Pith. sign in

REVIEW 3 major objections 6 minor 66 references

Exploring Non-Local Spatial-Angular Correlations with a Hybrid Mamba-Transformer Framework for Light Field Super-Resolution

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

Pith's one-line read The paper presents LFMT, a hybrid Mamba-Transformer network that claims state-of-the-art light field super-resolution by scanning each 2D subspace once and refining epipolar features with parallel Mamba and Transformer branches.

desk verdict Solid, well-ablated LFSR engineering whose SOTA claim is undercut by the authors' own reference: they cite LFTransMamba [56] but never compare to it, and the 'first time' hybrid claim is contradicted by that citation. read the letter →

arxiv 2509.04824 v1 pith:CDHZOXBM submitted 2025-09-05 cs.CV cs.AI

classification cs.CVcs.AI
keywords lightfieldsuper-resolutionstatespacemodelMambaTransformerepipolarplaneimagesubspacescanningspatial-angularcorrelationhybridarchitecture
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

This paper tackles light field image super-resolution (LFSR), the task of increasing the spatial resolution of every view in a 4D light field. The authors claim that existing Mamba-based LFSR methods waste computation by scanning the same pixels repeatedly across coupled subspaces, and that pure Transformer methods are too costly. They introduce LFMT, which slices the light field into spatial, angular, and epipolar-plane 2D subspaces, scans each with a single unidirectional pass (Sub-SS), then models non-local correlations in two stages: coarse spatial-angular extraction, then deep epipolar refinement with parallel Mamba and Transformer branches. On five standard benchmarks, they report an average 0.18 dB PSNR improvement over the previous state of the art LF-DET for both ×2 and ×4 upscaling, with 2.19M parameters and 66.72G FLOPs. If correct, this would make LFMT the current best-performing LFSR model and a template for hybrid Mamba-Transformer design in other light-field tasks.

What carries the argument

The central object is Subspace Simple Scanning (Sub-SS): instead of scanning the light field in multiple directions within each 2D subspace, LFMT applies one unidirectional scan per subspace, covering spatial, angular, and epipolar-plane views without re-scanning coupled pixels. The Subspace Simple Mamba Block (SSMB) wraps this scan, adding a symmetric convolution+SiLU branch to compensate for sequential information loss; Residual Subspace Mamba Blocks (RSMB) add channel attention and learnable residuals; and the stage-II dual branches — EPMB (Mamba) and EPTB (Transformer) — process horizontal and vertical epipolar planes in parallel, letting disparity-encoding slanted-line structures refine

What would settle it

A reader could select LFMT's hyperparameters using only the training split of the five datasets, freeze the configuration, and then compare against LF-DET on the 23 test scenes; if the average PSNR advantage falls below roughly 0.1 dB or changes sign on any scale, the reported 0.18 dB edge is not stable.

Watch

Extended reading notes

Core claim

LFMT slices the 4D light field into 2D subspaces — sub-aperture images (spatial domain), macro-pixel images (angular domain), and epipolar-plane images (epipolar domain) — and replaces the multi-directional scanning used in prior Mamba-based LFSR with a single unidirectional scan per subspace, called Subspace Simple Scanning (Sub-SS). The authors argue that multi-directional scanning redundantly models coupled pixels and amplifies noise, and they show that Sub-SS both lowers FLOPs and raises PSNR relative to bidirectional and four-way scanning baselines. On top of this, they build a two-stage architecture: Stage I coarsely extracts spatial-angular features using Spatial-Angular Residual Subs

Load-bearing premise

The authors chose the number of blocks and which features to fuse by maximizing average PSNR on the same five test benchmarks used for the final comparison, so the reported 0.18 dB gain may partly reflect tuning to those specific test sets rather than a general model advantage.

