Pith. sign in

REVIEW 4 major objections 5 minor 31 references

A multimodal fusion network that classifies point trajectories achieves state-of-the-art dynamic object segmentation and cleaner static reconstructions.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A multimodal trajectory-classification network plus a point-query SAM refinement step yields better dynamic masks and static reconstructions than DAS3R-style baselines on DAVIS.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection A credible multimodal fusion recipe for dynamic-object segmentation whose main flaw is an overbroad SOTA claim contradicted by its own PointOdyssey results. the 4 major comments →

arxiv 2607.18153 v1 pith:Q44KAUX5 submitted 2026-07-20 cs.CV

Robust Multimodal Dynamic Object Segmentation

Zhe Xin , Hanzhi Chang , Penghui Huang , Yinian Mao , Guoquan Huang This is my paper
classification cs.CV
keywords dynamic object segmentationfeature trajectoriesmultimodal fusionpoint tracking3D reconstructionSAM refinementstatic scene reconstructionmotion segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that dynamic object segmentation is best solved by classifying feature trajectories that fuse three complementary cues: 2D point tracks, 3D reconstruction outputs (depth, camera parameters, attention maps), and SAM semantic features. A Transformer combined with feature-clustering aggregation networks classifies each trajectory as static or dynamic, and a point-query-based iterative SAM refinement cleans the resulting mask. The paper reports state-of-the-art accuracy and recall on DAVIS2017, PointOdyssey, and Sintel, and shows that the improved masks translate into higher PSNR/SSIM for static scene reconstruction with 3D Gaussian Splatting. If true, this offers a practical recipe: a lightweight 2.8M-parameter model on top of off-the-shelf trackers and reconstruction backbones.

Core claim

The central claim is that no single cue is sufficient: 2D tracks capture motion but lack geometry, 3D reconstruction adds geometry but is error-prone for moving objects, and semantics help boundaries but are not motion-aware. The paper's discovery is that converting all cues into feature trajectories and classifying them with a Transformer plus aggregation network lets the model adaptively trust whichever cue is most reliable for each scene, while the point-query SAM refinement handles multiple dynamic objects within one mask. This yields the best-reported dynamic masks on four benchmarks, which in turn improve static scene reconstruction using 3DGS.

What carries the argument

The core object is the feature trajectory: for each tracked 2D point, it concatenates pixel position, per-frame displacement, depth, depth offset, interpolated semantic feature, and attention value, plus camera intrinsics/extrinsics and their relative transforms. The aggregation network performs differentiable clustering/unpooling to give each trajectory global context, and a two-layer Transformer encoder integrates interactions across trajectories. The classification is a sigmoid over the fused features. The SAM refinement (Algorithm 1) then iteratively samples individual dynamic pixels, queries SAM, and accepts the resulting segment if it overlaps the coarse mask by at least β and has at l

Load-bearing premise

The pipeline samples a fixed 200×200 grid of points from the first frame; any dynamic object or thin part that falls entirely between grid points has no trajectory and cannot be classified or refined.

