Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

JointSplat: Probabilistic Joint Flow-Depth Optimization for Sparse-View Gaussian Splatting

T0 review · 5 major / 6 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 The paper claims that gating depth-flow fusion by per-pixel optical-flow matching probability sets a new state of the art in sparse-view Gaussian splatting, with 27.53 dB PSNR on RealEstate10K at just 125M parameters.

desk verdict Incremental but plausible: a small PSNR gain from a probabilistic flow-depth mask, with reproducibility gaps and overclaimed generality. read the letter →

arxiv 2506.03872 v1 pith:YW4OK63E submitted 2025-06-04 cs.CV cs.AI

Yang Xiao , Guoan Xu , Qiang Wu , Wenjing Jia This is my paper
classification cs.CVcs.AI
keywords sparse-viewreconstructionnovelviewsynthesis3DGaussiansplattingjointflow-depthestimationopticalflowconfidencedepthconsistencylossfeed-forwardmulti-view
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 sparse-view 3D reconstruction, which usually builds Gaussian-splat scenes from predicted depth alone, is improved by jointly estimating optical flow and then trusting it only where a per-pixel matching probability says the flow is reliable. This probabilistic weighting is the proposed answer to a known failure: in low-texture, repetitive, or occluded regions, depth matching and optical flow are wrong in complementary ways, and naive fusion of the two propagates errors. JointSplat reports 27.53 dB PSNR on RealEstate10K, a new best among compared feed-forward methods, achieved with 125M parameters, one-third the size of the larger DepthSplat variant it also edges out. On ACID, a zero-shot outdoor benchmark, it reaches 28.37 dB PSNR, ranking second and matching the larger model. The practical target is fast, generalizable novel-view synthesis from sparse image sets for augmented reality, robotics, and autonomous driving.

What carries the argument

The load-bearing object is the flow probability mask $M_{\text{flow}}(u)=M_{\text{occ}}(u)\cdot f_c(u)$, where $f_c(u)=\max_{\delta\in\mathcal{N}(u)}\frac{\exp(s_{u,\delta})}{\sum_{\delta'}\exp(s_{u,\delta'})}$ is the maximum softmax matching probability inside a depth-adaptive search window and $M_{\text{occ}}$ is a depth-based occlusion mask obtained from feature correlation after depth-guided warping. This mask scales the fusion of hybrid and flow-triangulated depths and reweights the multi-view depth-consistency loss $L_{\text{mvd}}=M_{\text{flow}}\cdot|D^{\text{refine}}_i-\hat{D}^{\text{refine}}_i|$, so confident matches are amplified and unreliable or occluded ones suppressed.

What would settle it

Compute flow endpoint error against the predicted probability $f_c(u)$ on a held-out split where known depth and pose make the true correspondence exactly knowable; the central claim predicts a strong negative correlation, so the absence of that correlation, such as top-quintile $f_c$ pixels having no better endpoint error than bottom-quintile pixels, would falsify the mechanism.

Watch

Extended reading notes

Core claim

JointSplat's central claim is that optical flow and depth are complementary for sparse-view Gaussian splatting, but only if their fusion is gated by a probabilistic measure of flow reliability. The framework computes a depth-aware flow-matching probability map, the maximum softmax similarity over a depth-adaptive search window, multiplies it by a depth-based occlusion mask, and uses the resulting flow probability mask to steer both a residual depth refinement and a multi-view depth-consistency loss. On RealEstate10K, this probabilistic fusion reaches a PSNR of 27.53 dB, SSIM 0.889, and LPIPS 0.113, beating the DepthSplat ViT-B backbone by 0.19 dB and the ViT-L variant at 354M parameters despite using only 125M parameters. On ACID it matches DepthSplat-ViT-L at 28.37 dB PSNR while ranking just behind HiSplat. The authors attribute the gains to the probabilistic mechanism rather than to the backbone, since the same ViT-B backbone without flow-depth fusion scores lower.

Load-bearing premise

The load-bearing premise is that the maximum softmax matching probability inside the depth-adaptive search window faithfully indicates whether the flow estimate is correct, since a high score on a wrong peak, or a true match outside the window, would make the mask upweight exactly the wrong correspondences.