Editorial extensions

If this is right

  • Substituting per-subspace unidirectional scanning for multi-directional scanning is claimed to reduce redundant computation and improve PSNR simultaneously (Table III).
  • The two-stage coarse-to-fine design is claimed to increase angular consistency, measured by lower per-view PSNR standard deviation and better EPI texture, and to improve downstream depth estimation (Fig. 6).
  • Combining Mamba and Transformer branches in the epipolar-plane domain is claimed to outperform either mechanism alone at comparable parameter counts (Table VI).
  • The model's strength on large-disparity scenes (e.g., STFgantry) suggests the epipolar-plane refinement transfers disparity structure into super-resolved images.
  • The tiny variant (1.37M parameters, 42.33G FLOPs) matches LF-DET's PSNR, indicating the design can trade depth for efficiency on resource-constrained hardware.

Reading between the lines

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

  • The test-set-driven architecture search (Tables VII and VIII) probably inflates the reported 0.18 dB margin; a validation-split-based selection is needed to know the true gain.
  • The paper cites LFTransMamba as a hybrid Mamba-Transformer LFSR method but excludes it from the quantitative table, so the claimed margin over hybrids is unverified.
  • Because the epipolar-plane branch encodes disparity as slanted-line structure, the Sub-SS scanning principle could transfer to angular super-resolution and depth estimation with minor changes.
  • The ablation results in Table VIII show that removing the shallow F_SA feature from the fused set is at least as good as keeping it, while the text describes the opposite; this internal inconsistency suggests the fusion design is not fully understood.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes LFMT, a hybrid Mamba-Transformer network for light field image super-resolution (LFSR). The method introduces a Subspace Simple Scanning (Sub-SS) strategy to reduce redundancy in multi-directional subspace scanning, builds Subspace Simple Mamba Blocks (SSMB/RSMB), and uses a dual-stage modeling pipeline: Stage I extracts spatial-angular features via SA-RSMB, while Stage II refines them in the epipolar plane with parallel EPMB (Mamba) and EPTB (Transformer) branches. On five standard LFSR benchmarks, LFMT reports an average PSNR improvement of 0.18 dB over LF-DET for both ×2 and ×4 SR, with about 2.19M parameters and 66.72G FLOPs. Extensive ablation studies evaluate scanning strategies, component contributions, block counts, and feature fusion choices.

Significance. If the reported results are reproducible, the paper makes a useful empirical contribution: Sub-SS is a simple and plausible remedy for the redundancy of multi-directional scanning in LF Mamba models, and the hybrid Mamba-Transformer design is a timely direction. The paper provides a clearly written framework, releases code, and includes thorough ablations and qualitative evaluation. However, the central 'state-of-the-art' claim is not yet established because the directly relevant hybrid baseline LFTransMamba [56] is cited but excluded from all quantitative comparisons, and the architecture's hyperparameters are selected on the same test benchmarks used for final evaluation. These issues are fixable but require careful revision.

