acceptodds
Withdrawn from ICLR 2027

Beyond Ground-Truth Fidelity: Rethinking Exposure Correction through Input-Conditioned Perceptual Evaluation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Prevailing evaluation protocols for exposure correction (EC) prioritize fidelity to a designated ground truth (GT), implicitly treating the GT as a perfect perceptual reference. Yet a mis-exposed input can retain reliable scene content absent from the GT, raising the question of whether higher target fidelity necessarily implies a better correction. To examine this premise, we conduct an extensive human preference study with more than 30 observers, collecting over 60,000 pairwise judgments over 5,200 images. We generate visually diverse candidate corrections using representative EC models trained with different objectives that induce distinct trade-offs in exposure, color, and content preservation. In each trial, observers view the input alongside two candidate corrections and provide pairwise preferences among all three images. The resulting judgments reveal limited agreement between conventional metric rankings and human preferences, motivating us to reformulate EC quality in terms of input-conditioned perceptual improvement. We construct ECIQA-Pref from these judgments and develop ECIQA, a degraded-reference metric that models directed input–candidate relations using global appearance and native-resolution local structural evidence. ECIQA achieves the highest agreement with subjective scores among the evaluated full-, no-, and degraded-reference metrics, reaching 0.6482 SRCC and 0.8169 PLCC. A qualitative pilot study further suggests that ECIQA-guided training can better preserve input-visible structure, providing preliminary support for optimizing EC beyond target fidelity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Exposure correction, Degraded-reference metric, Human preference dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qf85OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper challenges the conventional evaluation of exposure correction methods, which primarily measures fidelity to a designated ground-truth image using metrics such as PSNR and SSIM. The authors argue that a mis-exposed input may preserve reliable scene details that are absent from the ground truth, meaning that greater numerical similarity does not necessarily indicate better perceptual quality.

Critical strengths

This paper identifies an important issue with existing exposure correction methods and proposes an approach to address it.

The authors have constructed a dataset, which required a considerable amount of effort.

The paper is well written and shows no obvious signs of AI-generated content.

Critical weaknesses

The authors appear to have conflated several concepts. Suppose that the dependent variable YY represents image quality, while the independent variables XX represent image attributes, such as exposure, color temperature, artifacts, and so on. Existing work on exposure correction can be divided into two settings:

  1. Improving YY by modifying only one component of XX, namely exposure, as in the SICE [1] dataset and UEC [2].
  2. For images degraded in terms of exposure, improving YY by modifying multiple components of XX, as in the MSEC dataset.

[1] Learning a Deep Single Image Contrast Enhancer from Multi-Exposure Images

[2] Unsupervised Exposure Correction

Your assessment framework seems to alternate between these two settings. First, you state that models trained on MSEC are used as candidates, which suggests that your work follows the second setting. Later, however, you construct the qMap based on MEF, which is more aligned with the first setting. Why are the other components of XX ignored at this stage? Furthermore, the first setting has already been investigated in prior work, such as P-IEANet [3] and IEMx [4].

[3] Rethinking No-reference Image Exposure Assessment from Holism to Pixel: Models, Datasets and Benchmarks

[4] Towards Automatic Image Exposure Level Assessment

In image enhancement, it may not be necessary to draw a strict distinction between these two settings, because enhancement models can potentially be applied across both. However, your work focuses on assessment, for which such ambiguity is problematic, since the problem formulation itself constitutes part of the contribution.

You also acknowledge that the variables XX considered in your work differ from those used in general-purpose IQA. This raises a fundamental question: What exactly does XX include in your formulation? This point is not clearly explained in the current manuscript. In contrast, [3] provides an explicit definition of the problem: it determines the image attributes XX to be assessed based on Adams' Zone System theory and subsequently develops corresponding scoring criteria.

Questions for authors

What exactly does XX include in your formulation?(Please see Critical Weaknesses for details)

Reviewer_U5tROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a perceptual preference formulation for exposure correction (EC) evaluation and introduces a degraded-reference metric called ECIQA, which utilizes evidence from global appearance and native-reference local structural evidence to model the directed input-candidate relations. Furthermore, the ECIQA-Pref dataset was constructed based on human preference judgments, and a two-stage training strategy combining structural knowledge distillation with preference learning was employed to improve alignment with human evaluations.

Critical strengths

The original input is used as scene evidence to evaluate the effectiveness of exposure correction, rather than being limited primarily to the degree of fidelity to the ground truth. The cross-attention mechanism combines global appearance with native-resolution local structural evidence at native resolution, supporting the assessment of perceptual quality improvements and content preservation. Furthermore, the ECIQA-Pref dataset provides supervision based on human preferences, enabling the original input to outperform results from failed corrections and thereby grounding the evaluation framework in human judgment.

Critical weaknesses

(1) The manuscript mentions that MEF-SSIM is used both to supervise the generation of candidate images and to pre-train the qMap encoder through knowledge distillation. However, this dual reliance may introduce a structural bias that causes the output to be optimized for MEF-SSIM. Therefore, please verify whether ECIQA can generalize to architectures that were not trained using the MEF-SSIM loss.

(2) The minimum-maximum normalization by scenario in the manuscript makes the target Elo scores set-dependent, whereas ECIQA evaluates isolated input-candidate pairs without considering the set context. The manuscript should clarify whether training based on these set-dependent targets leads to inconsistencies in learning. Furthermore, it is necessary to compare ECIQA with direct pairwise ranking baselines to demonstrate the superiority of the current methods.

(3) Although the ablation experiments in this manuscript demonstrate the necessity of input preprocessing by removing input features, they fail to prove that cross-attention from candidate features to the input outperforms simpler fusion strategies. Additional experiments using the same backbone network were conducted to compare the performance of cross-attention with that of standard feature fusion, and to contrast the performance of qMap with and without MEF-SSIM distillation.