Editorial extensions

If this is right

  • Sparse-view Gaussian splatting can be improved without scaling up the network, since a 125M-parameter model surpasses a 354M-parameter counterpart by replacing binary flow masks with continuous match-probability weighting.
  • Flow-based depth triangulation becomes usable for view consistency even though no ground-truth flow is available during training, because the flow probability mask suppresses erroneous correspondences before they enter the depth-consistency loss.
  • On low-texture and repetitive scenes, where depth-only cost volumes wander, the flow cue with probability gating recovers local structure, reducing artifacts and improving cross-view depth alignment.
  • The framework transfers zero-shot to outdoor aerial scenes on ACID at 28.37 dB PSNR, indicating the mechanism does not overfit to indoor RealEstate10K statistics.

Reading between the lines

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

  • Because the flow probability mask is a continuous soft weight derived inside the differentiable matching layer, the same gating could serve other geometric losses, such as pose, scene-flow, or stereo depth, wherever correspondences are known to be unreliable; the paper does not test these.
  • The search radii $r_{\min}$ and $r_{\max}$ in the depth-adaptive window are never reported, leaving the window size an untested free parameter; a natural extension would compare fixed, depth-linear, and learned radius schedules.
  • The KITTI zero-shot result (Abs Rel 64.58 versus 56.83 for DepthSplat) suggests the flow-depth probabilistic fusion can hurt in long-range outdoor driving scenes with sparse depth supervision, so the method's advantage may be domain-dependent rather than universal.
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

5 major / 6 minor

Summary. JointSplat proposes a feed-forward sparse-view 3D Gaussian splatting method that combines multi-view depth estimation with optical flow through a pixel-wise probabilistic weighting mechanism. The method builds on DepthSplat (ViT-B) and adds: (i) a depth-adaptive flow-matching confidence map computed as the maximum softmax over a depth-dependent search window, (ii) a depth-based occlusion mask that gates the confidence, (iii) flow-based triangulation used to refine the hybrid depth via a residual prediction, and (iv) a multi-view depth consistency loss reweighted by the flow probability mask. The model is trained with self-supervised flow losses and the standard rendering loss on RealEstate10K and evaluated on RealEstate10K, ACID, and zero-shot depth benchmarks (TartanAir, ScanNet, KITTI). The central claim is that the probabilistic flow-depth fusion, rather than the backbone, drives a 0.19 dB PSNR improvement over DepthSplat ViT-B (27.53 vs 27.34) and establishes a new SOTA on RealEstate10K.

Significance. If the reported gains are reproducible, the paper makes a modest but useful contribution: it demonstrates a simple confidence-gating scheme for merging optical flow and depth cues in sparse-view Gaussian splatting, with ablations suggesting that each component (occlusion mask, matching probability, depth refinement, multi-view consistency loss) contributes positively. The method's parameter count (125M) is competitive with DepthSplat ViT-B. However, the significance is limited by the small magnitude of the reported improvement, the absence of error bars, and the lack of reported values for key parameters (r_min, r_max) that the method's correctness depends on. The physical motivation for the depth-adaptive window appears to be backwards for perspective cameras, which casts doubt on the mechanism's stated rationale. The claims of consistently outperforming SOTA are also overstated given the ACID and KITTI results. The paper is not a breakthrough but could be a solid incremental work if the reproducibility issues are addressed.

