Pith. sign in

REVIEW 4 major objections 6 minor 61 references

A new dynamic MEF benchmark and a Householder projection network claim to remove motion ghosts while restoring high-dynamic-range detail.

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

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

A large-scale dynamic MEF benchmark with expert GTs plus a Householder-based network that decouples illumination pre-alignment from orthogonal ghost projection.

T0 review reviewed 2026-07-12 challenge →

load-bearing objection Solid MEF package: a genuinely larger motion-centric benchmark plus a clean geometric attention twist that beats recent SOTA on the usual tables; expert GT subjectivity is the real soft spot but does not sink the work. the 4 major comments →

arxiv 2607.03110 v1 pith:SXUY6ZEX submitted 2026-07-03 cs.CV

ExpoMotion: A Large-Scale Benchmark and A Householder Projection Network for Multi-Exposure Fusion

Yao Liu , Lishen Qu , Shihao Zhou , Jie Liang , Hui Zeng , Yabin Peng , Huipeng Lin , Lei Zhang
show 1 more author
Jufeng Yang
This is my paper
classification cs.CV
keywords Multi-Exposure FusionDynamic RangeDeghostingHouseholder transformationHDR imagingattention mechanismsbenchmark dataset
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Multi-exposure fusion can produce display-ready high-dynamic-range images, yet real camera and object motion create ghosts, and existing datasets either lack motion or lack trustworthy ground truth. This paper releases ExpoMotion, a 1,738-sequence collection spanning controlled, laboratory, and real-world motion with expert-refined labels, so that deghosting and perceptual quality can be measured together. It then introduces the Householder Orthogonal Projection network, which first harmonizes brightness with global statistics and then treats residual ghosts as vectors that can be reflected out of the feature space by a learned Householder operator. The authors show that models trained on the new data generalize better on both classic and in-the-wild test sets, and that their network recovers clearer detail with fewer artifacts than recent fusion and HDR transformers.

Core claim

The paper claims that multi-exposure fusion under motion is best solved by decoupling the problem into exposure pre-alignment followed by geometric ghost filtering: global illumination statistics first match auxiliary frames to the reference exposure, after which a dynamic Householder reflector projects misalignment artifacts orthogonally out of the feature manifold while leaving high-frequency detail intact. Together with the ExpoMotion benchmark that supplies reliable dynamic ground truth, this design yields higher full-reference scores and better no-reference perceptual quality than prior methods.

What carries the argument

Householder Orthogonal Attention (HOA): a learned Householder reflector that models ghosts as orthogonal perturbations and subtracts their projection from auxiliary features after global-prior illumination alignment.

Load-bearing premise

The expert-guided pipeline that turns algorithm candidates plus photographer tone-curve edits into ground truth is assumed to give unbiased labels that correctly capture both ghost-free fusion and human preference.

What would settle it

Train the same architecture on ExpoMotion versus a purely automatic or purely synthetic GT set of equal size and measure whether the claimed PSNR/SSIM and no-reference gains on held-out real-motion sequences disappear or reverse.

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

If this is right

  • Models trained on ExpoMotion should transfer more reliably to real handheld multi-exposure stacks than models trained only on static or tone-mapped HDR sets.
  • Treating ghost removal as geometric projection rather than similarity attention reduces residual tearing around moving objects and saturated light sources.
  • The three-frame HOP design can serve as a practical, relatively lightweight alternative to larger restoration transformers for on-device MEF.
  • Future variable-frame MEF systems can reuse the same exposure-pre-alignment plus orthogonal-filtering split once the architecture is generalized beyond three inputs.

Where Pith is reading between the lines

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

  • The same Householder rejection idea could be tried for other multi-frame restoration tasks where one wants complementary information yet must discard non-corresponding content, such as multi-frame denoising or burst deblurring.
  • If the expert GT process itself embeds photographer taste, perceptual metrics may over-reward methods that match that taste rather than purely physical radiance consistency.
  • A follow-up that releases the raw candidate fusion pool before expert selection would let the community quantify how much the human refinement step moves the benchmark.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces ExpoMotion, a large-scale multi-exposure fusion (MEF) benchmark of 1,738 sequences (10,909 images) spanning controlled, laboratory, and real-world motion under extreme exposure ranges, with ground truths produced by multi-algorithm candidate generation followed by expert selection and manual tone/color refinement. To address deghosting under large brightness gaps, it proposes HOP, which decouples the problem into Global Priors Illumination Alignment (GPIA) for exposure pre-alignment and Householder Orthogonal Attention (HOA/HOPU) that models ghosts as difference-derived orthogonal perturbations and rejects them via a learnable Householder-like reflector, plus a Gradient-Guided FFN with a fixed Laplacian branch. Experiments report state-of-the-art full-reference PSNR/SSIM on Kalantari17, RealHDRV, and a controlled-motion split, improved no-reference perceptual scores under cross-dataset training, and ablations supporting each module; code and data are promised publicly.