major comments (3)
  1. [Sec. II.C / Table I] The paper cites LFTransMamba [56] as a prior hybrid Mamba-Transformer model for LFSR, but this baseline is absent from all quantitative comparisons in Table I. Since [56] is the closest architectural competitor, the reported 0.18 dB margin over LF-DET does not substantiate the abstract's claim of 'significantly outperforms current state-of-the-art methods.' The Introduction's statement that the Mamba-Transformer combination is introduced 'for the first time' for LFSR is also directly contradicted by [56]. Please add LFTransMamba to the comparisons or explicitly temper the SOTA/novelty claims.
  2. [Sec. IV.C, Tables VII/VIII] The architecture is tuned on the evaluation benchmarks: Table VII selects the numbers of SA-RSMB/EPMB/EPTB blocks, and Table VIII selects the feature-fusion set (e.g., 'w/o F_SA' at 32.66 dB vs. 'with all' at 32.62 dB), both by maximizing average PSNR on the same five test sets used in Table I. This test-set model selection can inflate reported gains. Please provide a separate validation split or nested evaluation, and report variability across runs or scenes to support the significance claim.
  3. [Sec. IV.B, Table I] The reported improvements over the next-best method are small (0.05–0.18 dB) and no error bars or significance tests are provided. Given that the architecture is tuned on the test benchmarks, the claim of 'significant' outperformance is not statistically supported. Please include standard deviations over multiple training runs or per-scene variance, or rerun the final comparison on a held-out split.
minor comments (6)
  1. [Eq. (13)] The output of EPTB is denoted F_SAM in Eq. (13), but it should be F_SAT to match the notation in Eq. (11) and Fig. 3.
  2. [Table I, notes] The 'geometry assembling strategy' denoted by * is never described in the text. Please specify what it is (e.g., test-time self-ensemble) and whether it is used in any qualitative results.
  3. [Tables I and IV] There are small inconsistencies in reported complexity: LFMT ×2 is 2.06M/63.18G in Table I but 2.04M/62.76G in Table IV; ×4 params also differ slightly. Please reconcile the numbers.
  4. [Table I / Table IV / Table VII] Typos: '34.98/00.9602' in Table I; 'EPMT' in Table IV should be 'EPTB'; 'EPIM/EPIT' in Table VII should be 'EPMB/EPTB'; the ×4 SSIM gain over LF-DET is stated as 0.008 in Sec. IV.B.1 but the table shows 0.0008.
  5. [Abstract and Sec. IV.B.6] The limitation statement (0.25 dB below EPIT on STFgantry ×2) is honest, but it is somewhat in tension with the abstract's blanket 'significantly outperforms' phrasing. Please qualify the claim to reflect dataset-specific exceptions.
  6. [Throughout] Minor language issues: 'real-word' should be 'real-world' in the abstract; 'SS-Sub' in Sec. IV.C.1 should likely be 'Sub-SS'; please proofread for consistent terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LFMT's results are empirical benchmark measurements, not derivations that contain their own conclusions.

full rationale

This is an empirical architecture paper. LFMT's output is produced by a concrete network trained with L1 loss on the 144-scene training split and evaluated on the 23-scene test split (Sec. IV.A); the reported PSNR/SSIM gains are measured against independent baselines on standard benchmarks, not derived from the definition of any component. The design choices (Sub-SS, SA-RSMB, EPMB, EPTB) are motivated qualitatively and validated by ablations (Tables III–VIII), but no component is defined in terms of the target metric, and no equation reduces the reported result to its inputs. The self-citations to prior work by Hou/Zeng et al. (refs. [41]–[43]) appear only in the literature review and are not load-bearing. The Sec. IV.B.6 limitation that LFMT is about 0.25 dB below EPIT on STFgantry ×2, and the absence of the cited LFTransMamba [56] baseline from Table I, are legitimate concerns about the strength and completeness of the state-of-the-art claim, but they are comparison/attribution issues, not circularity. Similarly, choosing block counts and fusion features by test-set PSNR (Tables VII/VIII) is a model-selection/overfitting risk, not a self-referential derivation. No circular step can be exhibited, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. It builds on standard SSM/Transformer mechanisms and standard LF subspace assumptions from prior work. The main tunable components are architectural counts chosen directly on the test benchmarks.

free parameters (4)
  • Number of SA-RSMB blocks = 3
    Chosen via ablation on the test benchmarks (Table VII) to maximize average PSNR.
  • Number of EPMB/EPTB blocks = 3
    Chosen via ablation on the test benchmarks (Table VII); the count of EPMB and EPTB is kept equal.
  • Channel dimension C = 64
    Set by design following prior LFSR transformers [29], [32]; not swept.
  • Training schedule = Adam, lr 2e-4 halved every 15 epochs, 90 epochs, batch size 2
    Standard training settings; not central to the architecture claim.
assumptions (3)
  • domain assumption The 4D light field can be decoupled into 2D subspaces (SAI, MacPI, EPI) without information loss for super-resolution.
    Used throughout Section III.C; this is a standard premise in LF literature (e.g., [29], [32]).
  • domain assumption Epipolar-plane line patterns encode disparity and structure that can guide spatial super-resolution.
    Stated in Section II.A and used to justify Stage II. This is standard LF geometry from prior work [32].
  • ad hoc to paper Mamba's selective scan and Transformer's self-attention are complementary for LF features.
    Central design hypothesis of the Stage II dual-branch; supported only by internal ablation Table VI, not by external evidence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Non-Local Spatial-Angular Correlations with a Hybrid Mamba-Transformer Framework for Light Field Super-Resolution." pith.science (2026). https://pith.science/paper/CDHZOXBM

