Pith. sign in

REVIEW 3 major objections 3 minor 37 references

A no-reference network predicts PSNR for low-light enhanced images, selecting a near-oracle candidate in 89.5% of held-out scenes while all seven NR-IQA baselines achieve 0% top-1.

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 ↗

BlindPSNR selects the best enhancement candidate by training a no-reference network to regress exactly to analytic PSNR, achieving 89.5% top-1 selection on held-out RealX3D.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Useful no-reference PSNR predictor for LLIE candidate selection, with a clean internal evaluation and honest limitations, but the lack of a held-out enhancement family leaves the general-fidelity claim untested. the 3 major comments →

arxiv 2607.27628 v1 pith:GMEMGNCM submitted 2026-07-30 cs.CV

BlindPSNR: A No-Reference Fidelity Predictor for Low-Light Image Enhancement

Mingzhe Lyu , Jinqiang Cui , Hong Zhang This is my paper
classification cs.CV
keywords no-reference PSNR predictionlow-light image enhancementcandidate selectionno-reference image quality assessmentheteroscedastic regressionwindowed cross-attentionanalytic log-MSE targetshyperparameter tuning
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

The paper claims that PSNR, the standard signal-fidelity metric that normally requires a ground-truth reference, can be predicted by a lightweight network that sees only the enhanced image and its degraded low-light input. The network is trained with exact analytic log-MSE labels computed from paired data, so no separately trained teacher is needed. On held-out scenes from a calibrated multi-view benchmark, it selects the enhancement configuration with the highest true PSNR in 89.5% of scene groups, and the chosen candidate's PSNR is within 0.1 dB of oracle in 51 of 57 groups. The paper also shows that seven existing no-reference quality metrics, which target perceptual quality rather than signal fidelity, fail completely at this selection task, several preferring systematically lower-fidelity outputs. This matters because automating per-scene parameter selection could remove a laborious manual step when deploying low-light enhancement.

Core claim

With paired training data, ground-truth PSNR is analytically computable as a single global log-MSE scalar, and BlindPSNR uses these exact targets to supervise a no-reference scoring network. The student takes both the enhanced image and its degraded low-light input, fuses them through windowed cross-attention whose windows shrink with feature stride, and regresses the log-MSE through a heteroscedastic head plus a complementary smooth-L1 distillation loss on the same target. On a held-out multi-view benchmark with scene-unseen groups, BlindPSNR reaches SRCC 0.960, top-1 89.5%, top-5 100%, and a mean regret of 0.026 dB, effectively matching the oracle in 51 of 57 scene groups. In contrast, all

What carries the argument

The paper's central mechanism is the analytic global log-MSE target, a scalar computed directly from enhanced ground-truth pairs, which supplies exact regression supervision without any teacher network; carrying the argument is the windowed cross-attention fusion of enhanced and low-light features, whose shrinking window sizes align dense local error patterns, and a two-term objective that combines heteroscedastic regression with smooth-L1 distillation on the same target.

Load-bearing premise

The near-oracle selection claim rests on the training corpus (2,423 images, 145,788 candidates from four data families and five enhancement methods) matching deployment conditions; the paper's own Table IV shows top-1 drops to 12-38% on out-of-distribution data, so if real deployment uses unseen enhancement families or scene statistics, the ranking may not track PSNR.

What would settle it

On a fresh calibrated multi-view benchmark with the same five enhancement families but entirely new scenes and lighting, measure BlindPSNR's top-1 selection accuracy and mean regret; if top-1 falls below roughly 85% or mean regret exceeds 0.5 dB, the claimed near-oracle selection within its training distribution is falsified.

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

If this is right

  • Automatic per-scene tuning of LLIE parameters becomes feasible without ground truth, removing manual candidate inspection for deployment.
  • Perceptual NR-IQA metrics should not be used as proxies for PSNR-based selection in low-light enhancement; the paper's results show they can be worse than random.
  • The high top-5 and near-zero regret on in-distribution scenes imply that a small set of top-ranked candidates can safely replace exhaustive search.
  • Within-method ranking over parameter variants (top-1 >= 89.5%, regret <= 0.04 dB across five families) supports using the network for hyperparameter search inside a single enhancement family.
  • Because cross-domain accuracy drops (top-1 12-38%), the near-oracle guarantee is limited to distributions similar to the training corpus; deployment requires retraining or broader data.

