Pith. sign in

REVIEW 3 major objections 4 minor 40 references

MS-RAFT-3D: A Multi-Scale Architecture for Recurrent Image-Based Scene Flow

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Multi-scale recurrent refinement of SE(3) motion fields sets new state-of-the-art results for image-based scene flow on both KITTI and Spring benchmarks.

desk verdict Solid multi-scale scene flow architecture with honest ablations; the Spring 'SOTA' headline overstates a 5.9% gain over the authors' own baseline, but the core contribution is real. read the letter →

arxiv 2506.01443 v1 pith:CHOLGUJI submitted 2025-06-02 cs.CV

Jakob Schmid , Azin Jahedi , Noah Berenguel Senn , Andrés Bruhn This is my paper
classification cs.CV
keywords image-basedsceneflowrecurrentnetworkmulti-scalearchitecturecoarse-to-fineestimationSE(3)motionfieldrigidembeddingsKITTIbenchmarkSpring
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 argues that the single-scale recurrent architecture RAFT-3D can be upgraded into a multi-scale, coarse-to-fine estimator that produces state-of-the-art image-based scene flow. On KITTI, its best model reduces scene flow outliers from 4.26 to 3.89, an 8.7% relative improvement, and on Spring it cuts the endpoint error from 26.71 to 9.13, a 65.8% improvement. These gains come from a U-Net feature encoder, a lean top-down context encoder that replaces the heavier ResNet50, upsampling the SE(3) field and motion embeddings between scales, and a multi-scale training loss. If correct, this shows that multi-scale ideas that already helped optical flow and stereo also benefit the harder task of estimating 3D motion.

What carries the argument

The central mechanism is the coarse-to-fine SE(3) field initialization: at each scale, the rigid-motion embeddings and the SE(3) transformation field from the coarser scale are upsampled with a shared learned convex mask, with the transformation upsampled in the Lie algebra, and used to initialize the next finer scale's recurrent update. Around this, the paper places a U-Net feature encoder with multi-scale consolidation, a lean top-down context encoder whose per-scale outputs are unified by 1x1 convolutions, a two-level correlation pyramid with on-demand cost sampling, and the Dense-SE3 layer that re-estimates each pixel's transformation from neighboring pixels with similar rigid-motion embeddings and high confidence. The rigid-motion embeddings are smoothed by a bi-Laplacian term whose edge weights are predicted by the network, and ablations report that removing this smoothing raises scene flow error substantially.

What would settle it

Fine-tune the runner-up methods, such as CamLiRAFT or ScaleRAFT, on the Spring training split using the same protocol and submit them to the Spring benchmark; if any achieves a scene flow endpoint error below 9.13, the paper's '65.8% over SOTA' claim is contradicted. Alternatively, run MS-RAFT-3D+ on Spring while replacing CroCo-Stereo disparities with those of a weaker stereo method; if the endpoint error rises above 9.13, the stereo input, not the architecture alone, carries part of the claimed improvement.

Watch

Extended reading notes

Core claim

The central claim is that an image-based scene flow estimator can be organized as a pyramid of recurrent refinements: start with a coarse SE(3) field at 1/16 resolution, then repeatedly upsample the field and its rigid-motion embeddings with a shared convex mask (using Lie-algebra upsampling for the transform) to initialize each finer scale, ending at 1/4 resolution in the 3-scale model and 1/2 resolution in the 4-scale model. Combined with a U-Net feature encoder, a lean top-down context encoder that saves about 17 million parameters relative to RAFT-3D, and a multi-scale multi-iteration loss over optical flow, inverse-depth change, and revision terms, the resulting MS-RAFT-3D and MS-RAFT-3D+ produce sharper motion detail, especially in dynamic foreground regions. On KITTI the best model achieves a scene flow outlier rate of 3.89 versus 4.26 for the previous leading method, and on Spring it achieves 9.13 endpoint error versus 26.71 for the best fine-tuned baseline. The FlyingThings experiment, which uses ground-truth disparity, isolates the architecture's contribution and shows the multi-scale versions beat both RAFT-3D baselines across all reported metrics.

Load-bearing premise

The Spring state-of-the-art claim assumes that PWOC-3D is the strongest available comparison because it is the only published method fine-tuned on Spring; if other current methods were fine-tuned and still beat MS-RAFT-3D+, the 65.8% headline improvement would not hold.