Significance. If the claims hold, ExpoMotion would be a useful community resource: it is substantially larger and more motion-centric than prior real MEF/HDR sets, and the no-reference cross-dataset results (Table 2) give independent evidence that training on it improves generalization beyond the authors’ own GTs. The HOP design is a concrete, efficient alternative to flow/attention deghosting, with a clear modular ablation (Table 3) and favorable accuracy–parameter trade-offs versus Restormer and related transformers. Public release of dataset and code would make both contributions falsifiable and reusable. The main significance risk is that full-reference superiority and the “high-fidelity GT” claim rest on an unquantified expert pipeline; that does not erase the dataset scale or the method’s empirical gains, but it bounds how strongly the benchmark can currently be treated as an objective deghosting oracle.

major comments (4)
  1. Section 3 (Ground Truth Generation) is load-bearing for every full-reference claim and for the assertion that ExpoMotion is a reliable deghosting benchmark, yet no inter-rater agreement, residual-ghost audit, or comparison against independent deghosting oracles is reported. Candidate fusion, rater selection, and five experts’ manual tone-curve/color edits define the supervision that HOP is trained and scored against (Table 1 controlled-motion split). Without agreement statistics or a held-out audit of residual motion artifacts in the refined GTs, measured PSNR/SSIM gains could partly reflect stylistic agreement with the expert panel rather than objective ghost removal. Please quantify reliability (e.g., pairwise expert agreement on a subset, or side-by-side residual maps) and discuss how expert preference may couple to the training objective.
  2. Section 3 (Data Filtering) states that more than 80% of 14,235 captured sequences were discarded for blur, noise, handshake, or any intra-bracket motion that breaks the static-scene assumption used for GT synthesis. This filtering concentrates the retained distribution on scenes for which clean expert GTs are easy to produce and may under-represent the hard real-world motion regimes the abstract and Fig. 2 claim to cover. The controlled protocols (multi-view/multi-pose static bracketing and static–dynamic swap) further limit natural motion blur and occlusion. Please report the retained fraction by acquisition type (controlled / lab / real), characterize discarded failure modes, and either add a harder real-motion full-reference subset or clearly scope the benchmark’s coverage claims.
  3. Section 4.2 and Eqs. (4)–(7) rest on the modeling assumption that motion ghosts are approximately additive components along a difference-derived conflict direction v that can be orthogonally rejected from the clean feature manifold. This is a strong geometric prior; the paper offers intuition and ablations (Table 3, HOPU rows) but no diagnostic that v aligns with actual misalignment (e.g., correlation of |v| with optical-flow magnitude, or failure cases under large occlusion). Because HOA is the central methodological claim, please add at least a qualitative/quantitative check that the rejected component is ghost-like rather than high-frequency detail, and discuss when the orthogonality assumption breaks.
  4. Introduction criticizes tone-mapping HDR reconstruction sets as MEF supervision for contrast distortion, yet §5.1 trains on Kalantari17 and RealHDRV after Photomatix tone mapping and evaluates full-reference on those same tone-mapped targets (Table 1). This weakens the narrative contrast with prior practice and confounds whether gains on those sets come from better deghosting or better matching of Photomatix aesthetics. Please either justify the protocol consistently, report a pure ExpoMotion-only training ablation on the external sets, or separate “tone-mapped HDR” evaluation from the MEF-native ExpoMotion claims.
minor comments (6)
  1. Abstract and early pages contain concatenated words (e.g., “capableofcomprehensivelyevaluating”, “theextremebrightnessdisparities”); clean spacing and line-break artifacts throughout.
  2. Fig. 3 / Fig. 5 notation: F_SA, F_Aux, and F_Ref are not always defined before use; a short symbol table would help.
  3. Eq. (2) uses Up(·) without stating the upsampling operator; specify bilinear vs. nearest and whether R is broadcast per-channel.
  4. Table 1 reports single-run PSNR/SSIM with no seed variance or error bars; even a two-seed range would strengthen SOTA claims over Restormer (~0.5 dB).
  5. Limitation section notes the fixed three-input design; a short experiment with 5/7-frame stacks (even a simple multi-pass baseline) would better match the dataset’s exposure-length diversity advertised in Fig. 2.
  6. Related Work could more clearly separate MEF-native dynamic benchmarks from HDR deghosting sets to avoid conflating evaluation protocols.