Where Pith is reading between the lines

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

  • The same analytic-supervision recipe could extend to other reference-based fidelity metrics that are not closed-form, such as SSIM or custom perceptual losses, by distilling their scores from a paired corpus into a no-reference dual-input network.
  • The large contribution of the low-light input (+29.9 pp) suggests enhancement artifacts are only interpretable relative to the input; a testable extension is whether dual-input architectures transfer to candidate selection in denoising or deblurring.
  • If the model's ranking is reliable, it could serve as a training signal (e.g., a loss term or reward) for generative LLIE methods, replacing hand-tuned losses with a fidelity-oriented selector—though the paper does not explore this.
  • The complete failure of seven NR-IQA baselines (0% top-1, some with negative SRCC) indicates perceptual quality and signal fidelity diverge sharply after low-light enhancement; this raises a broader question about which metric should guide LLIE optimization for downstream vision tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes BlindPSNR, a no-reference network that predicts the PSNR of a low-light enhanced image relative to the (unknown) ground truth, using only the enhanced image and the original low-light input. The network is trained on analytically computed global log-MSE targets derived from paired training data, with a heteroscedastic regression loss plus a smooth-L1 distillation loss. A windowed cross-attention mechanism fuses the degraded input with the enhanced image at multiple scales. Experiments on held-out RealX3D scenes report SRCC 0.960, 89.5% top-1 selection accuracy, and mean regret 0.026 dB, outperforming seven NR-IQA baselines that all score 0% top-1. Cross-dataset results on LOL-v2-real, LSRW, and mip-NeRF-360 show lower but positive correlation (SRCC 0.61–0.67). A leave-one-out ablation assigns the largest gain to the low-light input and the distillation loss.

Significance. If the results are robust, the paper addresses a genuine gap: NR-IQA methods target perceptual quality and are poorly aligned with PSNR-based fidelity, so a no-reference PSNR predictor for candidate selection in LLIE is practically useful. The analytic supervision target is a clean design choice that avoids teacher networks, and the leave-one-out ablation is a constructive way to attribute contributions. The paper ships code, which supports reproducibility. However, the strength of the central claim—that BlindPSNR can replace reference-based PSNR for selecting among LLIE candidates—is currently limited by the absence of a held-out enhancement-family test and by the lack of a chance-level baseline for the selection metrics. The cross-dataset generalization is modest, and the paper itself acknowledges that enhancement-method diversity is a key factor, so the scope of the claim must be narrowed or the missing experiments supplied.

major comments (3)
  1. [§III.B, Table II] The claim that all seven NR-IQA baselines 'fail entirely' (0% top-1) is not interpretable without a chance-level baseline. With 233 candidates per group, random selection yields top-1 accuracy of about 0.43% (and higher if the 0.1 dB tolerance is considered), so 0% top-1 is essentially at chance. The meaningful comparisons are SRCC and regret, where baselines are indeed poor, but the paper's framing overstates the failure. Please report a random (or chance) baseline for top-1, top-5, and expected regret on the same candidate sets, and discuss the selection metrics relative to that baseline.
  2. [§III.A–III.B, Tables II and IV] The training corpus is built from exactly five enhancement families (HVI-CIDNet, CIDNet-SICE, CLAHE, Gamma+Denoise, MSR), and the held-out RealX3D evaluation uses candidates generated from the same families. The cross-dataset sets in Table IV also use familiar families (e.g., tone-curve/gamma candidates for mip-NeRF-360). No experiment tests a held-out enhancement family. Because the model sees both L and E, it could learn family- and parameter-specific artifact signatures rather than content-dependent PSNR. The ablation in Table III (removing L costs 29.9 pp top-1) is consistent with this possibility. The Limitations section explicitly states that diversity of training enhancement methods is important, but no quantitative test is provided. Please add a held-out-family experiment—e.g., train on any four families and test on the fifth, or test on an enhancement method absent from training
  3. [§III.B, Tables II and III] All headline numbers come from a single training run, with no seed variance reported. The ablation differences (e.g., 89.5% vs. 78.9% for removing L_gdist) are of the same magnitude as typical seed-to-seed variation in training a 34.9M-parameter network. Please report mean and standard deviation over at least three random seeds for the main selection metrics, and note whether the ablation gaps are significant relative to this variance.
minor comments (3)
  1. [§III.C, Table V] For within-method selection, some rows report SRCC only around 0.77–0.83 while top-1 is above 96% and regret is very small. It would help to explain how near-perfect selection can coexist with moderate rank correlation (e.g., whether the top-PSNR candidate is identifiable even when ranking of lower candidates is noisy). This is a clarity point, not a correctness concern.
  2. [§II.D, Eq. (4)] The notation λ is introduced but its value (0.5) appears only in the text; please state it explicitly near the equation for self-containedness.
  3. [§I] The claim 'no learning-based method addresses no-reference PSNR prediction for low-light image enhancement' is strong. Please qualify by the cited codec-based methods and any video-compression PSNR estimators, since those are 'no-reference' in a limited sense. The current wording could be misread as covering all image-processing domains.