Editorial extensions

If this is right

  • On KITTI, the best variant's 3.89 scene flow outlier rate is the lowest published result, reducing the previous best by 8.7%.
  • On Spring, the 4-scale model cuts the best fine-tuned baseline's endpoint error by 65.8%, demonstrating that high-resolution refinement recovers detail that single-scale 1/8-resolution methods lose.
  • The lean context encoder removes over 17 million parameters relative to RAFT-3D's ResNet50 while improving accuracy, so the multi-scale gains are not bought by a heavier context network.
  • Bi-Laplacian smoothing of rigid-motion embeddings is essential: ablations show scene flow error rises from 6.66 to 11.83 on KITTI (train) when it is disabled.
  • The multi-scale design works on all scene flow components jointly and shares the recurrent unit across scales, so it can be applied directly to other recurrent joint-motion estimators that use the same RAFT-style update block.

Reading between the lines

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

  • An unstated consequence is that the 65.8% Spring improvement may understate or overstate the true advantage over current public methods, because the comparison is restricted to fine-tuned approaches; fine-tuning other published methods on Spring could change that margin substantially.
  • Because the FlyingThings experiment with ground-truth disparity shows large gains from the architecture alone, the multi-scale recurrence rather than the improved stereo input carries most of the improvement; a direct check would be to evaluate MS-RAFT-3D+ on KITTI and Spring using a weaker stereo method.
  • The same coarse-to-fine SE(3) initialization could be transferred to point-cloud scene flow networks, whose multi-scale recurrent variants already exist, to test whether the benefit generalizes across input modalities.
  • A testable extension is to vary the number of scales and the Dense-SE3 neighborhood radius independently to map where the accuracy-versus-runtime tradeoff saturates.
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 / 4 minor

Summary. The paper presents MS-RAFT-3D, a multi-scale recurrent architecture for image-based scene flow estimation. It extends RAFT-3D with a three- or four-level coarse-to-fine scheme in which the SE(3) field and rigid-motion embeddings are upsampled between scales, a U-Net feature encoder, a lean top-down context encoder, and a multi-scale multi-iteration loss. The authors evaluate the method on FlyingThings, KITTI and Spring, reporting consistent improvements over RAFT-3D baselines and claiming state-of-the-art results on KITTI and Spring, with code and benchmark submissions available.

Significance. The architectural contribution is credible and useful. The FlyingThings comparison uses ground-truth disparity, so the gains there are attributable to the multi-scale design rather than to the improved CroCo-Stereo input; the ablations in Tables 4 and 5 separately support the feature encoder, context encoder, scale initialization, smoothing and neighborhood choices. The paper also reduces parameter count by roughly 17M relative to RAFT-3D and releases code, which aids reproducibility. However, the headline SOTA claims are not yet supported with the same rigor: the Spring claim relies on a single 2019 comparator and an unsupported fine-tuning premise, and the KITTI claim is a time-sensitive snapshot. If the claims are appropriately qualified, the paper would be a solid contribution to recurrent scene flow estimation.

major comments (3)
  1. [Section 3.1 (Table 3) and abstract] The Spring SOTA claim of a 65.8% improvement (9.13 vs. 26.71) is not established by the presented evidence. The comparison is made only to PWOC-3D, a 2019 method, and the sentence 'other methods on the benchmark including the original RAFT-3D method, are not fine-tuned on the dataset' is asserted without citing the benchmark's fine-tuning flags or otherwise documenting the status of current leaderboard entries. Moreover, PWOC-3D predates the Spring benchmark, so if it was fine-tuned by the authors for this comparison, the protocol should be described; if it was not, it is not the 'best fine-tuned approach.' Against the paper's own matched baseline RAFT-3D-CroCo, MS-RAFT-3D+ improves by 5.9% on Spring (9.13 vs. 9.70). The abstract and conclusion should therefore be revised to claim the architecture-driven improvement over the matched baseline, and to report the PWOC-3D comparison only with verifiable fine-tuning status.
  2. [Section 3.1 (Table 2) and abstract] The KITTI claim of outperforming 'the current SOTA' is a snapshot based on CamLiRAFT (2023). Benchmark leaderboards are time-varying, and the paper does not state the date of the comparison or define whether the reference is the full leaderboard or only published methods. The 8.7% figure should be qualified as 'best published method at the time of submission' or refreshed against the current leaderboard before publication, in both the abstract and the conclusion.
  3. [Section 3.1 (Table 2) and Section 1] The KITTI comparison includes CamLiFlow and CamLiRAFT, which are camera-LiDAR fusion methods, while the paper frames its contribution as image-based scene flow. If the KITTI benchmark permits LiDAR input, the phrase 'image-based scene flow SOTA' is ambiguous and should be stated as SOTA on the benchmark regardless of input modality; if the comparison is meant to be image-based, the table should be restricted to, or at least clearly separate, methods using the same RGB-D-from-stereo input as MS-RAFT-3D. This distinction directly affects the interpretation of the headline comparison.
minor comments (4)
  1. [Table 1] The notation '13.6x' and similar is unexplained; the caption says converted inlier metrics, but the reader cannot tell what the 'x' denotes or how the conversion was done. Please add units or a footnote.
  2. [Table 5] The header 'Context Encoder Channelsfinest to coarsest scale' is missing a space, and 'Neigborhood' is misspelled; the rows '32 Neighbors' and '256 Pixels' should state whether the comparison is by radius, neighborhood size, or covered image area.
  3. [Section 2, Flow Refinement] The fixed neighborhood radius of 32 across scales means the effective image area shrinks at finer scales; the text says 'the considered area gets smaller on finer scales' but the rationale is given only later in the ablation. A one-sentence explanation here would improve readability.
  4. [Conclusion] The final sentence repeats the unqualified SOTA claim; align it with the revised abstract and with the qualifications suggested in the major comments.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; architecture gains are evaluated against external benchmarks, and internal ablations are self-contained.