major comments (5)
  1. [Sec. 3.1 (Eqs. 1–5)] The depth-adaptive search window is a load-bearing component, but the paper never reports the values of r_min and r_max, and the monotonic relationship in Eq. (2) is physically backwards for standard perspective projection: nearby points produce larger image displacements than distant points under camera translation, so larger search windows should be assigned to small normalized depth, not large depth. With the stated dependence, true matches for close objects will fall outside N(u), and because f_c(u) in Eq. (5) is the maximum softmax over N(u), the network can assign high confidence to an incorrect offset, which is then used to upweight the wrong correspondences in Eq. (13). Please report the radii, correct the direction of Eq. (2) if needed, and provide at least a sensitivity analysis over the window size.
  2. [Abstract and Sec. 4.1 (Tab. 1) and Appendix A.2 (Tab. 4)] The statement that JointSplat 'consistently outperforms SOTA' is contradicted by the paper's own numbers. On ACID, JointSplat (PSNR 28.37, SSIM 0.848, LPIPS 0.141) ties DepthSplat on PSNR/LPIPS, is marginally above on SSIM, and is clearly below HiSplat (28.66, 0.850, 0.137); on KITTI, the method regresses sharply relative to DepthSplat (Abs Rel 64.58 vs 56.83; δ1 20.05 vs 46.26). The RealEstate10K SOTA claim may still be defensible, but the abstract and conclusion must be reworded to describe the scope accurately.
  3. [Sec. 4.1 (Tab. 2) and Sec. 4.2 (Tab. 3)] The improvement over the DepthSplat ViT-B backbone is 0.19 dB PSNR, 0.002 SSIM, and −0.003 LPIPS, yet no error bars, standard deviations, or multiple-seed experiments are provided. Given that the ablation table attributes 0.44 dB to the flow probability mask and that different mask designs span a 0.82 dB range (26.71–27.53), the reader cannot determine whether the headline 0.19 dB gain is statistically significant or within run-to-run variation. Please report variance or confidence intervals for the main comparison and for the key ablations.
  4. [Sec. 3.1 (Eq. 5) and Sec. 3.2 (Eq. 9)] The symbol f_c(u) is used ambiguously: in Sec. 3.1 it denotes both the expected-displacement flow vector and the per-pixel matching probability (maximum softmax value), while Eq. (9) uses F_i^flow for the optical flow without stating how it relates to f_c(u). This ambiguity makes it impossible to determine whether the flow used in triangulation comes from the pretrained GMFlow network or from the depth-adaptive matching layer, which is essential for understanding the method. Please fix the notation and explicitly define the flow estimate that enters Eq. (9).
  5. [Sec. 3.1 (Eq. 5)] The quantity f_c(u) is called a matching probability, but it is the maximum of a softmax over a fixed set of candidate offsets within N(u); it is a peakiness or relative-confidence measure and not a calibrated probability of the match being correct. In particular, if the true correspondence lies outside N(u), f_c(u) can be close to 1 for a wrong offset. The paper should explicitly acknowledge this limitation and quantify the risk by reporting the window radii and the fraction of pixels whose best match lands at the boundary of N(u).
minor comments (6)
  1. [Eq. (9)] There is a typo in the definition of the matched pixel: 'u^{t+1}_i = u^j_i + F_i^flow' should presumably be 'u^{i+1}_j = u^j_i + F_i^flow(u^j_i)'.
  2. [Tab. 3] The rows 'w/ Forward-Backward Consistency Mask' and 'w/o Flow Probability Mask' report identical numbers (27.09) and the text explains that they are the same setting; listing them twice is confusing and should be merged or clarified.
  3. [Eq. (7)] The occlusion mask uses a sigmoid threshold at τ=0.5, which is equivalent to requiring the normalized dot product to be positive; the choice is not justified or ablated.
  4. [Sec. 4, Implementation Details] The hyper-parameters λ_s1, λ_s2, λ_c, λ_g, λ_m are said to 'follow prior work', but no prior-work citation is given for these specific values; please provide one.
  5. [Sec. 3.3] The census and smoothness losses are mentioned without equations; given that the self-supervised flow training is part of the method, at least the smoothness formulation (first- and second-order derivatives) should be specified.
  6. [Sec. 1 and Sec. 4.1 (Tab. 2)] The parameter counts reported for DepthSplat (120M / 354M) and JointSplat (125M / 359M) differ only slightly; 'one-third of its counterpart' refers to the ViT-L variant and should be stated explicitly in the introduction to avoid the impression that it is one-third of the ViT-B baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the flow-confidence weighting is self-referential but not a fitted-input prediction, and the final PSNR is evaluated against held-out ground truth.

full rationale

JointSplat's derivation chain is not circular. The only candidate is the flow probability mask M_flow = M_occ · f_c, where f_c (Eq. 5) is the maximum softmax over the same depth-adaptive window used to form the flow vector. Using this self-confidence as a weight in the depth-consistency loss (Eq. 13) and in the residual fusion (Eq. 10) is a training-time reweighting scheme, not a fitted parameter renamed as a prediction. The paper's reported prediction — novel-view PSNR on RealEstate10K/ACID — is obtained by rendering from the refined depth and comparing to ground-truth images, an external benchmark that does not reduce to the mask by construction. No self-citation chain is load-bearing; the cited priors (GMFlow, GMDepth, DepthSplat) are third-party and the backbone is a published baseline. The unstated r_min/r_max values and the KITTI regression are correctness/generalization risks, not circularity. Therefore the claim that probabilistic flow-depth fusion, rather than the backbone, drives the +0.19 dB gain is a substantive empirical assertion, even if the confidence signal is self-referential.