Circularity Check

0 steps flagged

Empirical MEF paper with independent held-out and external benchmarks; expert GT is a soft assumption, not a definitional circularity.

full rationale

The paper is a standard empirical computer-vision contribution: a new dynamic MEF dataset (ExpoMotion) plus a network (HOP) that applies a classical Householder reflector to learned features. Full-reference claims are measured on held-out controlled-motion splits and on external datasets (Kalantari17, RealHDRV); no-reference claims use Sen/Tursen and a real-motion split. The Householder operator (Eqs. 3–7) is a standard linear-algebra construction whose parameters (conflict vector v, gate M, scale α) are learned from data, not defined in terms of the target PSNR/SSIM or perceptual scores. Ablations (Table 3) and cross-dataset generalization (Table 2) further separate architecture from training distribution. The only soft spot is the expert-guided GT pipeline (Section 3), which could inject stylistic bias into full-reference numbers; that is an assumption-quality concern, not a reduction of a claimed prediction to its own inputs by construction. No self-citation uniqueness theorems, no fitted-parameter-as-prediction, and no renaming of known results appear. Score 1 reflects the minor residual risk that expert GTs partly shape the very perceptual target the model is trained to match, without elevating it to circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

Standard deep-learning and image-processing assumptions plus a small set of architectural free parameters and modeling choices. No new physical entities; the main invented constructs are the modular network components and the geometric interpretation of ghosts as orthogonal perturbations.

free parameters (4)
  • alpha (Householder scaling vector) = learned (init 0)
    Learnable per-channel intensity of the Householder rejection; initialized to zero and optimized end-to-end (Eq. 4).
  • beta (GGFN Laplacian mix) = learned (init 0)
    Learnable scalar mixing the fixed Laplacian high-pass branch into the FFN (Eq. 9); init 0.
  • N1 (number of HHTB blocks) = 2 or 3
    Architectural depth hyper-parameter defining HOP-S (N1=2) vs HOP-B (N1=3).
  • GPIA pooling kernel and epsilon = 16x16, epsilon small positive
    16x16 average pool and additive epsilon for illumination ratio (Eqs. 1-2); chosen by design.
axioms (4)
  • domain assumption Illumination is predominantly low-frequency and can be approximated by global average-pooled statistics (Retinex-inspired).
    Invoked to justify GPIA pre-alignment (Section 4.1).
  • ad hoc to paper Motion-induced ghost artifacts can be modeled as additive components lying along a difference-derived conflict direction that is approximately orthogonal to the clean feature manifold.
    Core modeling choice of HOA / HOPU (Section 4.2, Geometric Intuition).
  • domain assumption Expert photographers can produce ground-truth fused images that simultaneously eliminate ghosts and match human perceptual preference better than any automatic method.
    Underpins the entire GT pipeline and all full-reference evaluation (Section 3).
  • standard math Standard Householder reflection H = I - 2vv^T (or its scaled version) is a valid geometric operator for feature-space rejection inside a neural network.
    Linear-algebra identity used in Eqs. 3-4 and 7.
invented entities (3)
  • Householder Orthogonal Attention (HOA) / HOPU no independent evidence
    purpose: Project motion ghosts out of auxiliary features via a dynamic Householder reflector built from the feature difference.
    New architectural module; independent evidence is only the empirical ablation and SOTA numbers inside this paper.
  • Global Priors Illumination Alignment (GPIA) no independent evidence
    purpose: Pre-align auxiliary exposures to the reference using global pooled statistics before attention.
    New module; evidence is internal ablation only.
  • Gradient-Guided FFN (GGFN) with fixed Laplacian branch no independent evidence
    purpose: Inject explicit high-frequency edge prior into the feed-forward path.
    New module combining learned DWConv with hard-coded Laplacian kernel; evidence internal.

reviewed 2026-07-12 · how reviews work

0 comments
Cite this review

Pith. "Pith review of ExpoMotion: A Large-Scale Benchmark and A Householder Projection Network for Multi-Exposure Fusion." pith.science (2026). https://pith.science/paper/SXUY6ZEX