full rationale

The paper's central claim is an architectural improvement (multi-scale coarse-to-fine scene flow) assessed on the external KITTI and Spring benchmarks. The improvements over the RAFT-3D-CroCo baseline are reported as raw benchmark numbers, not as quantities derived from fitted parameters. The multi-scale loss in Section 2 ('We generalize the idea of the multi-scale multi-iteration loss from [14] to the scene flow loss of RAFT-3D') borrows from the authors' own prior work, but the generalization is implemented and ablated within this paper, and the final benchmark numbers are externally grounded. Similarly, the U-Net feature encoder is taken from CCMR [15], another self-citation, yet its benefit is verified by an ablation in Table 4 comparing 'Top-down RAFT-like' versus 'CCMR' encoders, so the choice is not forced by citation alone. The Spring SOTA claim relies on an asserted premise in Section 3.1 that PWOC-3D is the best fine-tuned comparator because other methods are not fine-tuned on Spring; this is a correctness/competitiveness concern, not circularity, since PWOC-3D is an external 2019 method. No equation in the paper is defined in terms of the quantities it claims to predict, and no fitted value is renamed as a prediction. Overall, the derivation chain is self-contained: the architectural components are independently validated by ablation, and the headline results are benchmark comparisons.

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

No new particles, forces, or physical entities are introduced. The approach inherits the rigid-motion and depth-availability assumptions from RAFT-3D, and its quantitative results depend on a set of hand-chosen hyperparameters (loss weights, neighborhood radii, iteration counts). The benchmark claims are therefore conditional on these standard domain assumptions.

free parameters (5)
  • wd (inverse depth change loss weight) = 250
    Weight balancing the inverse depth change term in the multi-scale loss (Section 2, Supervision). Chosen by hand; affects the relative emphasis on depth vs. flow.
  • wrev (flow revision loss weight) = 0.2
    Weight for the flow revision term in the loss (Section 2, Supervision). Hand-chosen.
  • gamma (loss decay factor) = 0.8
    Exponential decay factor for the multi-scale multi-iteration loss (Section 2). Hand-chosen.
  • Dense-SE3 neighborhood radius = 32 (3-scale), 64 (4-scale)
    Neighborhood radius in the Dense-SE3 layer. Selected via ablations (Tables 5 and Section 3).
  • Iteration schedule = [4,6,8] and [4,5,5,6]
    Number of recurrent update iterations per scale for 3-scale and 4-scale models, chosen following prior work [14].
assumptions (3)
  • domain assumption Depth/disparity is available or can be accurately estimated (GT on FlyingThings, CroCo-Stereo on KITTI/Spring).
    The method consumes RGB-D input; on stereo benchmarks it inherits disparities from CroCo-Stereo (Section 2, first paragraph).
  • domain assumption The scene can be represented as a set of regions undergoing rigid SE(3) motion, justifying the Dense-SE3 parametrization and bi-Laplacian smoothing of motion embeddings.
    Inherited from RAFT-3D [9]; used in the refinement equations in Section 2.
  • standard math Upsampling the SE(3) field in Lie algebra preserves the group structure and is a valid coarse-to-fine transfer.
    Used in the scale initialization (Section 2, Method); standard result in Lie group theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MS-RAFT-3D: A Multi-Scale Architecture for Recurrent Image-Based Scene Flow." pith.science (2026). https://pith.science/paper/CHOLGUJI