@misc{pith2026250904824,
  author       = {Pith},
  title        = {Pith review of: Exploring Non-Local Spatial-Angular Correlations with a Hybrid Mamba-Transformer Framework for Light Field Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CDHZOXBM}},
  note         = {Machine review of arXiv:2509.04824}
}
read the original abstract

Recently, Mamba-based methods, with its advantage in long-range information modeling and linear complexity, have shown great potential in optimizing both computational cost and performance of light field image super-resolution (LFSR). However, current multi-directional scanning strategies lead to inefficient and redundant feature extraction when applied to complex LF data. To overcome this challenge, we propose a Subspace Simple Scanning (Sub-SS) strategy, based on which we design the Subspace Simple Mamba Block (SSMB) to achieve more efficient and precise feature extraction. Furthermore, we propose a dual-stage modeling strategy to address the limitation of state space in preserving spatial-angular and disparity information, thereby enabling a more comprehensive exploration of non-local spatial-angular correlations. Specifically, in stage I, we introduce the Spatial-Angular Residual Subspace Mamba Block (SA-RSMB) for shallow spatial-angular feature extraction; in stage II, we use a dual-branch parallel structure combining the Epipolar Plane Mamba Block (EPMB) and Epipolar Plane Transformer Block (EPTB) for deep epipolar feature refinement. Building upon meticulously designed modules and strategies, we introduce a hybrid Mamba-Transformer framework, termed LFMT. LFMT integrates the strengths of Mamba and Transformer models for LFSR, enabling comprehensive information exploration across spatial, angular, and epipolar-plane domains. Experimental results demonstrate that LFMT significantly outperforms current state-of-the-art methods in LFSR, achieving substantial improvements in performance while maintaining low computational complexity on both real-word and synthetic LF datasets.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 54 canonical work pages

  1. [56]

    Lftrans- mamba: A hybrid mamba-transformer model for light field image super- resolution,

    K. Jin, Z. Wei, A. Yang, D. Wu, M. Gao, and X. Zhou, “Lftrans- mamba: A hybrid mamba-transformer model for light field image super- resolution,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2025, pp. 1195–1204

  2. [1]

    Light field image processing: An overview,

    G. Wuet al., “Light field image processing: An overview,”IEEE J. Sel. Topics Signal Process, vol. 11, no. 7, pp. 926–954, Oct. 2017

  3. [2]

    Selective light field re- focusing for camera arrays using bokeh rendering and superresolution,

    Y . Wang, J. Yang, Y . Guo, C. Xiao, and W. An, “Selective light field re- focusing for camera arrays using bokeh rendering and superresolution,” IEEE Signal Process. Lett., vol. 26, no. 1, pp. 204–208, Jan. 2019

  4. [3]

    Multi-volumetric refocusing of light fields,

    S. S. Jayaweera, C. U. Edussooriya, C. Wijenayake, P. Agathoklis, and L. T. Bruton, “Multi-volumetric refocusing of light fields,”IEEE Signal Process. Lett., vol. 28, pp. 31–35, Dec. 2020

  5. [4]

    Depth from a light field image with learning-based matching costs,

    H.-G. Jeon, J. Park, G. Choe, J. Park, Y . Bok, Y .-W. Tai, and I. S. Kweon, “Depth from a light field image with learning-based matching costs,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 41, no. 2, pp. 297– 310, Feb. 2019

  6. [5]

    Depth estimation from light field using graph-based structure-aware analysis,

    Y . Zhang, W. Dai, M. Xu, J. Zou, X. Zhang, and H. Xiong, “Depth estimation from light field using graph-based structure-aware analysis,” IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 11, pp. 4269– 4283, Nov. 2020

  7. [6]

    Accurate light field depth esti- mation with superpixel regularization over partially occluded regions,