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

The central mechanism depends on hand-set search radii, a fixed occlusion threshold, and manual loss weights; on pretrained backbones (Depth Anything V2, GMDepth, GMFlow) treated as fixed inputs; on the static-scene assumption in flow triangulation; and on the untested assumption that a softmax maximum is a trustworthy confidence score. No new physical entities are introduced.

free parameters (4)
  • r_min, r_max = not reported
    Hand-set bounds of the depth-adaptive search radius (Eq. 2). They control whether large-displacement matches are ever considered; if too small, the key matching probability is computed over the wrong window.
  • tau (occlusion threshold) = 0.5
    Hand-set threshold in Eq. (7) that converts the sigmoid feature correlation into a binary occlusion mask. It directly determines which pixels are zeroed in the flow probability mask.
  • lambda_s1, lambda_s2, lambda_c, lambda_g, lambda_m = 0.0025, 0.0025, 0.1, 0.1, 0.1
    Weights of the smoothness, census, GCC, and multi-view depth consistency losses in Eq. (15). They balance the flow/depth supervision and were chosen following prior work, not derived.
  • lambda (LPIPS weight) = 0.05
    Weight of the perceptual loss in Eq. (14). Set by hand, not fitted to data or justified by sensitivity analysis.
assumptions (4)
  • domain assumption Static scene and camera ego-motion for flow-based triangulation
    Eq. (9) triangulates depth from optical flow under the assumption of no dynamic objects. The paper acknowledges this in Sec. 3.2: 'This triangulation assumes a static scene and camera ego-motion, and may fail in dynamic regions.'
  • domain assumption Pretrained GMFlow/GMDepth on FlyingThings transfer to RealEstate10K and ACID
    The method relies on optical flow and cost-volume features from networks trained on synthetic data. No fine-tuning of these backbones is described, so the entire pipeline depends on this transfer.
  • ad hoc to paper Maximum softmax matching probability is a valid per-pixel confidence measure
    Eq. (5) defines f_c as the max softmax over a depth-dependent window. The paper does not justify that this quantity tracks true match reliability, and it fails when the true match is outside the window.
  • domain assumption Hybrid depth from Depth Anything V2 and GMDepth provides enough geometric prior for warping and occlusion detection
    Eq. (6) and Eq. (7) warp features with predicted depth. If depth is poor in textureless regions, the occlusion mask will be poor, and the flow probability mask inherits the error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of JointSplat: Probabilistic Joint Flow-Depth Optimization for Sparse-View Gaussian Splatting." pith.science (2026). https://pith.science/paper/YW4OK63E

@misc{pith2026250603872,
  author       = {Pith},
  title        = {Pith review of: JointSplat: Probabilistic Joint Flow-Depth Optimization for Sparse-View Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YW4OK63E}},
  note         = {Machine review of arXiv:2506.03872}
}
read the original abstract

Reconstructing 3D scenes from sparse viewpoints is a long-standing challenge with wide applications. Recent advances in feed-forward 3D Gaussian sparse-view reconstruction methods provide an efficient solution for real-time novel view synthesis by leveraging geometric priors learned from large-scale multi-view datasets and computing 3D Gaussian centers via back-projection. Despite offering strong geometric cues, both feed-forward multi-view depth estimation and flow-depth joint estimation face key limitations: the former suffers from mislocation and artifact issues in low-texture or repetitive regions, while the latter is prone to local noise and global inconsistency due to unreliable matches when ground-truth flow supervision is unavailable. To overcome this, we propose JointSplat, a unified framework that leverages the complementarity between optical flow and depth via a novel probabilistic optimization mechanism. Specifically, this pixel-level mechanism scales the information fusion between depth and flow based on the matching probability of optical flow during training. Building upon the above mechanism, we further propose a novel multi-view depth-consistency loss to leverage the reliability of supervision while suppressing misleading gradients in uncertain areas. Evaluated on RealEstate10K and ACID, JointSplat consistently outperforms state-of-the-art (SOTA) methods, demonstrating the effectiveness and robustness of our proposed probabilistic joint flow-depth optimization approach for high-fidelity sparse-view 3D reconstruction.

Figures

Figures reproduced from arXiv: 2506.03872 by the authors.