@misc{pith2026260703110,
  author       = {Pith},
  title        = {Pith review of: ExpoMotion: A Large-Scale Benchmark and A Householder Projection Network for Multi-Exposure Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SXUY6ZEX}},
  note         = {Machine review of arXiv:2607.03110}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-Exposure Fusion (MEF) effectively extends dynamic range, but practical deployment is hindered by motion-induced ghosting and the scarcity of high-quality dynamic benchmarks. Current benchmarks largely neglect dynamic scenes and lack reliable ground truth, making it difficult to handle the complexity of real-world motions. In response, we introduce ExpoMotion, a large-scale benchmark designed to evaluate deghosting capabilities. Comprising 1,738 sequences and 10,909 images across diverse environments, it covers a wide range of motions and provides high-fidelity GTs constructed through an expert-guided acquisition pipeline. To tackle the complex dynamics and extreme conditions captured in this benchmark, we propose the Householder Orthogonal Projection network (HOP), which revisits MEF deghosting from a mathematical perspective via Householder transformation, decoupling multi-frame alignment into exposure pre-alignment and ghost filtering. Specifically, the Global Priors Illumination Alignment (GPIA) module first rectifies drastic dynamic range discrepancies by utilizing global statistics for exposure harmonization. Regarding ghost removal, our Householder Orthogonal Attention (HOA) models artifacts as orthogonal perturbations. By employing a dynamic Householder reflector, HOA effectively projects ghosts out of the feature manifold while preserving high-frequency details. Experiments demonstrate that our ExpoMotion dataset enables superior generalization and artifact-free detail restoration, while also validating the effectiveness and efficiency of the HOP method. The dataset and code are available at https://github.com/Leo-LiuYao/ExpoMotion.

Figures

Figures reproduced from arXiv: 2607.03110 by Huipeng Lin, Hui Zeng, Jie Liang, Jufeng Yang, Lei Zhang, Lishen Qu, Shihao Zhou, Yabin Peng, Yao Liu.

Figure 1
Figure 1. Figure 1: Overview of the proposed large-scale MEF dataset. Deviating from the conven￾tional focus on perceptual quality in static scenes, our work introduces a large-scale MEF dataset dedicated to simultaneous detail restoration and motion artifact sup￾pression. We integrate data from controlled and real-world motion, extreme lighting environments, and laboratory setups. Abstract. Multi-Exposure Fusion (MEF) effect… view at source ↗
Figure 2
Figure 2. Figure 2: Statistics and diversity of the collected ExpoMotion dataset. (a) Quantitative comparison showing our dataset is the largest among existing real-world MEF and HDR benchmarks. (b) Representative samples of the four data acquisition types. (c) T-SNE visualization showing superior feature diversity compared to related datasets. (d) Luminance range distribution. (e) A radar chart highlighting our comprehensive… view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline of the proposed Householder Orthogonal Projection Network (HOP). MEF is decomposed into exposure pre-alignment and ghost filtering stages. all while producing a perceptually pleasing result free of motion-induced ghosting artifacts. To address the issues of inconsistent brightness and motion artifacts across the three input images, we propose the Householder Orthogonal Projection Net￾work (HOP). W… view at source ↗
Figure 4
Figure 4. Figure 4: Geometric intuition of the Householder transformation. Geometric Intuition of Householder Re￾jection. Formally, a Householder matrix H ∈ R d×d is a fundamental linear algebra operator that performs an orthogonal reflection of a vec￾tor across a hyperplane defined by a unit normal vector v. This reflection can be expressed as: \mathbf {H} = \mathbf {I} - 2\mathbf {v}\mathbf {v}^T (3) where I is the identity… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of HHTB and GPIA. Specifically, HHTB is composed of two core components: Householder Orthogonal Attention (HOA) and Gradient-Guided Feed￾Forward Network (GGFN). This acts as a strong prior for ghost localization: \mathbf {v}_{raw} = \mathcal {G}_{dw}(\mathbf {D}), \quad \mathbf {v} = \frac {\mathbf {v}_{raw}}{||\mathbf {v}_{raw}||_2 + \epsilon } (5) where Gdw denotes a parameter-efficient dept… view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparisons on Kalantari17 [14] dataset and ours. Zoom in for a better view. 5.2 Comparison with Previous Work Compared methods. We compare our approach with representative CNN￾based methods, such as SAFNet [16] and AFUNet [20], and recent Transformer￾based restoration architectures, including SCTNet [42], Restormer [56], HDR￾Trans [24], and ASTv2 [61]. All methods are evaluated under the same proto… view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of cross-dataset generalization. The indices of the sub-figures are consistent with those in Tab. 2. left sub-figure, competing models struggle to distinguish foreground structures from moving backgrounds, resulting in severe ghosting artifacts and unnatural distortions around the lamp pole. Furthermore, in the challenging night scene on the right, while other models exhibit severe noise … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