    J. Chen, J. Hou, Y . Ni, and L.-P. Chau, “Accurate light field depth esti- mation with superpixel regularization over partially occluded regions,” IEEE Trans. Image Process., vol. 27, no. 10, pp. 4889–4900, 2018

  8. [7]

    Occlusion-aware unsupervised learning of depth from 4-d light fields,

    J. Jin and J. Hou, “Occlusion-aware unsupervised learning of depth from 4-d light fields,”IEEE Trans. Image Process., vol. 31, pp. 2216–2228, 2022

Show all 66 references
  1. [8]

    Light field depth estimation via stitched epipolar plane images,

    P. Zhou, L. Shi, X. Liu, J. Jin, Y . Zhang, and J. Hou, “Light field depth estimation via stitched epipolar plane images,”IEEE Trans. Vis. Comput. Graph., vol. 30, no. 10, pp. 6866–6879, 2023

  2. [9]

    Learning signed distance field for multi-view surface reconstruction

    J. Zhang, Y . Yao, and L. Quan, “Learning signed distance field for multi-view surface reconstruction.” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2021, pp. 6505–6514

  3. [10]

    A light-field journey to virtual reality,

    J. Yu, “A light-field journey to virtual reality,”IEEE MultiMedia, vol. 24, no. 2, pp. 104–112, Apr. 2017

  4. [11]

    Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays,

    S. Choi, M. Gopakumar, Y . Peng, J. Kim, and G. Wetzstein, “Neural 3d holography: Learning accurate wave propagation models for 3d holographic virtual and augmented reality displays,”ACM Trans. Graph., vol. 40, no. 6, pp. 1–12, 2021

  5. [12]

    Learning a deep convolutional network for light-field image super-resolution,

    Y . Yoon, H.-G. Jeon, D. Yoo, J.-Y . Lee, and I. So Kweon, “Learning a deep convolutional network for light-field image super-resolution,” in Proc. IEEE Int. Conf. Comput. Vis. Workshops, 2015, pp. 24–32

  6. [13]

    Residual networks for light field image super-resolution,

    S. Zhang, Y . Lin, and H. Sheng, “Residual networks for light field image super-resolution,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2019, pp. 11 046–11 055

  7. [14]

    Light field spatial super-resolution using deep efficient spatial-angular separable convolution,

    H. W. F. Yeung, J. Hou, X. Chen, J. Chen, Z. Chen, and Y . Y . Chung, “Light field spatial super-resolution using deep efficient spatial-angular separable convolution,”IEEE Trans. Image Process., vol. 28, no. 5, pp. 2319–2330, May 2019

  8. [15]

    Light field spatial super- resolution via deep combinatorial geometry embedding and structural consistency regularization,

    J. Jin, J. Hou, J. Chen, and S. Kwong, “Light field spatial super- resolution via deep combinatorial geometry embedding and structural consistency regularization,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2020, pp. 2260–2269

  9. [16]

    End-to-end light field spatial super- resolution network using multiple epipolar geometry,

    S. Zhang, S. Chang, and Y . Lin, “End-to-end light field spatial super- resolution network using multiple epipolar geometry,”IEEE Trans. Image Process., vol. 30, pp. 5956–5968, 2021

  10. [17]

    Light field super-resolution by jointly exploiting internal and external similarities,

    Z. Cheng, Z. Xiong, and D. Liu, “Light field super-resolution by jointly exploiting internal and external similarities,”IEEE Trans. Circuits Syst. Video Technol., vol. 30, no. 8, pp. 2604–2616, Aug. 2019. 14

  11. [18]

    Light field image super-resolution using deformable convolution,

    Y . Wang, J. Yang, L. Wang, X. Ying, T. Wu, W. An, and Y . Guo, “Light field image super-resolution using deformable convolution,”IEEE Trans. Image Process., vol. 30, pp. 1057–1071, Dec. 2020

  12. [19]