Circularity Check

0 steps flagged

No significant circularity: the PSNR supervision is an analytic function of ground truth, not of the model's own outputs, and the held-out evaluation is external, with generalization limitations honestly reported.

full rationale

The paper's derivation chain is non-circular. The training target is s = log(MSE(E,G)+epsilon), computed directly from the ground-truth G and the enhanced candidate E (Eq. 2), with PSNR recovered as -4.343s. This target does not depend on the network's parameters, predictions, or any fitted constant; it is an analytic full-reference label used to supervise a separate no-reference network. The model fuses E and L via windowed cross-attention and is trained with heteroscedastic regression plus a smooth-L1 distillation loss, both on the same analytic target. No equation in the paper defines the prediction in terms of the supervision or vice versa, and no parameter is fitted to the evaluation set. Evaluation on RealX3D is scene-unseen, and cross-dataset results on LOL-v2-real, LSRW, and mip-NeRF-360 are reported, with the paper explicitly acknowledging reduced top-1 accuracy (12-38%) and attributing the gap to domain shift. The skeptic's concern that the model may exploit enhancement-family identity rather than true fidelity is a generalization/memorization concern about external validity, not a definitional circularity: the supervision still measures true PSNR, and the paper does not claim to have tested an unseen enhancement family. References to prior work are standard external citations, not self-citations, and no uniqueness theorem or ansatz is imported through author self-citation. The paper's own limitations section is consistent with a genuine, albeit bounded, empirical claim rather than a circular construction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or formal entities. Its contributions are a learned regression model, an analytic training target, and an author-constructed candidate-selection benchmark; all load-bearing assumptions are domain assumptions about the representation and transferability of PSNR error patterns.

free parameters (4)
  • ConvNeXt-Tiny + head weights (~34.9M) = learned on training corpus
    All PSNR predictions are produced by these learned weights; the paper offers no closed-form or parameter-free route to PSNR.
  • Loss weight lambda for L_gdist = 0.5
    Hand-set balance between heteroscedastic regression and smooth-L1 distillation; not swept, only removed wholesale in ablation.
  • Cross-attention window sizes = 8, 4, 2 at strides 8/16/32
    Chosen by design to shrink with stride; no ablation varies window size.
  • Epsilon in log-MSE target = 10^-6
    Regularization constant preventing log(0); standard but hand-chosen, affects absolute scale only.
axioms (4)
  • domain assumption PSNR/log-MSE is the correct fidelity criterion for selecting LLIE candidates
    The entire evaluation treats the candidate with highest GT PSNR as the oracle; the paper does not validate PSNR selection against downstream task performance or perceptual preference.
  • domain assumption A function from (L,E) to PSNR(E,G) exists and generalizes across scenes within a training distribution
    Eq. (1) assumes the low-light input plus enhanced image contain enough information about the unknown reference for a learned network to rank candidates.
  • domain assumption The multi-source training corpus is representative of LLIE deployment
    Section III.A; the strongest results are scene-unseen within the same datasets, and Table IV shows severe top-1 degradation on domain-unseen data.
  • domain assumption ImageNet-pretrained ConvNeXt features provide a useful prior for pixel-level error prediction
    Backbone choice is inherited from prior literature; no experiment in the paper tests whether pretraining is required.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of BlindPSNR: A No-Reference Fidelity Predictor for Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/GMEMGNCM