What would settle it

Use a DAVIS video containing a small dynamic object (e.g., a ball a few pixels wide) placed in the first frame so that no grid vertex lands on it. The method's recall on that object should drop to near zero, while a dense optical-flow baseline like ParticleSfM segments it. Alternatively, a synthetic clip with a 3×3-pixel moving square between grid points would directly demonstrate the coverage gap.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, dynamic mask accuracy improves enough to raise static-scene reconstruction PSNR by roughly 2 dB over Easi3R on DAVIS2017.
  • The method's efficiency (0.28s inference, 2.8M parameters) makes it viable for real-time or embedded robotics and AR pipelines, in contrast to baselines with hundreds of millions of parameters.
  • The point-query SAM strategy removes the single-object assumption, so multi-object dynamic scenes no longer break post-processing.
  • Because the model uses relative offsets (Δx, Δd, camera motion) rather than absolute coordinates, it generalizes across datasets with different scales and camera intrinsics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The fixed 200×200 grid of sampling points is a coverage bottleneck: any dynamic object smaller than the grid spacing (or positioned between vertices) will have no trajectory and will be invisible to classification and refinement. Adaptive or learned point sampling could fix this.
  • The method inherits failure modes from MonST3R: the attention and depth features degrade on textureless or highly dynamic scenes (as seen in lower Sintel recall), so a more robust geometry backbone or uncertainty weighting would likely help.
  • The SAM refinement thresholds β and γ are dataset-specific heuristics; a learned decision rule for accepting SAM proposals could make the loop more stable across scenes.
  • The reconstruction claim depends on the 3DGS pipeline of DAS3R; the segmentation contribution could be isolated by evaluating masks directly with other downstream tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a multimodal dynamic object segmentation framework that fuses 2D point tracks from CoTracker, 3D reconstruction features and attention maps from MonST3R, and SAM semantic features. The features are assembled into per-trajectory representations, classified by a Transformer plus feature-clustering aggregation network, and refined by a point-query-based iterative SAM post-processing step. The predicted dynamic masks are then used for staticness-aware 3D Gaussian Splatting reconstruction. Experiments are reported on PointOdyssey, DAVIS 2017, and Sintel against P-SfM, MonST3R, Easi3R, and DAS3R, and the paper claims state-of-the-art performance in both segmentation and reconstruction.

Significance. If the results are reproducible, the core idea—a lightweight 2.8M-parameter fusion network that combines geometric and semantic trajectory features plus a point-query SAM refinement—is practically useful. The DAVIS segmentation results and the static-scene reconstruction improvements are strong, and the ablation in Table V supports the inclusion of the camera, attention, aggregation, and SAM components. However, the unqualified state-of-the-art claim is contradicted by the paper's own PointOdyssey results, and the efficiency claim is not supported by the full-sequence timing. These issues are correctable but need to be addressed before publication.

major comments (4)
  1. [Abstract / Table I] The abstract claims state-of-the-art performance in both dynamic object segmentation and static scene reconstruction. On PointOdyssey, however, Ours(+SAM) obtains IoU 64.38 and precision 73.79, while DAS3R obtains IoU 66.54 and precision 76.62 (Table I). Since PointOdyssey is the training dataset, this is not a cross-domain gap. The SOTA claim should be replaced with dataset-specific statements or a clearly defined ranking, and the discussion in Section V-A.1 should not describe the result as 'superior' based only on accuracy and recall when the primary IoU metric is lower.
  2. [Section V-D / Table VI] The efficiency claim 'shorter inference time' is not supported by Table VI. The network-only time is 0.2833s, but the full pipeline including SAM refinement is 32.44s, which is longer than DAS3R (27.55s) and P-SfM (14.07s). The paper should report the end-to-end inference time for the final method and should not compare the network-only time against baselines that include their post-processing.
  3. [Section III-B / Eq. (6)] The stated goal is to estimate per-frame masks {M^i}, but the model outputs a single label m^k per trajectory (Eq. 5), and training labels are sampled from the ground-truth mask of the first frame only (Eq. 6). This means a trajectory is forced to be classified as either static or dynamic for the entire clip. The framework therefore cannot represent an object that is static at the first frame but becomes dynamic later, or vice versa. Please clarify whether the task is clip-level object labeling or per-frame motion-state segmentation; if the latter, the loss in Eq. (6) is mismatched with the evaluation protocol.
  4. [Section III-A / Algorithm 1] The method samples a fixed 200×200 grid of points from the first frame as CoTracker queries. Any dynamic object or thin structure that does not intersect the grid is invisible to the classifier, and the SAM refinement cannot recover it because queries are drawn from the coarse mask. The paper should discuss this limitation and provide evidence (e.g., performance on small-object subsets of DAVIS) that the grid density is sufficient for the claimed generality.
minor comments (5)
  1. [Section V-C / Table V] The ablation study reports only accuracy and recall, while the comparison tables use IoU and precision as well. The ablation would be more informative if the same metrics were used, since the SAM refinement primarily improves recall and its effect on IoU cannot be assessed from Table V.
  2. [General] No error bars or multiple-seed results are reported. Some differences are small (e.g., Sintel IoU 54.66 vs 53.36 in Table III), so the significance of the improvements should be stated cautiously.
  3. [Algorithm 1] The notation Mcoarse vs M_coarse and SumAll(M_x) is informal; please define the size and overlap operations clearly.
  4. [Table VI] The row labels 'Ours' and 'Ours (Full Sequence)' are ambiguous. Clarify that 'Ours' is the network without SAM post-processing and 'Ours (Full Sequence)' is the complete method.
  5. [References] Reference [26] is formatted as 'in Proc. arXiv:2410.11831'; this should be corrected to a standard citation.