Figure 1
Figure 1. Comparison between JointSplat and previous methods. JointSplat outperforms prior feed-forward method [7] and joint flow-depth baseline [11] in both depth estimation and sparse-view synthesis, especially in regions with repetitive or weak textures (see the red box), and large displace￾ments or occlusions (see the blue box). By introducing a probabilistic optimization mechanism, our method enables more robust geometri… view at source ↗
Figure 2
Figure 2. The system flow of our proposed JointSplat. Given sparse-view images Ii with camera intrinsics K and relative poses Ti→i+1, we extract hybrid depth D hyb i and optical flow features F flow i using pretrained models [25–27]. A probabilistic optimization mechanism is then introduced to scale the information fusion between depth and flow to obtain refined depth estimation Dref i , which is back-projected to determine G… view at source ↗
Figure 3
Figure 3. Qualitative comparison with SOTA methods. cases, our method produces images with accurate alignment and finer details. Notably, in the first scene, our result shows fewer artifacts than DepthSplat [7] (see the blue box). This improvement can be attributed to the explicit incorporation of the probabilistic optimization mechanism in our depth estimation pipeline. By fully leveraging the complementarity between optical… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with DepthSplat [7] on depth estimation. Blue boxes highlight regions with finer structural details revealed by the proposed JointSplat. Specifically, JointSplat produces sharper object boundaries, better preserves geometric structures, and more …
Figure 5
Figure 5. Figure 5: Qualitative comparison of occlusion mask designs. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Sparse-View 3D Reconstruction: Recent Advances and Open Challenges

    cs.CV 2025-07 conditional novelty 3.0 of 10

    A comprehensive survey that organizes sparse-view 3D reconstruction methods into geometry-based, NeRF, 3DGS, and diffusion-based categories, with benchmarks and open challenges.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Visual Language Maps for Robot Navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual Language Maps for Robot Navigation,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 10 608–10 615

    work page 2023

  2. [2]

    ActiveGAMER: Active GAussian Mapping through Efficient Rendering,

    L. Chen, H. Zhan, K. Chen, X. Xu, Q. Yan, C. Cai, and Y . Xu, “ActiveGAMER: Active GAussian Mapping through Efficient Rendering,”arXiv preprint arXiv:2501.06897, 2025

    arXiv 2025

  3. [3]

    Toward Real-world BEV Perception: Depth Uncertainty Estimation via Gaussian Splatting,

    S.-W. Lu, Y .-H. Tsai, and Y .-T. Chen, “Toward Real-world BEV Perception: Depth Uncertainty Estimation via Gaussian Splatting,”arXiv preprint arXiv:2504.01957, 2025

    arXiv 2025

  4. [4]

    3D Gaussian Splatting for Real-time Radiance Field Rendering,

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3D Gaussian Splatting for Real-time Radiance Field Rendering,”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    work page 2023

  5. [5]

    pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction,

    D. Charatan, S. L. Li, A. Tagliasacchi, and V . Sitzmann, “pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 19 457–19 467

    work page 2024

  6. [6]

    MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images,

    Y . Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T.-J. Cham, and J. Cai, “MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 370–386

    work page 2024

  7. [7]

    DepthSplat: Connecting Gaussian Splatting and Depth,

    H. Xu, S. Peng, F. Wang, H. Blum, D. Barath, A. Geiger, and M. Pollefeys, “DepthSplat: Connecting Gaussian Splatting and Depth,”arXiv preprint arXiv:2410.13862, 2024

    arXiv 2024

  8. [8]

    TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers,

    C. Zhang, Y . Zou, Z. Li, M. Yi, and H. Wang, “TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 9, 2025, pp. 9869–9877

    work page 2025

Show all 46 references
  1. [9]

    DCPI-Depth: Explicitly Infusing Dense Correspondence Prior to Unsupervised Monocular Depth Estimation,

    M. Zhang, Y . Feng, Q. Chen, and R. Fan, “DCPI-Depth: Explicitly Infusing Dense Correspondence Prior to Unsupervised Monocular Depth Estimation,”arXiv preprint arXiv:2405.16960, 2024

    2024 arXiv

  2. [10]

    Invariant Feature Matching in Spacecraft Rendezvous and Docking Optical Imaging Based on Deep Learning,

    D. Guo, S. Wu, D. Weng, C. Gao, and W. Li, “Invariant Feature Matching in Spacecraft Rendezvous and Docking Optical Imaging Based on Deep Learning,”Remote Sensing, vol. 16, no. 24, p. 4690, 2024

    2024

  3. [11]