@misc{pith2026260727628,
  author       = {Pith},
  title        = {Pith review of: BlindPSNR: A No-Reference Fidelity Predictor for Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GMEMGNCM}},
  note         = {Machine review of arXiv:2607.27628}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Low-light image enhancement (LLIE) methods involve tunable parameters that are typically fixed, often leading to performance degradation when applied across scenes. Manually selecting the best configuration, however, can be time-consuming and not always practical. Peak signal-to-noise ratio (PSNR) is the natural fidelity criterion for automating parameter selection, yet it requires a ground-truth reference that is typically unavailable. To our knowledge, no learning-based method addresses no-reference PSNR prediction for low-light image enhancement; the natural surrogate, no-reference image quality assessment (NR-IQA), targets perceptual quality rather than signal fidelity, and all seven baselines we test achieve 0% top-1 selection accuracy on our benchmark. With paired training data, the ground-truth PSNR is analytically computable, providing exact supervision without a separate teacher network. Building on this, we propose BlindPSNR, a lightweight no-reference network that fuses the enhanced image with the degraded low-light input via windowed cross-attention and estimates PSNR through heteroscedastic regression. While a scalar-regression baseline achieves top-1 accuracy of 54.4%, BlindPSNR raises this to 89.5% with regret dropping from 1.62 dB to 0.026 dB, and generalizes to unseen datasets (SRCC = 0.61-0.67).

Figures

Figures reproduced from arXiv: 2607.27628 by Hong Zhang, Jinqiang Cui, Mingzhe Lyu.

Figure 1
Figure 1. Figure 1: Overview of BlindPSNR. Top (blue, dashed, training only): an analytic full-reference module computes the global log-MSE target s from enhanced– ground-truth pairs. Bottom (green, solid, deployed): the no-reference scoring network takes the enhanced image E and low-light input L through a shared ConvNeXt backbone, windowed cross-attention fusion, and an FPN decoder; global average pooling (GAP) over the res… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative candidate selection on two held-out scenes (rows 1–3: CIDNet-SICE [ [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Held-out RealX3D evaluation. (a) Predicted vs. GT PSNR for all [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] 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

37 extracted references

  1. [1]

    FeatEn- Hancer: Enhancing hierarchical features for object detection and beyond under low-light vision,

    K. A. Hashmi, G. Kallempudi, D. Stricker, and M. Z. Afzal, “FeatEn- Hancer: Enhancing hierarchical features for object detection and beyond under low-light vision,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 12 830–12 840

    2023

  2. [2]

    LoLI-Street: Benchmarking low-light image enhancement and beyond,

    M. T. Islam, B. M. S. Nazmul Karim, A. Jain, G. Bertasius, and A. K. M. Mohiuddin, “LoLI-Street: Benchmarking low-light image enhancement and beyond,” inProc. Asian Conference on Computer Vision (ACCV), 2024, pp. 341–358

    2024

  3. [3]

    Contrast limited adaptive histogram equalization,

    K. Zuiderveld, “Contrast limited adaptive histogram equalization,” in Graphics Gems IV. Academic Press, 1994, pp. 474–485

    1994

  4. [4]

    Deep Retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep Retinex decomposition for low-light enhancement,” inProc. British Machine Vision Conference (BMVC), 2018

    2018

  5. [5]

    HVI: A new color space for low-light image enhancement,

    Q. Yan, Y . Feng, C. Zhang, G. Pang, K. Shi, P. Wu, W. Dong, J. Sun, and Y . Zhang, “HVI: A new color space for low-light image enhancement,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 5678–5687

    2025

  6. [6]

    Retinex- former: One-stage Retinex-based transformer for low-light image en- hancement,

    Y . Cai, H. Bian, J. Lin, H. Wang, R. Timofte, and Y . Zhang, “Retinex- former: One-stage Retinex-based transformer for low-light image en- hancement,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 12 504–12 513

    2023

  7. [7]

    LIME: Low-light image enhancement via illumination map estimation,

    X. Guo, Y . Li, and H. Ling, “LIME: Low-light image enhancement via illumination map estimation,”IEEE Transactions on Image Processing, vol. 26, no. 2, pp. 982–993, 2017