Circularity Check

0 steps flagged

No significant circularity: the approach is a supervised fusion of independently pretrained external features, with no fitted parameter relabeled as a prediction and no load-bearing self-citation chain.

full rationale

The paper's derivation chain is a standard supervised learning pipeline, not a circular one. Inputs are (1) 2D tracks from CoTracker, (2) depth/camera/attention features from MonST3R, and (3) SAMv2 semantic features. The classifier Φ_sem/Φ maps F_traj and F_sem to per-trajectory dynamic scores via Eq. (5), and is trained by cross-entropy against ground-truth masks in Eq. (6). There is no equation in which the output is defined in terms of a fitted parameter that is later reported as a prediction; the trained model is evaluated on held-out DAVIS/Sintel data and on the PointOdyssey test split. The cited external methods (MonST3R, Easi3R, DAS3R, CoTracker, SAM) are used as feature extractors or baselines, not as self-citations, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The SAM refinement in Algorithm 1 deliberately takes the union of accepted SAM masks with the coarse mask, so it is recall-preserving by construction; this is an explicit algorithmic design choice, not a hidden equivalence that makes the learning result circular. The disagreement between the abstract's 'state-of-the-art' claim and Table I (DAS3R has higher IoU/precision on PointOdyssey) is a correctness/claim-consistency issue, not a circularity issue. The grid-sampling coverage limitation is an architectural assumption, not a circular step. Overall, no specific reduction of a claimed prediction to its own inputs can be exhibited from the paper's equations or citations.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. The central claim rests on upstream pretrained models (MonST3R, CoTracker, SAM), a set of manually chosen thresholds, and the transferability of synthetic training. The most consequential free parameter is the 200x200 point grid, which bounds recall by construction.

free parameters (5)
  • Point grid density 200x200
    Chosen in Section III-A; determines the minimum detectable object size and track coverage.
  • Clip length N = 10
    Set in Implementation Details; no sensitivity analysis is provided.
  • SAM overlap threshold beta = 0.3
    Hyperparameter in Algorithm 1; no sensitivity analysis is provided.
  • SAM minimum size gamma = 5
    Hyperparameter in Algorithm 1; no sensitivity analysis is provided.
  • alpha = 0.7
    Listed in Implementation Details but never appears in Algorithm 1 or the loss; its role is unclear.
axioms (5)
  • domain assumption MonST3R depth, pose, and attention estimates are accurate enough to support dynamic mask prediction.
    The whole geometric feature stream is built from MonST3R outputs (Section III-A); if these are poor in a scene, the fusion network has no independent geometric signal.
  • domain assumption CoTracker tracks sampled points reliably across 10 frames in dynamic scenes.
    The 2D point trajectories are the backbone of the representation and the tracked points are used to sample ground-truth labels during training (Section III-B).
  • domain assumption SAMv2 semantic features align with object boundaries and are stable for static/dynamic classification.
    Semantic features are interpolated from SAM feature maps and are used both in the trajectory and in the final refinement; if SAM misses an object, it is unlikely to appear in the mask.
  • domain assumption MonST3R attention maps carry dynamic-object signal, as in Easi3R.
    Attention maps are aggregated across layers and pairs (Section III-A) and fed as input; this assumes the signal used by Easi3R transfers to the new task.
  • domain assumption Training on PointOdyssey synthetic data transfers to DAVIS and Sintel.
    The model is trained only on PointOdyssey and evaluated on DAVIS and Sintel; generalization is assumed rather than explicitly studied.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Multimodal Dynamic Object Segmentation." pith.science (2026). https://pith.science/paper/Q44KAUX5