    Light field reconstruction via deep adaptive fusion of hybrid lenses,

    J. Jin, M. Guo, J. Hou, H. Liu, and H. Xiong, “Light field reconstruction via deep adaptive fusion of hybrid lenses,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 10, pp. 12 050–12 067, 2023

  13. [20]

    Intra-inter view interaction network for light field image super-resolution,

    G. Liu, H. Yue, J. Wu, and J. Yang, “Intra-inter view interaction network for light field image super-resolution,”IEEE Trans. Multimedia, vol. 25, pp. 256–266, Nov. 2021

  14. [21]

    Light field super-resolution using decoupled selective matching,

    Y . Liu, Z. Cheng, Z. Xiao, and Z. Xiong, “Light field super-resolution using decoupled selective matching,”IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 5, pp. 3313–3326, May 2023

  15. [22]

    Un- elf: Unconstrained neural light field for self-supervised angular super- resolution,

    M. Zhao, H. Sheng, R. Chen, R. Cong, Z. Cui, and D. Yang, “Un- elf: Unconstrained neural light field for self-supervised angular super- resolution,”IEEE Trans. Circuits Syst. Video Technol., 2025

  16. [23]

    Deep spatial-angular regularization for light field imaging, denoising, and super-resolution,

    M. Guo, J. Hou, J. Jin, J. Chen, and L.-P. Chau, “Deep spatial-angular regularization for light field imaging, denoising, and super-resolution,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 10, pp. 6094–6110, 2021

  17. [24]

    Probabilistic-based feature embedding of 4-d light fields for compressive imaging and denoising,

    X. Lyu and J. Hou, “Probabilistic-based feature embedding of 4-d light fields for compressive imaging and denoising,”Int. J. Comput. Vis., vol. 132, no. 6, pp. 2255–2275, 2024

  18. [25]

    Light field super-resolution via adaptive feature remixing,

    K. Ko, Y . J. Koh, S. Chang, and C.-S. Kim, “Light field super-resolution via adaptive feature remixing,”IEEE Trans. Image Process., vol. 30, pp. 4114–4128, 2021

  19. [26]

    Spatial-angular versatile convolution for light field reconstruction,

    Z. Cheng, Y . Liu, and Z. Xiong, “Spatial-angular versatile convolution for light field reconstruction,”IEEE Trans. Comput. Imag., vol. 8, pp. 1131–1144, 2022

  20. [27]

    Dense dual-attention network for light field image super-resolution,

    Y . Mo, Y . Wang, C. Xiao, J. Yang, and W. An, “Dense dual-attention network for light field image super-resolution,”IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 7, pp. 4431–4443, Jul. 2022

  21. [28]

    Light field image super- resolution network via joint spatial-angular and epipolar information,

    V . Van Duong, T. N. Huu, J. Yim, and B. Jeon, “Light field image super- resolution network via joint spatial-angular and epipolar information,” IEEE Trans. Comput. Imag., vol. 9, pp. 350–366, 2023

  22. [29]

    Light field image super-resolution with transformers,

    Z. Liang, Y . Wang, L. Wang, J. Yang, and S. Zhou, “Light field image super-resolution with transformers,”IEEE Signal Process. Lett., vol. 29, pp. 563–567, 2022

  23. [30]

    Detail-preserving transformer for light field image super-resolution,

    S. Wang, T. Zhou, Y . Lu, and H. Di, “Detail-preserving transformer for light field image super-resolution,” inProc. AAAI Conf. Artif. Intell., 2022, pp. 1–9

  24. [31]

    Exploiting spatial and angular correlations with deep efficient transformers for light field image super-resolution,

    R. Cong, H. Sheng, D. Yang, Z. Cui, and R. Chen, “Exploiting spatial and angular correlations with deep efficient transformers for light field image super-resolution,”IEEE Trans. Multimedia, vol. 26, pp. 1421– 1435, Jun. 2023

  25. [32]

    Learn- ing non-local spatial-angular correlation for light field image super- resolution,

    Z. Liang, Y . Wang, L. Wang, J. Yang, S. Zhou, and Y . Guo, “Learn- ing non-local spatial-angular correlation for light field image super- resolution,” inProc. IEEE Int. Conf. Comput. Vis. (ICCV), 2023, pp. 12 376–12 386