    2017

  8. [8]

    Zero- reference deep curve estimation for low-light image enhancement,

    C. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, and R. Cong, “Zero- reference deep curve estimation for low-light image enhancement,” in Proc. IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2020, pp. 1780–1789

    2020

  9. [9]

    Low-light image and video enhancement using deep learning: A sur- vey,

    C. Li, C. Guo, L. Han, J. Jiang, M.-M. Cheng, J. Gu, and C. C. Loy, “Low-light image and video enhancement using deep learning: A sur- vey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 12, pp. 9396–9416, 2022

    2022

  10. [10]

    Sparse gradient regularized deep Retinex network for robust low-light image enhance- ment,

    W. Yang, W. Wang, H. Huang, S. Wang, and J. Liu, “Sparse gradient regularized deep Retinex network for robust low-light image enhance- ment,”IEEE Transactions on Image Processing, vol. 30, pp. 2072–2086, 2021

    2072

  11. [11]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal Processing Letters, vol. 20, no. 3, pp. 209–212, 2013

    2013

  12. [12]

    No-reference image quality assessment in the spatial domain,

    A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,”IEEE Transactions on Image Processing, vol. 21, no. 12, pp. 4695–4708, 2012

    2012

  13. [13]

    Blind image quality evaluation using perception based features,

    N. Venkatanath, D. Praneeth, B. M. Chandrasekhar, S. S. Channappayya, and S. S. Medasani, “Blind image quality evaluation using perception based features,” inProc. National Conference on Communications (NCC). IEEE, 2015, pp. 1–6

    2015

  14. [14]

    MUSIQ: Multi-scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “MUSIQ: Multi-scale image quality transformer,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 5128–5137

    2021

  15. [15]

    MANIQA: Multi-dimension attention network for no-reference image quality assessment,

    S. Yang, T. Wu, S. Shi, S. Lao, Y . Gong, M. Cao, J. Wang, and Y . Yang, “MANIQA: Multi-dimension attention network for no-reference image quality assessment,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2022, pp. 1191–1200

    2022

  16. [16]

    Blindly assess image quality in the wild guided by a self-adaptive hyper network,

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

    2020

  17. [17]

    TOPIQ: A top-down approach from semantics to distortions for image quality assessment,

    C. Chen, J. Mo, J. Hou, H. Wu, L. Liao, W. Sun, Q. Yan, and W. Lin, “TOPIQ: A top-down approach from semantics to distortions for image quality assessment,”IEEE Transactions on Image Processing, vol. 33, pp. 2404–2418, 2024

    2024

  18. [18]

    A comprehensive experiment- based review of low-light image enhancement methods and benchmark- ing low-light image quality assessment,

    M. T. Rasheed, D. Shi, and H. Khan, “A comprehensive experiment- based review of low-light image enhancement methods and benchmark- ing low-light image quality assessment,”Signal Processing, vol. 204, p. 108821, 2023

    2023

  19. [19]

    Learning con- ditional knowledge distillation for degraded-reference image quality assessment,

    H. Zheng, H. Yang, J. Fu, Z.-J. Zha, and J. Luo, “Learning con- ditional knowledge distillation for degraded-reference image quality assessment,” inProc. IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 10 242–10 251

    2021

  20. [20]

    Content-variant reference image quality assessment via knowledge distillation,

    G. Yin, W. Wang, Z. Yuan, C. Han, W. Ji, S. Sun, and C. Wang, “Content-variant reference image quality assessment via knowledge distillation,” inProc. AAAI Conference on Artificial Intelligence, vol. 36, no. 3, 2022, pp. 3134–3142

    2022

  21. [21]

    Omnidirectional image quality as- sessment with knowledge distillation,

    L. Liu, P. Ma, C. Wang, and D. Xu, “Omnidirectional image quality as- sessment with knowledge distillation,”IEEE Signal Processing Letters, vol. 30, pp. 1562–1566, 2023

    2023

  22. [22]

    No- reference VMAF: A deep neural network-based approach to blind video quality assessment,

    A. De Decker, J. De Cock, P. Lambert, and G. Van Wallendael, “No- reference VMAF: A deep neural network-based approach to blind video quality assessment,”IEEE Transactions on Broadcasting, vol. 70, no. 3, pp. 844–861, 2024