61 extracted references · 1 linked inside Pith

  1. [1]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Bai, H., Zhang, J., Zhao, Z., Deng, L., Cui, Y., Xu, S.: Retinex-mef: Retinex- based glare effects aware unsupervised multi-exposure image fusion. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 7251–7261 (2025)

    2025

  2. [2]

    IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV)

    Barua, H.B., Stefanov, K., Wong, K., Dhall, A., Krishnasamy, G.: Gta-hdr: A large- scale synthetic dataset for hdr image reconstruction. IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV). pp. 7876–7886 (2025)

    2025

  3. [3]

    Computers & Graphics39, 12–23 (2014)

    Bruce, N.D.: Expoblend: Information preserving exposure blending based on nor- malized log-domain entropy. Computers & Graphics39, 12–23 (2014)

    2014

  4. [4]

    IEEE Transactions on Image Processing (TIP).27(4), 2049–2062 (2018)

    Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing (TIP).27(4), 2049–2062 (2018)

    2049

  5. [5]

    AAAI Conference on Artificial Intelligence (AAAI).35(2), 973–981 (2021) 16 Y

    Chan, K.C., Wang, X., Yu, K., Dong, C., Loy, C.C.: Understanding deformable alignment in video super-resolution. AAAI Conference on Artificial Intelligence (AAAI).35(2), 973–981 (2021) 16 Y. Liu et al

    2021

  6. [6]

    IEEE Transactions on Image Processing (TIP).31, 2661–2672 (2022)

    Chen, J., Yang, Z., Chan, T.N., Li, H., Hou, J., Chau, L.P.: Attention-guided progressive neural texture fusion for high dynamic range image restoration. IEEE Transactions on Image Processing (TIP).31, 2661–2672 (2022)

    2022

  7. [7]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Chen, Z., Wang, Y., Cai, X., You, Z., Lu, Z., Zhang, F., Guo, S., Xue, T.: Ultra- fusion: Ultra high dynamic imaging using exposure fusion. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). (2025)

    2025

  8. [8]

    Debevec, P.E., Malik, J.: Recovering high dynamic range radiance maps from pho- tographs.ACMSIGGRAPHAnnualConference(SIGGRAPH).pp.369–378(1997)

    1997

  9. [9]

    IEEE Transactions on Image Processing (TIP).29, 1127–1138 (2019)

    Fang, Y., Zhu, H., Ma, K., Wang, Z., Li, S.: Perceptual evaluation for multi- exposure image fusion of dynamic scenes. IEEE Transactions on Image Processing (TIP).29, 1127–1138 (2019)

    2019

  10. [10]

    Vision, Modeling and Visualization, RWTH Aachen

    Grosch, T., et al.: Fast and robust high dynamic range image generation with camera and object movement. Vision, Modeling and Visualization, RWTH Aachen. pp. 277–284 (2006)

    2006

  11. [11]

    Asian Conference on Computer Vision pp

    Heo, Y.S., Lee, K.M., Lee, S.U., Moon, Y., Cha, J.: Ghost-free high dynamic range imaging. Asian Conference on Computer Vision pp. 486–500 (2010)

    2010

  12. [12]

    Hu, J., Gallo, O., Pulli, K., Sun, X.: Hdr deghosting: How to deal with saturation? IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1163–1170 (2013)

    2013

  13. [13]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Jiang, T., Wang, C., Li, X., Li, R., Fan, H., Liu, S.: Meflut: Unsupervised 1d lookup tables for multi-exposure image fusion. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10542–10551 (2023)

    2023

  14. [14]

    ACM Transactions on Graphics (TOG).36(4) (2017)

    Kalantari, N.K., Ramamoorthi, R.: Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (TOG).36(4) (2017)

    2017

  15. [15]

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image qual- itytransformer.IEEE/CVFInternationalConferenceonComputerVision(ICCV). pp. 5148–5157 (2021)

