acceptodds
Under review as a conference paper at ICLR 2027

R2AS: Relative-to-Absolute Supervision for Real-World Dehazing

Abstract

Adapting image dehazing models to unlabeled real images requires supervision beyond physical reconstruction: different combinations of transmission and clean radiance can explain the same hazy image. Under a shared-atmospheric-light model, controlled haze interventions prescribe exact relative optical-depth shifts between views. This separates the inverse problem into a known relative structure and one unknown absolute offset per pixel. We propose Relative-to-Absolute Supervision (R2AS) for real-world dehazing. First, Relative-State Calibration (RSC) constructs views with known transmission scaling and trains a teacher to match their prescribed optical-depth differences while enforcing reconstruction compatibility with the observations. Second, Absolute-State Anchoring (ASA) projects the teacher's aligned transmission and radiance predictions onto a shared, observation-compatible family, producing joint targets through a closed-form scalar solution at each pixel. The calibrated teacher remains frozen while these targets supervise a student that retains the backbone architecture and requires a single dehazer evaluation at inference. We establish a bound on relative optical-depth error and prove that anchoring does not increase weighted joint target error when the true state lies in the feasible family. Experiments on six real-world benchmarks demonstrate that R2AS achieves the highest MUSIQ on all four unpaired datasets and the highest PSNR on both paired datasets among the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Single-image dehazing, Real-world image restoration, Unsupervised domain adaptation, Self-supervised image restoration, Inverse problems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SsuwOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the supervision ambiguity in unlabeled real-image dehazing: satisfying the atmospheric scattering reconstruction relation does not mean that the predicted transmission corresponds to the true scene behind the clear image; even enforcing consistency across different constructed views may retain a common absolute-state offset. The authors accordingly propose Relative-to-Absolute Supervision (R2AS), which separates the relative information that can be determined by controlled haze intervention from the absolute-state selection that requires prior guidance from a teacher.

Specifically, RSC performs relative calibration using the known optical depth differences between constructed views, combined with observation reconstruction constraints. ASA then projects the aligned multi-view teacher predictions onto a one-dimensional feasible set to obtain a shared absolute-state supervision target, and trains the student model via distillation. The paper provides a relative error bound and a conditional projection guarantee, and validates the method on four unpaired datasets and two paired datasets, together with experiments on modules, losses, the number of views, and atmospheric light perturbation.

Critical strengths
  1. The authors do not equate reconstruction consistency with recovery correctness; instead, they separate the identifiable relative information from the prior-guided absolute selection. RSC and ASA each take on one of these responsibilities, so the problem analysis maps directly onto the training objective and the algorithmic pipeline. In my view, this decomposition of the supervision design is the paper's most commendable contribution.

  2. The paper compares the baseline, RSC-only, ASA-only, and the full model, and further ablates the RSC losses, tests intervention strengths outside the training range, and analyzes the number of views and atmospheric light perturbation. The reduction in relative optical depth error and observation error means the experiments go beyond no-reference quality scores. The results on the paired datasets also provide supplementary support for the recovery quality.

Critical weaknesses

The overall contribution of ASA is supported, but the independent contribution of the projection operation relative to ordinary distillation and multi-view aggregation has not been fully isolated. The existing module ablations show that adding ASA improves several metrics, and the multi-view experiments show that adding teacher evidence is beneficial. However, ASA simultaneously involves teacher supervision, multi-view information integration, feasible-set projection, and additional student training. Therefore, the current results cannot yet fully distinguish which step the gains mainly come from.

Questions for authors
  1. The existing anchoring-strength experiments cover multiple non-zero strengths, but do not yet directly show the corresponding result for no projection at all. With everything else held constant, how does the method perform when using the raw teacher targets, or only the aligned multi-view averaged targets?

  2. The full model improves on several metrics, but, for example, TOPIQ on RTTS is lower than the RSC-only result. The authors have already reported this phenomenon. Can representative examples be analyzed to determine whether this corresponds to a trade-off in residual haze, color, texture, or contrast, so as to analyze the specific quality dimensions that ASA improves?