    2024

  23. [23]

    Blind multimodal quality assessment of low-light images,

    M. Wang, Z. Xu, M. Xu, and W. Lin, “Blind multimodal quality assessment of low-light images,”International Journal of Computer Vision, vol. 133, pp. 1665–1688, 2025

    2025

  24. [24]

    Gap- closing matters: Perceptual quality evaluation and optimization of low- light image enhancement,

    B. Chen, L. Zhu, H. Zhu, W. Yang, L. Song, and S. Wang, “Gap- closing matters: Perceptual quality evaluation and optimization of low- light image enhancement,”IEEE Transactions on Multimedia, vol. 26, pp. 3430–3443, 2024

    2024

  25. [25]

    No reference PSNR estimation for compressed pictures,

    D. S. Turaga, Y . Chen, and J. Caviedes, “No reference PSNR estimation for compressed pictures,”Signal Processing: Image Communication, vol. 19, no. 2, pp. 173–184, 2004

    2004

  26. [26]

    A method of estimating coding PSNR using quantized DCT coefficients,

    A. Ichigaya, M. Kurosawa, and H. Hara, “A method of estimating coding PSNR using quantized DCT coefficients,”IEEE Trans. Circuits Syst. Video Technol., vol. 16, no. 2, pp. 251–259, 2006

    2006

  27. [27]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11 976–11 986

    2022

  28. [28]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProc. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 2117–2125

    2017

  29. [29]

    What uncertainties do we need in Bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in Bayesian deep learning for computer vision?” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  30. [30]

    RealX3D: A physically-degraded 3D benchmark for multi-view visual restoration and reconstruction,

    S. Liu, C. Bao, Z. Cui, Y . Liu, X. Chu, L. Gu, M. V . Conde, R. Umagami, T. Hashimoto, Z. Hu, T. Xu, Y . Gan, Y . Kurose, and T. Harada, “RealX3D: A physically-degraded 3D benchmark for multi-view visual restoration and reconstruction,” arXiv preprint arXiv:2512.23437, 2025

    arXiv 2025

  31. [31]

    Aleth-NeRF: Illumination adaptive NeRF with concealing field assumption,

    Z. Cui, L. Gu, X. Sun, X. Ma, Y . Qiao, and T. Harada, “Aleth-NeRF: Illumination adaptive NeRF with concealing field assumption,” inProc. AAAI Conf. Artif. Intell. (AAAI), vol. 38, no. 2, 2024

    2024

  32. [32]

    Learning a deep single image contrast enhancer from multi-exposure images,

    J. Cai, S. Gu, and L. Zhang, “Learning a deep single image contrast enhancer from multi-exposure images,”IEEE Transactions on Image Processing, vol. 27, no. 4, pp. 2049–2062, 2018

    2049

  33. [33]

    A multiscale Retinex for bridging the gap between color images and the human observation of scenes,

    D. J. Jobson, Z.-u. Rahman, and G. A. Woodell, “A multiscale Retinex for bridging the gap between color images and the human observation of scenes,”IEEE Transactions on Image Processing, vol. 6, no. 7, pp. 965–976, 1997

    1997

  34. [34]

    Mip-NeRF 360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman, “Mip-NeRF 360: Unbounded anti-aliased neural radiance fields,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022, pp. 5470–5479

    2022

  35. [35]

    R2RNet: Low-light image enhancement via real-low to real-normal network,

    J. Hai, Z. Xuan, R. Yang, Y . Hao, F. Zou, F. Lin, and S. Han, “R2RNet: Low-light image enhancement via real-low to real-normal network,”J. Vis. Commun. Image Represent., vol. 90, p. 103712, 2023

    2023

  36. [36]

    From RankNet to LambdaRank to LambdaMART: An overview,

    C. J. C. Burges, “From RankNet to LambdaRank to LambdaMART: An overview,” Microsoft Research, Tech. Rep. MSR-TR-2010-82, 2010

    2010

  37. [37]

    IQA-PyTorch: PyTorch toolbox for image quality assessment,

    C. Chen and J. Mo, “IQA-PyTorch: PyTorch toolbox for image quality assessment,” https://github.com/chaofengc/IQA-PyTorch, 2022

    2022

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