    2021

  16. [16]

    European Conference on Computer Vision (ECCV)

    Kong, L., Li, B., Xiong, Y., Zhang, H., Gu, H., Chen, J.: Safnet: Selective alignment fusion network for efficient hdr imaging. European Conference on Computer Vision (ECCV). (2024)

    2024

  17. [17]

    IEEE International Conference on Multimedia and Expo (ICME)

    Kou,F.,Li,Z.,Wen,C.,Chen,W.:Multi-scaleexposurefusionviagradientdomain guided image filtering. IEEE International Conference on Multimedia and Expo (ICME). pp. 1105–1110 (2017)

    2017

  18. [18]

    IEEE Transactions on Image Processing (TIP)

    Li,H.,Ma,K.,Yong,H.,Zhang,L.:Fastmulti-scalestructuralpatchdecomposition for multi-exposure image fusion. IEEE Transactions on Image Processing (TIP). 29, 5805–5816 (2020)

    2020

  19. [19]

    IEEE Transactions on Image Processing (TIP).22(7), 2864–2875 (2013)

    Li, S., Kang, X., Hu, J.: Image fusion with guided filtering. IEEE Transactions on Image Processing (TIP).22(7), 2864–2875 (2013)

    2013

  20. [20]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Li, X., Ni, Z., Yang, W.: Afunet: Cross-iterative alignment-fusion synergy for hdr reconstruction via deep unfolding paradigm. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10666–10675 (2025)

    2025

  21. [21]

    IEEE Transactions on Image Processing (TIP).23(10), 4372–4382 (2014)

    Li, Z., Zheng, J., Zhu, Z., Wu, S.: Selectively detail-enhanced fusion of differ- ently exposed images with moving objects. IEEE Transactions on Image Processing (TIP).23(10), 4372–4382 (2014)

    2014

  22. [22]

    AAAI Conference on Artificial Intelligence (AAAI).37(2), 1710–1718 (2023)

    Liu, R., Li, C., Cao, H., Zheng, Y., Zeng, M., Cheng, X.: Emef: Ensemble multi- exposure image fusion. AAAI Conference on Artificial Intelligence (AAAI).37(2), 1710–1718 (2023)

    2023

  23. [23]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, S., Zhang, X., Sun, L., Liang, Z., Zeng, H., Zhang, L.: Joint hdr denoising and fusion: A real-world mobile hdr image dataset. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 13966–13975 (2023) ExpoMotion 17

    2023

  24. [24]

    European Conference on Computer Vision (ECCV)

    Liu, Z., Wang, Y., Zeng, B., Liu, S.: Ghost-free high dynamic range imaging with context-aware transformer. European Conference on Computer Vision (ECCV). pp. 344–360 (2022)

    2022

  25. [25]

    IEEE Transactions on Image Processing (TIP).32, 1529–1540 (2023)

    Luo, J., Ren, W., Gao, X., Cao, X.: Multi-exposure image fusion via deformable self-attention. IEEE Transactions on Image Processing (TIP).32, 1529–1540 (2023)

    2023

  26. [26]

    IEEE Transactions on Image Processing (TIP).29, 2808–2819 (2019)

    Ma, K., Duanmu, Z., Zhu, H., Fang, Y., Wang, Z.: Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing (TIP).29, 2808–2819 (2019)

    2019

  27. [27]

    IEEE Transactions on Image Processing (TIP).26(5), 2519–2532 (2017)

    Ma, K., Li, H., Yong, H., Wang, Z., Meng, D., Zhang, L.: Robust multi-exposure image fusion: a structural patch decomposition approach. IEEE Transactions on Image Processing (TIP).26(5), 2519–2532 (2017)

    2017

  28. [28]

    Pacific Conference on Computer Graphics and Applications (PG)

    Mertens, T., Kautz, J., Van Reeth, F.: Exposure fusion. Pacific Conference on Computer Graphics and Applications (PG). pp. 382–390 (2007)

    2007

  29. [29]

    Computer graphics forum28(1), 161–171 (2009)

    Mertens, T., Kautz, J., Van Reeth, F.: Exposure fusion: A simple and practical alternative to high dynamic range photography. Computer graphics forum28(1), 161–171 (2009)

    2009

  30. [30]

    European Confer- ence on Computer Vision (ECCV)

    Prabhakar, K.R., Agrawal, S., Singh, D.K., Ashwath, B., Babu, R.V.: Towards practical and efficient high-resolution hdr deghosting with cnn. European Confer- ence on Computer Vision (ECCV). pp. 497–513 (2020)