@misc{pith2026260718153,
  author       = {Pith},
  title        = {Pith review of: Robust Multimodal Dynamic Object Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q44KAUX5}},
  note         = {Machine review of arXiv:2607.18153}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Dynamic object segmentation plays a critical role in many visual applications such as static scene reconstruction from dynamic videos. However, existing optical flow-based methods fail to ensure consistent static/dynamic segmentation along object boundaries, while 3D reconstruction-based approaches are highly sensitive to reconstruction errors. To address these limitations, we present a dynamic object segmentation framework that can generate both precise and complete dynamic masks by integrating multimodal cues including 2D point tracks, 3D reconstruction, and semantic information. We design a network combining Transformer architectures with feature clustering aggregation modules to perform static/dynamic classification of multimodal feature trajectories. It enables the model to adaptively determine which type of feature should dominate based on the characteristics of each scene, while also mitigating the impact of feature degradation. Additionally, we introduce a novel point-query-based SAM post-processing method capable of handling multiple objects within a single mask. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in both dynamic object segmentation and static scene reconstruction tasks.

Figures

Figures reproduced from arXiv: 2607.18153 by Guoquan Huang, Hanzhi Chang, Penghui Huang, Yinian Mao, Zhe Xin.

Figure 1
Figure 1. Figure 1: Overview. Given a dynamic video containing several image frames, Our method leverages 2D point tracking, 3D reconstruction, and semantic priors as inputs, to accomplish the tasks of dynamic object segmentation and static scene reconstruction from dynamic videos. directly reflects pixel-level movements that are intrinsically correlated with object motions. However, their pixel-wise nature fails to ensure co… view at source ↗
Figure 2
Figure 2. Figure 2: Overall Architecture. Given a video clip with N frames, we first employ MonST3R [9] to perform coarse 3D reconstruction of the dynamic scene, obtaining depth information, camera parameters, and attention maps. Simultaneously, we use CoTracker [26] to densely sample 2D points on the images and obtain their motion trajectories across frames. These 2D point trajectories are then associated with corresponding … view at source ↗
Figure 3
Figure 3. Figure 3: Visual Comparison of Dynamic Mask Prediction. We provide quantitative comparisons for dynamic mask prediction tasks. Our method produces more precise and complete masks compared to other approaches. Original RGB GT Mask Ours (+SAM) DAS3R Monst3R [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual Comparison of Static Scene Reconstruction. We provide quantitative comparisons for the tasks of static scene reconstruction from dynamic videos using the DAVIS2017 dataset. Our method achieves more photorealistic novel view synthesis results compared to other methods, demonstrating its effectiveness in dynamic object segmentation. dynamic objects are removed. These factors make scene reconstruction … view at source ↗
Figure 5
Figure 5. Figure 5: Visual Comparison in the Ablation Study. We conduct ablation study on different feature combinations and processing strategies. Our full model achieves the best performance among all variants. RGB Images GT Ours (Point-based) Mask-based [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of Different SAM refinement strategy. We show comparisons between ours and mask-based SAM refinement strategy. Previous mask-based refinement strategy suffers from incomplete mask generation due to the assumption that dynamic pixels belong to a single object. background have been precisely and effectively separated. C. Ablation Study TABLE V. Motion Mask Accuracy of DAVIS2017 dataset with differ… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 5 linked inside Pith

  1. [1]

    Motion segmentation and appearance change detection based 2d hand tracking,

    J. H. Hammer, M. V oit, and J. Beyerer, “Motion segmentation and appearance change detection based 2d hand tracking,” in2016 19th International Conference on Information Fusion (FUSION). IEEE, 2016, pp. 1743–1750