  26. [33]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv:2312.00752, 2023

  27. [34]

    Mamba-based light field super- resolution with efficient subspace scanning,

    R. Gao, Z. Xiao, and Z. Xiong, “Mamba-based light field super- resolution with efficient subspace scanning,” inProc. Asian Conf. Comput. Vis. (ACCV), 2024, pp. 531–547

  28. [35]

    Lfmamba: Light field image super-resolution with state space model,

    W. Xiaet al., “Lfmamba: Light field image super-resolution with state space model,”arXiv:2406.12463, 2024

  29. [36]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,”arXiv:2401.09417, 2024

  30. [37]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” inNeurIPS, 2024

  31. [38]

    Light field rendering,

    M. Levoy and P. Hanrahan, “Light field rendering,” inSeminal Graphics Papers: Pushing the Boundaries, Volume 2, 2023, pp. 441–452

  32. [39]

    Spatial-angular interaction for light field image super-resolution,

    Y . Wang, L. Wang, J. Yang, W. An, J. Yu, and Y . Guo, “Spatial-angular interaction for light field image super-resolution,” inProc. Eur. Conf. Comput. Vis.Cham, Switzerland: Springer, 2020, pp. 290–308

  33. [40]

    Disentangling light fields for super-resolution and disparity estimation,

    Y . Wang, L. Wang, G. Wu, J. Yang, W. An, J. Yu, and Y . Guo, “Disentangling light fields for super-resolution and disparity estimation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 1, pp. 425–443, Jan. 2023

  34. [41]

    Content-aware warping for view synthesis,

    M. Guo, J. Hou, J. Jin, H. Liu, H. Zeng, and J. Lu, “Content-aware warping for view synthesis,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 8, pp. 9486–9503, 2023

  35. [42]

    Learning spatial- angular fusion for compressive light field imaging in a cycle-consistent framework,

    X. Lyu, Z. Zhu, M. Guo, J. Jin, J. Hou, and H. Zeng, “Learning spatial- angular fusion for compressive light field imaging in a cycle-consistent framework,” inProc. ACM Int. Conf. Multimedia (ACM MM), 2021, pp. 4613–4621

  36. [43]

    Deep coarse-to- fine dense light field reconstruction with flexible sampling and geometry- aware fusion,

    J. Jin, J. Hou, J. Chen, H. Zeng, S. Kwong, and J. Yu, “Deep coarse-to- fine dense light field reconstruction with flexible sampling and geometry- aware fusion,”IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), vol. 44, no. 4, pp. 1819–1836, 2020

  37. [44]

    Incorporating degradation estimation in light field spatial super-resolution,

    Z. Xiao and Z. Xiong, “Incorporating degradation estimation in light field spatial super-resolution,”Comput. Vis. Image Underst. (CVIU), vol. 252, p. 104295, 2025

  38. [45]

    Occlusion-embedded hybrid transformer for light field super-resolution,

    Z. Xiao, Z. Li, and W. Jia, “Occlusion-embedded hybrid transformer for light field super-resolution,” inProc. AAAI Conf. Artif. Intell. (AAAI), vol. 39, no. 8, 2025, pp. 8700–8708

  39. [46]

    Cutmib: Boosting light field super-resolution via multi-view image blending,

    Z. Xiao, Y . Liu, R. Gao, and Z. Xiong, “Cutmib: Boosting light field super-resolution via multi-view image blending,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 1672–1682

  40. [47]

    A new approach to linear filtering and prediction problems,

    R. E. Kalman, “A new approach to linear filtering and prediction problems,”J. Basic Eng., vol. 82, no. 1, Mar. 1960

  41. [48]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,”arXiv:2111.00396, 2021

  42. [49]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. W. Linderman, “Simplified state space layers for sequence modeling,”arXiv:2208.04933, 2022

  43. [50]