    2020

  31. [31]

    IEEE Inter- national Conference on Computational Photography (ICCP)

    Prabhakar, K.R., Arora, R., Swaminathan, A., Singh, K.P., Babu, R.V.: A fast, scalable, and reliable deghosting method for extreme exposure fusion. IEEE Inter- national Conference on Computational Photography (ICCP). pp. 1–8 (2019)

    2019

  32. [32]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Qin, R., Liu, X., Liu, X., Liu, J., Shi, J., Lin, L., Yang, J.: No pains, more gains: Recycling sub-salient patches for efficient high-resolution image recognition. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 14965–14975 (2025)

    2025

  33. [33]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Qin, R., Liu, X., Shi, J., Lin, L., Yang, J.: Boosting the dual-stream architecture in ultra-high resolution segmentation with resolution-biased uncertainty estimation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 25960–25970 (2025)

    2025

  34. [34]

    arXiv preprint arXiv:2606.27905 (2026)

    Qu, L., Liu, Y., Zhou, S., Liang, J., Zeng, H., Zhang, L., Yang, J.: There and back again: A flexible-frame transformer for multi-exposure fusion. arXiv preprint arXiv:2606.27905 (2026)

    Pith/arXiv arXiv 2026

  35. [35]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Ram Prabhakar, K., Sai Srikar, V., Venkatesh Babu, R.: Deepfuse: A deep unsuper- vised approach for exposure fusion with extreme exposure image pairs. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 4714–4722 (2017)

    2017

  36. [36]

    Seminal Graphics Papers: Pushing the Boundaries pp

    Reinhard, E., Stark, M., Shirley, P., Ferwerda, J.: Photographic tone reproduction for digital images. Seminal Graphics Papers: Pushing the Boundaries pp. 661–670 (2023)

    2023

  37. [37]

    ACM Transactions on Graphics (TOG)

    Sen, P., Kalantari, N.K., Yaesoubi, M., Darabi, S., Goldman, D.B., Shechtman, E.: Robust patch-based hdr reconstruction of dynamic scenes. ACM Transactions on Graphics (TOG). pp. 203–1 (2012)

    2012

  38. [38]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Shu, Y., Shen, L., Hu, X., Li, M., Zhou, Z.: Towards real-world hdr video recon- struction: A large-scale benchmark dataset and a two-stage alignment network. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2879–2888 (2024)

    2024

  39. [39]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Shu, Y., Shen, L., Hu, X., Li, M., Zhou, Z.: Towards real-world hdr video recon- struction: A large-scale benchmark dataset and a two-stage alignment network. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2879–2888 (June 2024) 18 Y. Liu et al

    2024

  40. [40]

    European Con- ference on Computer Vision (ECCV)

    Song, J.W., Park, Y.I., Kong, K., Kwak, J., Kang, S.J.: Selective transhdr: Transformer-based selective hdr imaging using ghost region mask. European Con- ference on Computer Vision (ECCV). pp. 288–304 (2022)

    2022

  41. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Su, S., Yan, Q., Zhu, Y., Zhang, C., Ge, X., Sun, J., Zhang, Y.: Blindly assess image quality in the wild guided by a self-adaptive hyper network. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3667–3676 (2020)

    2020

  42. [42]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Tel, S., Wu, Z., Zhang, Y., Heyrman, B., Demonceaux, C., Timofte, R., Ginhac, D.: Alignment-free hdr deghosting with semantics consistent transformer. IEEE/CVF International Conference on Computer Vision (ICCV). (2023)

    2023

  43. [43]

    IEEE/CVF International Conference on Computer Vision (ICCV)

    Tel, S., Wu, Z., Zhang, Y., Heyrman, B., Demonceaux, C., Timofte, R., Ginhac, D.: Alignment-free hdr deghosting with semantics consistent transformer. IEEE/CVF International Conference on Computer Vision (ICCV). pp. 12836–12845 (2023)

    2023

  44. [44]

    Computer Graphics Forum35(2), 139–152 (2016)

    Tursun, O.T., Akyüz, A.O., Erdem, A., Erdem, E.: An objective deghosting quality metric for hdr images. Computer Graphics Forum35(2), 139–152 (2016)

    2016

  45. [45]

    AAAI Con- ference on Artificial Intelligence (AAAI).39(20), 21278–21286 (2025)