@misc{pith2026250601443,
  author       = {Pith},
  title        = {Pith review of: MS-RAFT-3D: A Multi-Scale Architecture for Recurrent Image-Based Scene Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CHOLGUJI}},
  note         = {Machine review of arXiv:2506.01443}
}
read the original abstract

Although multi-scale concepts have recently proven useful for recurrent network architectures in the field of optical flow and stereo, they have not been considered for image-based scene flow so far. Hence, based on a single-scale recurrent scene flow backbone, we develop a multi-scale approach that generalizes successful hierarchical ideas from optical flow to image-based scene flow. By considering suitable concepts for the feature and the context encoder, the overall coarse-to-fine framework and the training loss, we succeed to design a scene flow approach that outperforms the current state of the art on KITTI and Spring by 8.7%(3.89 vs. 4.26) and 65.8% (9.13 vs. 26.71), respectively. Our code is available at https://github.com/cv-stuttgart/MS-RAFT-3D.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 39 canonical work pages

  1. [1]

    INTRODUCTION With applications such as autonomous driving [1], scene un- derstanding [2] and robot-assisted surgery [3] the estimation of 3D motion from RGB-D or stereo image sequences plays an important role in computer vision. The corresponding mo- tion field, which is also referred to as scene flow [4], provides valuable information on the interaction ...

    work page Pith review arXiv 2025

  2. [2]

    In case of stereo images, i.e

    METHOD Given a pair of RGB-D images our approach predicts a dense field of SE(3) transformations. In case of stereo images, i.e. when the depth is not available, it is estimated with a stereo method as in [9, 10, 11, 21]. In our approach we consider CroCo-Stereo [22] – a recent highly accurate stereo method. Conceptually, our approach is based on RAFT-3D ...

    work page

  3. [3]

    After multiple refinement iterations on the coarsest scale, the SE(3) motion field and the rigid motion embeddings are UpdateNiter(1) MatchingCosts Dense-SE3 SE(3)-Fld Mo-Emb Features 1Features 2 Context Depth 1Depth 2 ⇑ ×2 SE(3)-Fld Mo-Emb ↑ ×8 SE(3)-FldLoss Initialization 116×(h, w) UpdateNiter(2) MatchingCosts Dense-SE3 SE(3)-Fld Mo-Emb Features 1Featu...

    work page

  4. [4]

    As in [9], upsampling of the SE(3) field is done in the Lie algebra

    Because the edges of the motion field and the rigid motion embeddings are expected to coincide, the upsampling mask is shared between both. As in [9], upsampling of the SE(3) field is done in the Lie algebra. Note that while we emphasize here on the 3-scale model, we also present a 4-scale variant (finest scale at 1

    work page

  5. [5]

    Encoders The matching process starts by extracting image features from both frames and context features from the first frame

    in Section 3. Encoders The matching process starts by extracting image features from both frames and context features from the first frame. For computing the image features we use a Siamese U- Net-based encoder as in [15]. While the feature consolidation in this type of encoder is beneficial for image features, based on our ablations, it is not helpful fo...

    work page

  6. [6]

    For the latter, we increase the Dense-SE3 neighbor- hood radius to 64 due to the higher resolution

    RESULTS AND EXPERIMENTS Models We investigate a 3-scale model (MS-RAFT-3D), op- erating on 1 16, 1 8 and 1 4 of the original image resolution and a 4-scale variant (MS-RAFT-3D+) that adds a scale with res- olution 1 2 (see architecture in supplementary material : Sec- tion 6). For the latter, we increase the Dense-SE3 neighbor- hood radius to 64 due to th...

    work page

  7. [7]

    PWOC-3D: Deep occlusion-aware end-to- end scene flow estimation,

    R. Saxena, R. Schuster, O. Wasenmuller, and D. Stricker, “PWOC-3D: Deep occlusion-aware end-to- end scene flow estimation,” in IEEE IV, 2019, pp. 324– 331

    work page 2019

  8. [8]

    CONCLUSION We presented a novel recurrent multi-scale image-based scene flow approach that leverages successful concepts from optical flow to the field of scene flow estimation. Using a U-Net- based feature encoder, a lean context encoder, a coarse-to-fine estimation scheme based on a joint upsampling of the SE (3) field and the motion embeddings, a multi...