    2016

  2. [2]

    Moving object segmentation using optical flow and depth information,

    J. Klappstein, T. Vaudrey, C. Rabe, A. Wedel, and R. Klette, “Moving object segmentation using optical flow and depth information,” in Pacific-Rim symposium on image and video technology. Springer, 2009, pp. 611–623

    2009

  3. [3]

    Motion and depth augmented semantic segmentation for autonomous navigation,

    H. Rashed, A. El Sallab, S. Yogamani, and M. ElHelw, “Motion and depth augmented semantic segmentation for autonomous navigation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 0–0

    2019

  4. [4]

    Dymslam: 4d dynamic scene reconstruction based on geometrical motion segmentation,

    C. Wang, B. Luo, Y . Zhang, Q. Zhao, L. Yin, W. Wang, X. Su, Y . Wang, and C. Li, “Dymslam: 4d dynamic scene reconstruction based on geometrical motion segmentation,”IEEE Robotics and Automation Letters, vol. 6, no. 2, pp. 550–557, 2020

    2020

  5. [5]

    Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild,

    W. Zhao, S. Liu, H. Guo, W. Wang, and Y .-J. Liu, “Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 523–542

    2022

  6. [6]

    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

    2024

  7. [7]

    Self- supervised video object segmentation by motion grouping,

    C. Yang, H. Lamdouar, E. Lu, A. Zisserman, and W. Xie, “Self- supervised video object segmentation by motion grouping,” inPro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 7177–7188

    2021

  8. [8]

    Matnet: Motion- attentive transition network for zero-shot video object segmentation,

    T. Zhou, J. Li, S. Wang, R. Tao, and J. Shen, “Matnet: Motion- attentive transition network for zero-shot video object segmentation,” IEEE transactions on image processing, vol. 29, pp. 8326–8338, 2020

    2020

  9. [9]

    Monst3r: A simple approach for estimating geometry in the presence of motion,

    J. Zhang, C. Herrmann, J. Hur, V . Jampani, T. Darrell, F. Cole, D. Sun, and M.-H. Yang, “Monst3r: A simple approach for estimating geometry in the presence of motion,”arXiv preprint arXiv:2410.03825, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Easi3r: Estimating disentangled motion from dust3r without training,

    X. Chen, Y . Chen, Y . Xiu, A. Geiger, and A. Chen, “Easi3r: Estimating disentangled motion from dust3r without training,”arXiv preprint arXiv:2503.24391, 2025

    arXiv 2025

  11. [11]

    Das3r: Dynamics- aware gaussian splatting for static scene reconstruction,

    K. Xu, T. H. E. Tse, J. Peng, and A. Yao, “Das3r: Dynamics- aware gaussian splatting for static scene reconstruction,”arXiv preprint arXiv:2412.19584, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Dynaslam: Tracking, mapping, and inpainting in dynamic scenes,

    B. Bescos, J. M. F ´acil, J. Civera, and J. Neira, “Dynaslam: Tracking, mapping, and inpainting in dynamic scenes,”IEEE robotics and automation letters, vol. 3, no. 4, pp. 4076–4083, 2018

    2018

  13. [13]

    Dynamic-slam: Semantic monocular visual localization and mapping based on deep learning in dynamic environment,

    L. Xiao, J. Wang, X. Qiu, Z. Rong, and X. Zou, “Dynamic-slam: Semantic monocular visual localization and mapping based on deep learning in dynamic environment,”Robotics and Autonomous Systems, vol. 117, pp. 1–16, 2019

    2019

  14. [14]

    Robust 3d gaussian splatting for novel view synthesis in presence of distractors,

    P. Ungermann, A. Ettenhofer, M. Nießner, and B. Roessle, “Robust 3d gaussian splatting for novel view synthesis in presence of distractors,” inDAGM German Conference on Pattern Recognition. Springer, 2024, pp. 153–167

    2024

  15. [15]

