [24]

    Fastflownet: A lightweight network for fast optical flow estimation,

    L. Kong, C. Shen, and J. Yang, “Fastflownet: A lightweight network for fast optical flow estimation,” in 2021 IEEE International Confer- ence on Robotics and Automation (ICRA) . IEEE, 2021, pp. 10 310– 10 316

    2021

  17. [25]

    Mambaout: Do we really need mamba for vision?

    W. Yu and X. Wang, “Mambaout: Do we really need mamba for vision?” arXiv preprint arXiv:2405.07992 , 2024

    2024 arXiv

  18. [26]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition . Ieee, 2009, pp. 248–255

    2009

  19. [27]

    Landmark based bronchoscope localization for needle insertion under respiratory deformation,

    I. Fried, J. Hoelscher, J. A. Akulian, S. Pizer, and R. Alterovitz, “Landmark based bronchoscope localization for needle insertion under respiratory deformation,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 6593– 6600

    2023

  20. [28]

    Geometry- aware learning of maps for camera localization,

    S. Brahmbhatt, J. Gu, K. Kim, J. Hays, and J. Kautz, “Geometry- aware learning of maps for camera localization,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2018, pp. 2616–2625

    2018

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.