    work page

Show all 40 references
  1. [9]

    Object scene flow for au- tonomous vehicles,

    M. Menze and A. Geiger, “Object scene flow for au- tonomous vehicles,” in CVPR, 2015, pp. 3061–3070

    2015

  2. [10]

    Semantically coherent 4d scene flow of dynamic scenes,

    A. Mustafa and A. Hilton, “Semantically coherent 4d scene flow of dynamic scenes,”IJCV, vol. 128, pp. 319– 335, 2020

    2020

  3. [11]

    Surgem: A vision-based surgery environment model- ing framework for constructing a digital twin toward au- tonomous soft tissue manipulation,

    J. Chen, E. Kobayashi, I. Sakuma, and N. Tomi, “Surgem: A vision-based surgery environment model- ing framework for constructing a digital twin toward au- tonomous soft tissue manipulation,” IEEE RA-L, vol. 9, pp. 9789–9796, 2020

    2020

  4. [12]

    Three-dimensional scene flow,

    S. Vedula, S. Baker, R. Collins, T. Kanade, and P. Ran- der, “Three-dimensional scene flow,” in CVPR, 1999, pp. 722–729

    1999

  5. [13]

    Multi-scale 3d scene flow from binocular stereo sequences,

    R. Li and S. Sclaroff, “Multi-scale 3d scene flow from binocular stereo sequences,” CVIU, vol. 110, no. 1, pp. 75–90, 2008

    2008

  6. [14]

    Multi-view scene flow estimation: a view centered variational approach,

    T. Basha, Y . Moses, and N. Kiryati, “Multi-view scene flow estimation: a view centered variational approach,” in CVPR, 2010, pp. 1506–1513

    2010

  7. [15]

    CCMR: High resolution optical flow estimation via coarse-to- fine context-guided motion reasoning,

    A. Jahedi, M. Luz, M. Rivinius, and A. Bruhn, “CCMR: High resolution optical flow estimation via coarse-to- fine context-guided motion reasoning,” in WACV, 2024, pp. 6899–6908

    2024

  8. [16]

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

    Z. Teed and J. Deng, “RAFT: Recurrent all-pairs field transforms for optical flow,” in ECCV, 2020, pp. 402– 419

    2020

  9. [17]

    RAFT-3D: Scene flow using rigid-motion embeddings,

    Z. Teed and J. Deng, “RAFT-3D: Scene flow using rigid-motion embeddings,” in CVPR, 2021, pp. 8375– 8384

    2021

  10. [18]

    Cam- LiFlow: bidirectional camera-lidar fusion for joint opti- cal flow and scene flow estimation,

    H. Liu, T. Lu, Y . Xu, J. Liu, W. Li, and L. Chen, “Cam- LiFlow: bidirectional camera-lidar fusion for joint opti- cal flow and scene flow estimation,” inCVPR, 2022, pp. 5791–5801

    2022

  11. [19]

    Learning optical flow and scene flow with bidirectional camera- lidar fusion,

    H. Liu, T. Lu, Y . Xu, J. Liu, and L. Wang, “Learning optical flow and scene flow with bidirectional camera- lidar fusion,” IEEE TPAMI, 2023

    2023

  12. [20]

    At- tentive multimodal fusion for optical and scene flow,

    Y . Zhou, G. Mei, Y . Wang, F. Poiesi, and Y . Wan, “At- tentive multimodal fusion for optical and scene flow,” IEEE RA-L, vol. 8, pp. 6091–9098, 2023

    2023

  13. [21]

    ScaleRAFT: Cross-scale recur- rent all-pairs field transforms for 3d motion estimation,

    H. Ling and Q. Sun, “ScaleRAFT: Cross-scale recur- rent all-pairs field transforms for 3d motion estimation,” arXiv preprint arXiv:2407.09797v1, 2024

    2024 arXiv

  14. [22]

    MS-RAFT+: High resolution multi-scale RAFT,

    A. Jahedi, M. Luz, M. Rivinius, L. Mehl, and A. Bruhn, “MS-RAFT+: High resolution multi-scale RAFT,” IJCV, vol. 132, no. 5, pp. 1835–1856, 2024

    2024

  15. [23]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in CVPR, 2016, pp. 770–778

    2016

  16. [24]

    We then proceed with fine-tuning on KITTI [1] or Spring [25]

    for 200k iterations with a batch size of 3 using 368 × 768 patches and a maximum learning rate of 2 · 10−4. We then proceed with fine-tuning on KITTI [1] or Spring [25]. During fine-tuning, we train for 30k iterations with a batch size of 4 and a maximum learning rate of 6.25 ...