    Bootstrapping objectness from videos by relaxed common fate and visual grouping,

    L. Lian, Z. Wu, and S. X. Yu, “Bootstrapping objectness from videos by relaxed common fate and visual grouping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 582–14 591

    2023

  16. [16]

    Segmenting moving objects via an object-centric layered representation,

    J. Xie, W. Xie, and A. Zisserman, “Segmenting moving objects via an object-centric layered representation,”Advances in neural information processing systems, vol. 35, pp. 28 023–28 036, 2022

    2022

  17. [17]

    Learning segmentation from point trajectories,

    L. Karazija, I. Laina, C. Rupprecht, and A. Vedaldi, “Learning segmentation from point trajectories,”Advances in Neural Information Processing Systems, vol. 37, pp. 112 573–112 597, 2024

    2024

  18. [18]

    Object segmentation by long term analysis of point trajectories,

    T. Brox and J. Malik, “Object segmentation by long term analysis of point trajectories,” inEuropean conference on computer vision. Springer, 2010, pp. 282–295

    2010

  19. [19]

    Structure and motion from casual videos,

    Z. Zhang, F. Cole, Z. Li, M. Rubinstein, N. Snavely, and W. T. Freeman, “Structure and motion from casual videos,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 20–37

    2022

  20. [20]

    Romo: Robust motion segmentation improves structure from motion,

    L. Goli, S. Sabour, M. Matthews, M. Brubaker, D. Lagun, A. Ja- cobson, D. J. Fleet, S. Saxena, and A. Tagliasacchi, “Romo: Robust motion segmentation improves structure from motion,”arXiv preprint arXiv:2411.18650, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Fitting conic sections to “very scattered

    P. D. Sampson, “Fitting conic sections to “very scattered” data: An iterative refinement of the bookstein algorithm,”Computer graphics and image processing, vol. 18, no. 1, pp. 97–108, 1982

    1982

  22. [22]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,”arXiv preprint arXiv:2408.00714, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 697–20 709

    2024

  24. [24]

    Sea-raft: Simple, efficient, accurate raft for optical flow,

    Y . Wang, L. Lipson, and J. Deng, “Sea-raft: Simple, efficient, accurate raft for optical flow,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 36–54

    2024

  25. [25]

    Vision transformers for dense prediction,

    R. Ranftl, A. Bochkovskiy, and V . Koltun, “Vision transformers for dense prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 179–12 188

    2021

  26. [26]

    Cotracker3: Simpler and better point tracking by pseudo-labelling real videos,

    N. Karaev, I. Makarov, J. Wang, N. Neverova, A. Vedaldi, and C. Rupprecht, “Cotracker3: Simpler and better point tracking by pseudo-labelling real videos,” inProc. arXiv:2410.11831, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Spars3r: Semantic prior align- ment and regularization for sparse 3d reconstruction,

    Y . Tang, Y . Guo, D. Li, and C. Peng, “Spars3r: Semantic prior align- ment and regularization for sparse 3d reconstruction,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 26 810–26 821

    2025

  28. [28]

    Pointodyssey: A large-scale synthetic dataset for long-term point tracking,

    Y . Zheng, A. W. Harley, B. Shen, G. Wetzstein, and L. J. Guibas, “Pointodyssey: A large-scale synthetic dataset for long-term point tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 855–19 865

    2023

  29. [29]

    A benchmark dataset and evaluation method- ology for video object segmentation,

    F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung, “A benchmark dataset and evaluation method- ology for video object segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 724– 732

    2016

  30. [30]

    A naturalistic open source movie for optical flow evaluation,

    D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black, “A naturalistic open source movie for optical flow evaluation,” inEuropean confer- ence on computer vision. Springer, 2012, pp. 611–625

    2012

  31. [31]

    Learning two-view correspondences and geometry using order-aware network,

    J. Zhang, D. Sun, Z. Luo, A. Yao, L. Zhou, T. Shen, Y . Chen, L. Quan, and H. Liao, “Learning two-view correspondences and geometry using order-aware network,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5845–5854

    2019

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.