    Hungry hungry hippos: Towards language modeling with state space models,

    D. Y . Fu, T. Dao, K. K. Saab, A. W. Thomas, A. Rudra, and C. R ´e, “Hungry hungry hippos: Towards language modeling with state space models,”arXiv:2212.14052, 2022

  44. [51]

    Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,

    Z. Xing, T. Ye, Y . Yang, G. Liu, and L. Zhu, “Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 578–588

  45. [52]

    Videomamba: State space model for efficient video understanding,

    K. Li, X. Li, Y . Wang, Y . He, Y . Wang, L. Wang, and Y . Qiao, “Videomamba: State space model for efficient video understanding,” in Proc. Eur. Conf. Comput. Vis. (ECCV). Springer, 2024, pp. 237–255

  46. [53]

    Video mamba suite: State space model as a versatile alternative for video understanding,

    G. Chen, Y . Huang, J. Xu, B. Pei, Z. Chen, Z. Li, J. Wang, K. Li, T. Lu, and L. Wang, “Video mamba suite: State space model as a versatile alternative for video understanding,”arXiv:2403.09626, 2024

  47. [54]

    Pointmamba: A simple state space model for point cloud analysis,

    D. Liang, X. Zhou, W. Xu, X. Zhu, Z. Zou, X. Ye, X. Tan, and X. Bai, “Pointmamba: A simple state space model for point cloud analysis,” arXiv:2402.10739, 2024

  48. [55]

    Mamba3d: Enhancing local fea- tures for 3d point cloud analysis via state space model,

    X. Han, Y . Tang, Z. Wang, and X. Li, “Mamba3d: Enhancing local fea- tures for 3d point cloud analysis via state space model,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 4995–5004

  49. [57]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” inProc. Eur. Conf. Comput. Vis. (ECCV). Springer, 2024, pp. 222–241

  50. [58]

    Enhanced deep residual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee, “Enhanced deep residual networks for single image super-resolution,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops, 2017, pp. 136–144

  51. [59]

    Image super- resolution using very deep residual channel attention networks,

    Y . Zhang, K. Li, K. Li, L. Wang, B. Zhong, and Y . Fu, “Image super- resolution using very deep residual channel attention networks,” inProc. Eur. Conf. Comput. Vis. (ECCV), 2018, pp. 286–301

  52. [60]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”ICLR, 2021

  53. [61]

    New light field image dataset,

    M. Rerabek and T. Ebrahimi, “New light field image dataset,” inProc. 8th Int. Conf. Qual. Multimedia Exp., 2016, pp. 1–2

  54. [62]

    Datasets and benchmarks for densely sampled 4d light fields

    S. Wanner, S. Meister, and B. Goldluecke, “Datasets and benchmarks for densely sampled 4d light fields.” inVMV, vol. 13, 2013, pp. 225–226

  55. [63]

    A dataset and evaluation methodology for depth estimation on 4d light fields,

    K. Honauer, O. Johannsen, D. Kondermann, and B. Goldluecke, “A dataset and evaluation methodology for depth estimation on 4d light fields,” inProc. Asian Conf. Comput. Vis. (ACCV). Springer, 2017, pp. 19–34

  56. [64]

    Light field inpainting prop- agation via low rank matrix completion,

    M. Le Pendu, X. Jiang, and C. Guillemot, “Light field inpainting prop- agation via low rank matrix completion,”IEEE Trans. Image Process., vol. 27, no. 4, pp. 1981–1993, Apr. 2018

  57. [65]

    The (new) stanford light field archive,

    V . Vaish and A. Adams, “The (new) stanford light field archive,” Comput. Graph. Lab., Stanford Univ., vol. 6, no. 7, 2008

  58. [66]

    Robust depth estimation for light field via spinning parallelogram operator,

    S. Zhang, H. Sheng, C. Li, J. Zhang, and Z. Xiong, “Robust depth estimation for light field via spinning parallelogram operator,”Comput. Vis. Image Understand., vol. 145, pp. 148–159, Apr. 2016. 15 Haosong Liureceived the B.E. degree from the School of Information Science and ...

Pith tools

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