    Wang, X., Zeng, H., Chen, J., Liu, S., Chen, Y., Chao, G.: Otlrm: Orthogonal learning-based low-rank metric for multi-dimensional inverse problems. AAAI Con- ference on Artificial Intelligence (AAAI).39(20), 21278–21286 (2025)

    2025

  46. [46]

    IEEE Transactions on Image Processing (TIP)

    Wang, X., Zeng, H., Sun, B., Cao, J., Zhang, K., Shen, Q., Chen, Y.: Deep lora- unfolding networks for image restoration. IEEE Transactions on Image Processing (TIP). (2026)

    2026

  47. [47]

    arXiv preprint arXiv:2504.07667 (2025)

    Wang, Y., Wu, J., Bian, Y., Zhang, F., Xue, T.: S2r-hdr: A large-scale rendered dataset for hdr fusion. arXiv preprint arXiv:2504.07667 (2025)

    arXiv 2025

  48. [48]

    European Conference on Computer Vision (ECCV)

    Wu, S., Xu, J., Tai, Y.W., Tang, C.K.: Deep high dynamic range imaging with large foreground motions. European Conference on Computer Vision (ECCV). pp. 117–132 (2018)

    2018

  49. [49]

    European Conference on Computer Vision (ECCV)

    Wu, S., Xu, J., Tai, Y.W., Tang, C.K.: Deep high dynamic range imaging with large foreground motions. European Conference on Computer Vision (ECCV). pp. 120–135 (2018)

    2018

  50. [50]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).44(1), 502–518 (2020)

    Xu, H., Ma, J., Jiang, J., Guo, X., Ling, H.: U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).44(1), 502–518 (2020)

    2020

  51. [51]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yan, Q., Gong, D., Shi, Q., Hengel, A.v.d., Shen, C., Reid, I., Zhang, Y.: Attention- guided network for ghost-free high dynamic range imaging. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1751–1760 (2019)

    2019

  52. [52]

    Yan, Q., Hu, T., Sun, Y., Tang, H., Zhu, Y., Dong, W., Van Gool, L., Zhang, Y.: Towardhigh-qualityhdrdeghostingwithconditionaldiffusionmodels.IEEETrans- actions on Circuits and Systems for Video Technology34(5), 4011–4026 (2023)

    2023

  53. [53]

    IEEE Transactions on Image Processing (TIP).29, 4308– 4322 (2020)

    Yan, Q., Zhang, L., Liu, Y., Zhu, Y., Sun, J., Shi, Q., Zhang, Y.: Deep hdr imaging via a non-local network. IEEE Transactions on Image Processing (TIP).29, 4308– 4322 (2020)

    2020

  54. [54]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Ying, Z., Niu, H., Gupta, P., Mahajan, D., Ghadiyaram, D., Bovik, A.: From patches to pictures (paq-2-piq): Mapping the perceptual space of picture quality. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3575–3585 (2020)

    2020

  55. [55]

    IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

    You, Z., Cai, X., Gu, J., Xue, T., Dong, C.: Teaching large language models to regress accurate image quality scores using score distribution. IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR). pp. 14483–14494 (2025) ExpoMotion 19

    2025

  56. [56]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5728–5739 (2022)

    2022

  57. [57]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zeng, H., Wang, X., Chen, Y., Su, J., Liu, J.: Vision-language gradient descent- driven all-in-one deep unfolding networks. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 7524–7533 (2025)

    2025

  58. [58]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhang, N., Ye, Y., Zhao, Y., Wang, R.: Revisiting the stack-based inverse tone mapping. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9162–9171 (2023)

    2023

  59. [59]

    Information Fusion74, 111–131 (2021)

    Zhang, X.: Benchmarking and comparing multi-exposure image fusion algorithms. Information Fusion74, 111–131 (2021)

    2021

  60. [60]

    International Conference on Machine Learning (ICML)

    Zhao,Z.,Deng,L.,Bai,H.,Cui,Y.,Zhang,Z.,Zhang,Y.,Qin,H.,Chen,D.,Zhang, J., Wang, P., Van Gool, L.: Image fusion via vision-language model. International Conference on Machine Learning (ICML). (2024)

    2024

  61. [61]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

    Zhou,S.,Pan,J.,Yang,J.:Learninganadaptivesparsetransformerforefficientim- age restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI). (2025)

    2025

This paper was first reviewed by grok-4.5 on July 12, 2026.