    CoherentGS: Sparse Novel View Synthesis with Coherent 3D Gaussians,

    A. Paliwal, W. Ye, J. Xiong, D. Kotovenko, R. Ranjan, V . Chandra, and N. K. Kalantari, “CoherentGS: Sparse Novel View Synthesis with Coherent 3D Gaussians,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 19–37

    2024

  4. [12]

    Unsupervised Learning of Depth, Optical Flow and Pose with Occlusion from 3D Geometry,

    G. Wang, C. Zhang, H. Wang, J. Wang, Y . Wang, and X. Wang, “Unsupervised Learning of Depth, Optical Flow and Pose with Occlusion from 3D Geometry,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 1, pp. 308–320, 2020

    2020

  5. [13]

    Learning Optical Flow, Depth, and Scene Flow without Real-World Labels,

    V . Guizilini, K.-H. Lee, R. Ambru¸ s, and A. Gaidon, “Learning Optical Flow, Depth, and Scene Flow without Real-World Labels,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 3491–3498, 2022

    2022

  6. [14]

    Dense Point Trajectories by GPU-accelerated Large Displacement Optical Flow,

    N. Sundaram, T. Brox, and K. Keutzer, “Dense Point Trajectories by GPU-accelerated Large Displacement Optical Flow,” inEuropean conference on computer vision. Springer, 2010, pp. 438–451

    2010

  7. [15]

    Learning a Confidence Measure for Optical Flow,

    O. Mac Aodha, A. Humayun, M. Pollefeys, and G. J. Brostow, “Learning a Confidence Measure for Optical Flow,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 5, pp. 1107–1120, 2012

    2012

  8. [16]

    Stereo Magnification: Learning View Synthesis using Multiplane Images,

    T. Zhou, R. Tucker, J. Flynn, G. Fyffe, and N. Snavely, “Stereo Magnification: Learning View Synthesis using Multiplane Images,”arXiv preprint arXiv:1805.09817, 2018

    2018 arXiv

  9. [17]

    Infinite Nature: Perpetual View Generation of Natural Scenes from a Single Image,

    A. Liu, R. Tucker, V . Jampani, A. Makadia, N. Snavely, and A. Kanazawa, “Infinite Nature: Perpetual View Generation of Natural Scenes from a Single Image,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 14 458–14 467

    2021

  10. [18]

    Soft 3D Reconstruction for View Synthesis,

    E. Penner and L. Zhang, “Soft 3D Reconstruction for View Synthesis,”ACM Transactions on Graphics (TOG), vol. 36, no. 6, pp. 1–11, 2017

    2017

  11. [19]

    Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations,

    X. Zheng and D. Xu, “Flow-NeRF: Joint Learning of Geometry, Poses, and Dense Flow within Unified Neural Representations,”arXiv preprint arXiv:2503.10464, 2025

    2025 arXiv

  12. [20]

    HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction,

    S. Tang, W. Ye, P. Ye, W. Lin, Y . Zhou, T. Chen, and W. Ouyang, “HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction,”arXiv preprint arXiv:2410.06245, 2024. 10

    2024 arXiv

  13. [21]

    GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose,

    Z. Yin and J. Shi, “GeoNet: Unsupervised Learning of Dense Depth, Optical Flow and Camera Pose,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1983–1992

    2018

  14. [22]

    CamLiFlow: Bidirectional Camera-LiDAR Fusion for Joint Optical Flow and Scene Flow Estimation,

    H. Liu, T. Lu, Y . Xu, J. Liu, W. Li, and L. Chen, “CamLiFlow: Bidirectional Camera-LiDAR Fusion for Joint Optical Flow and Scene Flow Estimation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5791–5801

    2022

  15. [23]

    RPEFlow: Multimodal Fusion of RGB-PointCloud-Event for Joint Optical Flow and Scene Flow Estimation,

    Z. Wan, Y . Mao, J. Zhang, and Y . Dai, “RPEFlow: Multimodal Fusion of RGB-PointCloud-Event for Joint Optical Flow and Scene Flow Estimation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 030–10 040

    2023

  16. [24]

    Real-Time Simultaneous 3D Reconstruction and Optical Flow Estimation,

    M. Roxas and T. Oishi, “Real-Time Simultaneous 3D Reconstruction and Optical Flow Estimation,” in 2018 IEEE Winter Conference on Applications of Computer Vision (WACV). IEEE, 2018, pp. 885–893

    2018

  17. [25]

    Depth Anything V2,

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao, “Depth Anything V2,”Advances in Neural Information Processing Systems, vol. 37, pp. 21 875–21 911, 2024