  17. [25]

    Unifying flow, stero and depth estima- tion,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, F. Yu, D. Tao, and A. Geiger, “Unifying flow, stero and depth estima- tion,” IEEE TPAMI, vol. 45, no. 11, pp. 13941–13958, 2023

    2023

  18. [26]

    RAFT-stereo: Multi- level recurrent field transforms for stereo matching,

    L. Lipson, Z. Teed, and J. Deng, “RAFT-stereo: Multi- level recurrent field transforms for stereo matching,” in IEEE 3DV, 2021, pp. 218–227

    2021

  19. [27]

    Practical stereo matching via cascaded recurrent network with adaptive correlation,

    J. Li, P. Wang, P. Xiong, T. Cai, Z. Yan, L. Yang, J. Liu, H. Fan, and S. Liu, “Practical stereo matching via cascaded recurrent network with adaptive correlation,” in CVPR, 2022, pp. 16242–16251

    2022

  20. [28]

    Multi-scale bidirectional re- current network with hybrid correlation for point cloud based scene flow estimation,

    W. Cheng and J. H. Ko, “Multi-scale bidirectional re- current network with hybrid correlation for point cloud based scene flow estimation,” in ICCV, 2023

    2023

  21. [29]

    Combining RAFT- based stereo disparity and optical flow models for scene flow estimation,

    H. Pan, L. Li, S. An, and H. Xie, “Combining RAFT- based stereo disparity and optical flow models for scene flow estimation,” in IEEE ICIP, 2024, pp. 3505–3511

    2024

  22. [30]

    SF2SE3: Clus- tering scene flow into SE(3)-motions via proposal and selection,

    L. Sommer, P. Schr ¨oppel, and T. Brox, “SF2SE3: Clus- tering scene flow into SE(3)-motions via proposal and selection,” in GCPR, 2022, pp. 215–229

    2022

  23. [31]

    CroCo v2: Improved cross-view completion pre-training for stereo matching and optical flow,

    P. Weinzaepfel, T. Lucas, V . Leroy, Y . Cabon, V . Arora, R. Br´egier, G. Csurka, L. Antsfeld, B. Chidlovskii, and J. Revaud, “CroCo v2: Improved cross-view completion pre-training for stereo matching and optical flow,” in ICCV, 2023, pp. 17969–17980

    2023

  24. [32]

    A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,

    N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox, “A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation,” in CVPR, 2016, pp. 4040–4048

    2016

  25. [33]

    Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo,

    L. Mehl, J. Schmalfuss, A. Jahedi, Y . Nalivayko, and A. Bruhn, “Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo,” in CVPR, 2023

    2023

  26. [34]

    Learning to segment rigid motions from two frames,

    G. Yang and D. Ramanan, “Learning to segment rigid motions from two frames,” in CVPR, 2021

    2021

  27. [35]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Dollar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in CVPR, 2017

    2017

  28. [36]

    Object scene flow for au- tonomous vehicles,

    M. Menze and A. Geiger, “Object scene flow for au- tonomous vehicles,” in CVPR, 2015. 6 Supplementary Material In the following, we first show the architecture of our 4-scale model. Then we elaborate on our employed context encoder and finally, we demonstrate more visual resul...

    2015

  29. [37]

    and the Spring [25] benchmark

  30. [38]

    It can be seen that in addition to the three scales at [ 1 16, 1 8, 1 4], the SE(3) field is also refined at 1 2 resolution

    ARCHITECTURE OF MS-RAFT-3D+ Figure 3 shows the architecture of our 4-scale MS-RAFT-3D+ model. It can be seen that in addition to the three scales at [ 1 16, 1 8, 1 4], the SE(3) field is also refined at 1 2 resolution. This allows to capture more details from images. Besides, ...

  31. [39]

    The architecture is shown in Figure 4

    CONTEXT ENCODER We use a simple top-down feature extractor to compute con- text features. The architecture is shown in Figure 4. The num- bers in brackets show the number of channels that is output by each module. Note that the number of context encoder channels in the ablatio...

  32. [40]

    In both cases, our approach achieves detailed re- sults and lower errors

    QUALITATIVE RESULTS We present more qualitative results of our method from the Spring benchmark in Figure 5 and from the KITTI benchmark in Figure 6. In both cases, our approach achieves detailed re- sults and lower errors. Importantly, in the case of KITTI, as the top 80 pixe...

Pith tools

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