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

his paper presents R2AS, a two‑stage unsupervised domain‑adaptation framework for real‑world single‑image dehazing built upon HazeFlow. The work addresses the intrinsic ambiguity of atmospheric scattering model: one hazy observation corresponds to infinite pairs of transmission and scene radiance. The authors decompose the ill‑posed problem into relative‑state calibration (RSC) and absolute‑state anchoring (ASA). RSC constructs synthetic haze‑modified views and enforces prescribed optical‑depth difference constraints; ASA projects multi‑view teacher predictions onto a feasible convex set to produce closed‑form absolute restoration targets for student distillation. Theoretical propositions bound relative optical‑depth error and prove projection‑induced error reduction under feasible‑state assumptions.

Critical strengths

The paper well articulates the core non‑identifiability issue under atmospheric scattering model: only relative optical‑depth shift can be determined from controlled haze interventions, while absolute offset remains unknown. This conceptual separation between relative constraint and absolute anchoring is insightful for physics‑driven unsupervised dehazing.

Critical weaknesses
  1. Real physical haze in raw sensor data brings non‑ideal factors: spatially‑variant atmospheric light, multiple scattering, sensor noise, saturation, vignetting, nonlinear camera response. The paper’s core theory strictly assumes global constant atmospheric light and linear radiance, which hardly holds for raw real sensor data. The paper only tests processed sRGB compressed benchmark images, and lacks any raw‑sensor real‑world evaluation.

  2. The perturbation test only adds small numerical offsets to pre‑computed DCP atmospheric‑light value in simulation. It does not test real erroneous atmospheric‑light estimation failures that frequently occur in heavy haze, sky‑free scenes or saturated regions. Real‑world A estimation failure mode is not studied.

  3. Several recent unfolding / physical‑constraint self‑supervised dehazing approaches are absent in related work and comparison table. For example, physics‑based self‑supervised dehazing using consistency constraint, real‑world unfolding dehazing, and other augmentation‑based view‑consistency dehazing methods. The view‑augmentation idea used in RSC is not entirely new; several prior works adopt haze‑image augmentation and multi‑view consistency loss for real‑world adaptation, which are insufficiently discussed.

Questions for authors

Several recent unfolding / physical‑constraint self‑supervised dehazing approaches are absent in related work and comparison table. For example, physics‑based self‑supervised dehazing using consistency constraint, real‑world unfolding dehazing, and other augmentation‑based view‑consistency dehazing methods. The view‑augmentation idea used in RSC is not entirely new; several prior works adopt haze‑image augmentation and multi‑view consistency loss for real‑world adaptation, which are insufficiently discussed.

Reviewer_srrLOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes R2AS, a framework for adapting dehazing models to real-world images without paired ground truth. It first uses controlled haze changes to learn relative optical depth differences (RSC), then projects multi-view teacher predictions onto a physically consistent state to generate pseudo-labels (ASA). Experiments on six real-world benchmarks demonstrate its effectiveness.

Critical strengths
  1. The paper addresses an important issue in real-world dehazing: the ambiguity of recovering transmission and clean images from physical reconstruction alone.
  2. The proposed framework is physically motivated and supported by theoretical analysis.
  3. The method achieves promising results across multiple real-world benchmarks without introducing additional inference overhead.
Critical weaknesses
  1. The qualitative improvements of R2AS are not always convincing. In the first example of Fig. 3 and the second examples of Figs. 4 and 5, R2AS shows no clear visual advantage over other methods, with some residual haze still visible. The qualitative results do not fully reflect the advantages seen in the quantitative metrics.
  2. Insufficient Analysis of ASA's Contribution. Although Table 3 shows the effectiveness of ASA, it is unclear how much of the improvement comes from the proposed projection rather than the additional teacher-student distillation. Comparisons with direct distillation and simple multi-view averaging would help clarify its contribution.
  3. ASA aims to produce more accurate pseudo-labels, but their quality is not directly evaluated. Comparing teacher predictions before and after projection on paired datasets would provide stronger evidence for its effectiveness.
Questions for authors
  1. Since Eq. (24) shows that the observation consistency loss is essentially a scaled reconstruction loss, is there any practical advantage over directly using a standard physical reconstruction loss? A direct comparison would help justify this design.
  2. RSC uses only two views, although Appendix A.1 provides a general multi-view formulation. Have the authors explored whether using more views during RSC could further improve calibration?
1 comment
Reviewer_srrLOpenReview ↗

no response here

I didn't see a response here!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.