(4) Conditional processing based on degraded inputs should ensure that the metric is capable of distinguishing valid scene structures from exposure artifacts; however, this paper does not clarify whether ECIQA has actually learned to make this distinction or merely exhibits a preference for conservative corrections. To verify this premise, perturbation stress tests (assessing sensitivity to isolated noise, color bias, and detail loss) can be conducted, and the classification accuracy can be evaluated to determine whether the candidate results are superior to the original inputs.

(5) Appendix A.1 notes that due to systematic green bias, the output of CoTF-ME requires manual color correction. However, the qMap encoder relies entirely on brightness-based MEF-SSIM supervision, and its ability to penalize chrominance errors has not yet been verified. The manuscript should evaluate ECIQA under controlled global and local color perturbations to demonstrate its sensitivity to color distortion.

(6) The visualization in Section 5.5 does not demonstrate whether ECIQA-guided training preserves true details or merely amplifies noise and sharpening artifacts. Furthermore, relying exclusively on ECIQA’s self-assessment scores to validate ECIQA remains insufficient; additional experiments evaluating details, color, and artifacts are needed to justify the validity of ECIQA as a training loss.

Questions for authors

(1) The manuscript mentions that MEF-SSIM is used both to supervise the generation of candidate images and to pre-train the qMap encoder through knowledge distillation. However, this dual reliance may introduce a structural bias that causes the output to be optimized for MEF-SSIM. Therefore, please verify whether ECIQA can generalize to architectures that were not trained using the MEF-SSIM loss.

(2) The minimum-maximum normalization by scenario in the manuscript makes the target Elo scores set-dependent, whereas ECIQA evaluates isolated input-candidate pairs without considering the set context. The manuscript should clarify whether training based on these set-dependent targets leads to inconsistencies in learning. Furthermore, it is necessary to compare ECIQA with direct pairwise ranking baselines to demonstrate the superiority of the current methods.

(3) Although the ablation experiments in this manuscript demonstrate the necessity of input preprocessing by removing input features, they fail to prove that cross-attention from candidate features to the input outperforms simpler fusion strategies. Additional experiments using the same backbone network were conducted to compare the performance of cross-attention with that of standard feature fusion, and to contrast the performance of qMap with and without MEF-SSIM distillation.

(4) Conditional processing based on degraded inputs should ensure that the metric is capable of distinguishing valid scene structures from exposure artifacts; however, this paper does not clarify whether ECIQA has actually learned to make this distinction or merely exhibits a preference for conservative corrections. To verify this premise, perturbation stress tests (assessing sensitivity to isolated noise, color bias, and detail loss) can be conducted, and the classification accuracy can be evaluated to determine whether the candidate results are superior to the original inputs.

(5) Appendix A.1 notes that due to systematic green bias, the output of CoTF-ME requires manual color correction. However, the qMap encoder relies entirely on brightness-based MEF-SSIM supervision, and its ability to penalize chrominance errors has not yet been verified. The manuscript should evaluate ECIQA under controlled global and local color perturbations to demonstrate its sensitivity to color distortion.

(6) The visualization in Section 5.5 does not demonstrate whether ECIQA-guided training preserves true details or merely amplifies noise and sharpening artifacts. Furthermore, relying exclusively on ECIQA’s self-assessment scores to validate ECIQA remains insufficient; additional experiments evaluating details, color, and artifacts are needed to justify the validity of ECIQA as a training loss.

Reviewer_MsdAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether results that are closer to the ground truth in EC truly correspond to better perceptual quality. The authors first construct the ECIQA-Pref dataset and use human preference experiments to examine the discrepancy between traditional image quality assessment and subjective quality. They then propose ECIQA, which predicts human preferences by jointly modeling the original input and the corrected result.

Critical strengths
  1. The research motivation is clear, and it addresses a genuine limitation in EC evaluation.
  2. The data collection was relatively large in scale, which will help with further research on EC evaluation.
  3. Input-conditioned evaluation allows the model to take into account the structural and color information in the original image.
  4. The proposed ECIQA can be used to train and optimize other EC models.
Critical weaknesses
  1. The paper motivates its evaluation through trade-offs in exposure, color, and content preservation, However, during the collection and evaluation process, Observers were not reminded to consider these three perspectives. I am concerned whether the observers have fully taken these factors into account, such as whether someone might have overlooked content preservation and only focused on exposure and color. While this protocol is reasonable for measuring overall preference, it does not reveal how observers balance these factors. Additional attribute-level annotations or analyses would help clarify which factors in EC evolution have a greater impact on overall preference.
  2. Insufficient comparison with aesthetic assessment methods. The comparisons focus on IQA methods and do not explicitly evaluate specialized aesthetic assessment models, such as ArtiMuse [1] and FGAesQ [2].
  3. Limited analysis of inter-annotator agreement: The paper does not report quantitative measures of inter-annotator agreement, which are important for assessing the reliability of the subjective preference labels used for training and evaluation.
  4. Limited quantitative validation of ECIQA-guided training. A quantitative result, such as a user study, is needed to demonstrate the effectiveness of using ECIQA to train EC models.

[1] Cao S, Ma N, Li J, et al. Artimuse: Fine-grained image aesthetics assessment with joint scoring and expert-level understanding[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026: 15313-15322.

[2] Yang Z, Wang J, Zhang Z, et al. Fine-grained Image Aesthetic Assessment: Learning Discriminative Scores from Relative Ranks[J]. arXiv preprint arXiv:2603.03907, 2026.

Questions for authors

See weakness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.