    2024

  18. [26]

    Unifying Flow, Stereo and Depth Estimation,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, F. Yu, D. Tao, and A. Geiger, “Unifying Flow, Stereo and Depth Estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 11, pp. 13 941–13 958, 2023

    2023

  19. [27]

    GMFlow: Learning Optical Flow via Global Matching,

    H. Xu, J. Zhang, J. Cai, H. Rezatofighi, and D. Tao, “GMFlow: Learning Optical Flow via Global Matching,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8121–8130

    2022

  20. [28]

    U-Net: Convolutional Networks for Biomedical Image Segmen- tation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmen- tation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. ...

    2015

  21. [29]

    High-Resolution Image Synthesis with Latent Diffusion Models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-Resolution Image Synthesis with Latent Diffusion Models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  22. [30]

    FlowNet: Learning Optical Flow with Convolutional Networks,

    A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas, V . Golkov, P. Van Der Smagt, D. Cremers, and T. Brox, “FlowNet: Learning Optical Flow with Convolutional Networks,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 2758–2766

    2015

  23. [31]

    Learning Feature Descriptors using Camera Pose Supervision,

    Q. Wang, X. Zhou, B. Hariharan, and N. Snavely, “Learning Feature Descriptors using Camera Pose Supervision,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16. Springer, 2020, pp. 757–774

    2020

  24. [32]

    End-to-End Learning of Geometry and Context for Deep Stereo Regression,

    A. Kendall, H. Martirosyan, S. Dasgupta, P. Henry, R. Kennedy, A. Bachrach, and A. Bry, “End-to-End Learning of Geometry and Context for Deep Stereo Regression,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 66–75

    2017

  25. [33]

    Attention Is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention Is All You Need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  26. [34]

    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

  27. [35]

    UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss,

    S. Meister, J. Hur, and S. Roth, “UnFlow: Unsupervised Learning of Optical Flow with a Bidirectional Census Loss,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  28. [36]

    Occlusion Aware Unsupervised Learning of Optical Flow,

    Y . Wang, Y . Yang, Z. Yang, L. Zhao, P. Wang, and W. Xu, “Occlusion Aware Unsupervised Learning of Optical Flow,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4884–4893

    2018

  29. [37]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The Unreasonable Effectiveness of Deep Features as a Perceptual Metric,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 586–595

    2018

  30. [38]

    Adam: A Method for Stochastic Optimization,

    D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,”arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  31. [39]

    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,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, p...

    2016

  32. [40]

    pixelNeRF: Neural Radiance Fields from One or Few Images,

    A. Yu, V . Ye, M. Tancik, and A. Kanazawa, “pixelNeRF: Neural Radiance Fields from One or Few Images,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 4578–4587

    2021

  33. [41]

    Generalizable Patch-Based Neural Rendering,

    M. Suhail, C. Esteves, L. Sigal, and A. Makadia, “Generalizable Patch-Based Neural Rendering,” in European Conference on Computer Vision. Springer, 2022, pp. 156–174

    2022

  34. [42]

    Learning to Render Novel Views from Wide-Baseline Stereo Pairs,

    Y . Du, C. Smith, A. Tewari, and V . Sitzmann, “Learning to Render Novel Views from Wide-Baseline Stereo Pairs,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4970–4980

    2023

  35. [43]

    MuRF: Multi- Baseline Radiance Fields,

    H. Xu, A. Chen, Y . Chen, C. Sakaridis, Y . Zhang, M. Pollefeys, A. Geiger, and F. Yu, “MuRF: Multi- Baseline Radiance Fields,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 041–20 050

    2024

  36. [44]

    Vision meets robotics: The KITTI dataset,

    A. Geiger, P. Lenz, C. Stiller, and R. Urtasun, “Vision meets robotics: The KITTI dataset,”The international journal of robotics research, vol. 32, no. 11, pp. 1231–1237, 2013

    2013

  37. [45]

    TartanAir: A Dataset to Push the Limits of Visual SLAM,

    W. Wang, D. Zhu, X. Wang, Y . Hu, Y . Qiu, C. Wang, Y . Hu, A. Kapoor, and S. Scherer, “TartanAir: A Dataset to Push the Limits of Visual SLAM,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 4909–4916

    2020

  38. [46]

    ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5828–5839. A Appendix A.1 Limitations Despite th...

    2017

Pith tools

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